bachelor thesis ai

The Future of AI Research: 20 Thesis Ideas for Undergraduate Students in Machine Learning and Deep Learning for 2023!

A comprehensive guide for crafting an original and innovative thesis in the field of ai..

By Aarafat Islam on 2023-01-11

“The beauty of machine learning is that it can be applied to any problem you want to solve, as long as you can provide the computer with enough examples.” — Andrew Ng

This article provides a list of 20 potential thesis ideas for an undergraduate program in machine learning and deep learning in 2023. Each thesis idea includes an  introduction , which presents a brief overview of the topic and the  research objectives . The ideas provided are related to different areas of machine learning and deep learning, such as computer vision, natural language processing, robotics, finance, drug discovery, and more. The article also includes explanations, examples, and conclusions for each thesis idea, which can help guide the research and provide a clear understanding of the potential contributions and outcomes of the proposed research. The article also emphasized the importance of originality and the need for proper citation in order to avoid plagiarism.

1. Investigating the use of Generative Adversarial Networks (GANs) in medical imaging:  A deep learning approach to improve the accuracy of medical diagnoses.

Introduction:  Medical imaging is an important tool in the diagnosis and treatment of various medical conditions. However, accurately interpreting medical images can be challenging, especially for less experienced doctors. This thesis aims to explore the use of GANs in medical imaging, in order to improve the accuracy of medical diagnoses.

2. Exploring the use of deep learning in natural language generation (NLG): An analysis of the current state-of-the-art and future potential.

Introduction:  Natural language generation is an important field in natural language processing (NLP) that deals with creating human-like text automatically. Deep learning has shown promising results in NLP tasks such as machine translation, sentiment analysis, and question-answering. This thesis aims to explore the use of deep learning in NLG and analyze the current state-of-the-art models, as well as potential future developments.

3. Development and evaluation of deep reinforcement learning (RL) for robotic navigation and control.

Introduction:  Robotic navigation and control are challenging tasks, which require a high degree of intelligence and adaptability. Deep RL has shown promising results in various robotics tasks, such as robotic arm control, autonomous navigation, and manipulation. This thesis aims to develop and evaluate a deep RL-based approach for robotic navigation and control and evaluate its performance in various environments and tasks.

4. Investigating the use of deep learning for drug discovery and development.

Introduction:  Drug discovery and development is a time-consuming and expensive process, which often involves high failure rates. Deep learning has been used to improve various tasks in bioinformatics and biotechnology, such as protein structure prediction and gene expression analysis. This thesis aims to investigate the use of deep learning for drug discovery and development and examine its potential to improve the efficiency and accuracy of the drug development process.

5. Comparison of deep learning and traditional machine learning methods for anomaly detection in time series data.

Introduction:  Anomaly detection in time series data is a challenging task, which is important in various fields such as finance, healthcare, and manufacturing. Deep learning methods have been used to improve anomaly detection in time series data, while traditional machine learning methods have been widely used as well. This thesis aims to compare deep learning and traditional machine learning methods for anomaly detection in time series data and examine their respective strengths and weaknesses.

Photo by  Joanna Kosinska  on  Unsplash

6. Use of deep transfer learning in speech recognition and synthesis.

Introduction:  Speech recognition and synthesis are areas of natural language processing that focus on converting spoken language to text and vice versa. Transfer learning has been widely used in deep learning-based speech recognition and synthesis systems to improve their performance by reusing the features learned from other tasks. This thesis aims to investigate the use of transfer learning in speech recognition and synthesis and how it improves the performance of the system in comparison to traditional methods.

7. The use of deep learning for financial prediction.

Introduction:  Financial prediction is a challenging task that requires a high degree of intelligence and adaptability, especially in the field of stock market prediction. Deep learning has shown promising results in various financial prediction tasks, such as stock price prediction and credit risk analysis. This thesis aims to investigate the use of deep learning for financial prediction and examine its potential to improve the accuracy of financial forecasting.

8. Investigating the use of deep learning for computer vision in agriculture.

Introduction:  Computer vision has the potential to revolutionize the field of agriculture by improving crop monitoring, precision farming, and yield prediction. Deep learning has been used to improve various computer vision tasks, such as object detection, semantic segmentation, and image classification. This thesis aims to investigate the use of deep learning for computer vision in agriculture and examine its potential to improve the efficiency and accuracy of crop monitoring and precision farming.

9. Development and evaluation of deep learning models for generative design in engineering and architecture.

Introduction:  Generative design is a powerful tool in engineering and architecture that can help optimize designs and reduce human error. Deep learning has been used to improve various generative design tasks, such as design optimization and form generation. This thesis aims to develop and evaluate deep learning models for generative design in engineering and architecture and examine their potential to improve the efficiency and accuracy of the design process.

10. Investigating the use of deep learning for natural language understanding.

Introduction:  Natural language understanding is a complex task of natural language processing that involves extracting meaning from text. Deep learning has been used to improve various NLP tasks, such as machine translation, sentiment analysis, and question-answering. This thesis aims to investigate the use of deep learning for natural language understanding and examine its potential to improve the efficiency and accuracy of natural language understanding systems.

Photo by  UX Indonesia  on  Unsplash

11. Comparing deep learning and traditional machine learning methods for image compression.

Introduction:  Image compression is an important task in image processing and computer vision. It enables faster data transmission and storage of image files. Deep learning methods have been used to improve image compression, while traditional machine learning methods have been widely used as well. This thesis aims to compare deep learning and traditional machine learning methods for image compression and examine their respective strengths and weaknesses.

12. Using deep learning for sentiment analysis in social media.

Introduction:  Sentiment analysis in social media is an important task that can help businesses and organizations understand their customers’ opinions and feedback. Deep learning has been used to improve sentiment analysis in social media, by training models on large datasets of social media text. This thesis aims to use deep learning for sentiment analysis in social media, and evaluate its performance against traditional machine learning methods.

13. Investigating the use of deep learning for image generation.

Introduction:  Image generation is a task in computer vision that involves creating new images from scratch or modifying existing images. Deep learning has been used to improve various image generation tasks, such as super-resolution, style transfer, and face generation. This thesis aims to investigate the use of deep learning for image generation and examine its potential to improve the quality and diversity of generated images.

14. Development and evaluation of deep learning models for anomaly detection in cybersecurity.

Introduction:  Anomaly detection in cybersecurity is an important task that can help detect and prevent cyber-attacks. Deep learning has been used to improve various anomaly detection tasks, such as intrusion detection and malware detection. This thesis aims to develop and evaluate deep learning models for anomaly detection in cybersecurity and examine their potential to improve the efficiency and accuracy of cybersecurity systems.

15. Investigating the use of deep learning for natural language summarization.

Introduction:  Natural language summarization is an important task in natural language processing that involves creating a condensed version of a text that preserves its main meaning. Deep learning has been used to improve various natural language summarization tasks, such as document summarization and headline generation. This thesis aims to investigate the use of deep learning for natural language summarization and examine its potential to improve the efficiency and accuracy of natural language summarization systems.

Photo by  Windows  on  Unsplash

16. Development and evaluation of deep learning models for facial expression recognition.

Introduction:  Facial expression recognition is an important task in computer vision and has many practical applications, such as human-computer interaction, emotion recognition, and psychological studies. Deep learning has been used to improve facial expression recognition, by training models on large datasets of images. This thesis aims to develop and evaluate deep learning models for facial expression recognition and examine their performance against traditional machine learning methods.

17. Investigating the use of deep learning for generative models in music and audio.

Introduction:  Music and audio synthesis is an important task in audio processing, which has many practical applications, such as music generation and speech synthesis. Deep learning has been used to improve generative models for music and audio, by training models on large datasets of audio data. This thesis aims to investigate the use of deep learning for generative models in music and audio and examine its potential to improve the quality and diversity of generated audio.

18. Study the comparison of deep learning models with traditional algorithms for anomaly detection in network traffic.

Introduction:  Anomaly detection in network traffic is an important task that can help detect and prevent cyber-attacks. Deep learning models have been used for this task, and traditional methods such as clustering and rule-based systems are widely used as well. This thesis aims to compare deep learning models with traditional algorithms for anomaly detection in network traffic and analyze the trade-offs between the models in terms of accuracy and scalability.

19. Investigating the use of deep learning for improving recommender systems.

Introduction:  Recommender systems are widely used in many applications such as online shopping, music streaming, and movie streaming. Deep learning has been used to improve the performance of recommender systems, by training models on large datasets of user-item interactions. This thesis aims to investigate the use of deep learning for improving recommender systems and compare its performance with traditional content-based and collaborative filtering approaches.

20. Development and evaluation of deep learning models for multi-modal data analysis.

Introduction:  Multi-modal data analysis is the task of analyzing and understanding data from multiple sources such as text, images, and audio. Deep learning has been used to improve multi-modal data analysis, by training models on large datasets of multi-modal data. This thesis aims to develop and evaluate deep learning models for multi-modal data analysis and analyze their potential to improve performance in comparison to single-modal models.

I hope that this article has provided you with a useful guide for your thesis research in machine learning and deep learning. Remember to conduct a thorough literature review and to include proper citations in your work, as well as to be original in your research to avoid plagiarism. I wish you all the best of luck with your thesis and your research endeavors!

Continue Learning

The depth i: stereo calibration and rectification, pooling layer — short and simple.

Here's all the information you should know about Pooling Layer in CNN

Using Chains and Agents for LLM Application Development

Step-by-step guide to using chains and agents in LangChain

Generative AI vs. Discriminative AI

Breaking Down the Difference Between Generative AI and Discriminative AI and How Industry LLM Giants Cohere and OpenAI are Changing the Game in Generative AI

Future Trends in Conversational AI: What's Next?

The future of Conversational AI: from NLU advancements to ethical considerations. Discover emerging trends shaping software development.

Znote AI: The perfect sandbox for prototyping and deploying code

Completed Theses

State space search solves navigation tasks and many other real world problems. Heuristic search, especially greedy best-first search, is one of the most successful algorithms for state space search. We improve the state of the art in heuristic search in three directions.

In Part I, we present methods to train neural networks as powerful heuristics for a given state space. We present a universal approach to generate training data using random walks from a (partial) state. We demonstrate that our heuristics trained for a specific task are often better than heuristics trained for a whole domain. We show that the performance of all trained heuristics is highly complementary. There is no clear pattern, which trained heuristic to prefer for a specific task. In general, model-based planners still outperform planners with trained heuristics. But our approaches exceed the model-based algorithms in the Storage domain. To our knowledge, only once before in the Spanner domain, a learning-based planner exceeded the state-of-the-art model-based planners.

A priori, it is unknown whether a heuristic, or in the more general case a planner, performs well on a task. Hence, we trained online portfolios to select the best planner for a task. Today, all online portfolios are based on handcrafted features. In Part II, we present new online portfolios based on neural networks, which receive the complete task as input, and not just a few handcrafted features. Additionally, our portfolios can reconsider their choices. Both extensions greatly improve the state-of-the-art of online portfolios. Finally, we show that explainable machine learning techniques, as the alternative to neural networks, are also good online portfolios. Additionally, we present methods to improve our trust in their predictions.

Even if we select the best search algorithm, we cannot solve some tasks in reasonable time. We can speed up the search if we know how it behaves in the future. In Part III, we inspect the behavior of greedy best-first search with a fixed heuristic on simple tasks of a domain to learn its behavior for any task of the same domain. Once greedy best-first search expanded a progress state, it expands only states with lower heuristic values. We learn to identify progress states and present two methods to exploit this knowledge. Building upon this, we extract the bench transition system of a task and generalize it in such a way that we can apply it to any task of the same domain. We can use this generalized bench transition system to split a task into a sequence of simpler searches.

In all three research directions, we contribute new approaches and insights to the state of the art, and we indicate interesting topics for future work.

Greedy best-first search (GBFS) is a sibling of A* in the family of best-first state-space search algorithms. While A* is guaranteed to find optimal solutions of search problems, GBFS does not provide any guarantees but typically finds satisficing solutions more quickly than A*. A classical result of optimal best-first search shows that A* with admissible and consistent heuristic expands every state whose f-value is below the optimal solution cost and no state whose f-value is above the optimal solution cost. Theoretical results of this kind are useful for the analysis of heuristics in different search domains and for the improvement of algorithms. For satisficing algorithms a similarly clear understanding is currently lacking. We examine the search behavior of GBFS in order to make progress towards such an understanding.

We introduce the concept of high-water mark benches, which separate the search space into areas that are searched by GBFS in sequence. High-water mark benches allow us to exactly determine the set of states that GBFS expands under at least one tie-breaking strategy. We show that benches contain craters. Once GBFS enters a crater, it has to expand every state in the crater before being able to escape.

Benches and craters allow us to characterize the best-case and worst-case behavior of GBFS in given search instances. We show that computing the best-case or worst-case behavior of GBFS is NP-complete in general but can be computed in polynomial time for undirected state spaces.

We present algorithms for extracting the set of states that GBFS potentially expands and for computing the best-case and worst-case behavior. We use the algorithms to analyze GBFS on benchmark tasks from planning competitions under a state-of-the-art heuristic. Experimental results reveal interesting characteristics of the heuristic on the given tasks and demonstrate the importance of tie-breaking in GBFS.

Classical planning tackles the problem of finding a sequence of actions that leads from an initial state to a goal. Over the last decades, planning systems have become significantly better at answering the question whether such a sequence exists by applying a variety of techniques which have become more and more complex. As a result, it has become nearly impossible to formally analyze whether a planning system is actually correct in its answers, and we need to rely on experimental evidence.

One way to increase trust is the concept of certifying algorithms, which provide a witness which justifies their answer and can be verified independently. When a planning system finds a solution to a problem, the solution itself is a witness, and we can verify it by simply applying it. But what if the planning system claims the task is unsolvable? So far there was no principled way of verifying this claim.

This thesis contributes two approaches to create witnesses for unsolvable planning tasks. Inductive certificates are based on the idea of invariants. They argue that the initial state is part of a set of states that we cannot leave and that contains no goal state. In our second approach, we define a proof system that proves in an incremental fashion that certain states cannot be part of a solution until it has proven that either the initial state or all goal states are such states.

Both approaches are complete in the sense that a witness exists for every unsolvable planning task, and can be verified efficiently (in respect to the size of the witness) by an independent verifier if certain criteria are met. To show their applicability to state-of-the-art planning techniques, we provide an extensive overview how these approaches can cover several search algorithms, heuristics and other techniques. Finally, we show with an experimental study that generating and verifying these explanations is not only theoretically possible but also practically feasible, thus making a first step towards fully certifying planning systems.

Heuristic search with an admissible heuristic is one of the most prominent approaches to solving classical planning tasks optimally. In the first part of this thesis, we introduce a new family of admissible heuristics for classical planning, based on Cartesian abstractions, which we derive by counterexample-guided abstraction refinement. Since one abstraction usually is not informative enough for challenging planning tasks, we present several ways of creating diverse abstractions. To combine them admissibly, we introduce a new cost partitioning algorithm, which we call saturated cost partitioning. It considers the heuristics sequentially and uses the minimum amount of costs that preserves all heuristic estimates for the current heuristic before passing the remaining costs to subsequent heuristics until all heuristics have been served this way.

In the second part, we show that saturated cost partitioning is strongly influenced by the order in which it considers the heuristics. To find good orders, we present a greedy algorithm for creating an initial order and a hill-climbing search for optimizing a given order. Both algorithms make the resulting heuristics significantly more accurate. However, we obtain the strongest heuristics by maximizing over saturated cost partitioning heuristics computed for multiple orders, especially if we actively search for diverse orders.

The third part provides a theoretical and experimental comparison of saturated cost partitioning and other cost partitioning algorithms. Theoretically, we show that saturated cost partitioning dominates greedy zero-one cost partitioning. The difference between the two algorithms is that saturated cost partitioning opportunistically reuses unconsumed costs for subsequent heuristics. By applying this idea to uniform cost partitioning we obtain an opportunistic variant that dominates the original. We also prove that the maximum over suitable greedy zero-one cost partitioning heuristics dominates the canonical heuristic and show several non-dominance results for cost partitioning algorithms. The experimental analysis shows that saturated cost partitioning is the cost partitioning algorithm of choice in all evaluated settings and it even outperforms the previous state of the art in optimal classical planning.

Classical planning is the problem of finding a sequence of deterministic actions in a state space that lead from an initial state to a state satisfying some goal condition. The dominant approach to optimally solve planning tasks is heuristic search, in particular A* search combined with an admissible heuristic. While there exist many different admissible heuristics, we focus on abstraction heuristics in this thesis, and in particular, on the well-established merge-and-shrink heuristics.

Our main theoretical contribution is to provide a comprehensive description of the merge-and-shrink framework in terms of transformations of transition systems. Unlike previous accounts, our description is fully compositional, i.e. can be understood by understanding each transformation in isolation. In particular, in addition to the name-giving merge and shrink transformations, we also describe pruning and label reduction as such transformations. The latter is based on generalized label reduction, a new theory that removes all of the restrictions of the previous definition of label reduction. We study the four types of transformations in terms of desirable formal properties and explain how these properties transfer to heuristics being admissible and consistent or even perfect. We also describe an optimized implementation of the merge-and-shrink framework that substantially improves the efficiency compared to previous implementations.

Furthermore, we investigate the expressive power of merge-and-shrink abstractions by analyzing factored mappings, the data structure they use for representing functions. In particular, we show that there exist certain families of functions that can be compactly represented by so-called non-linear factored mappings but not by linear ones.

On the practical side, we contribute several non-linear merge strategies to the merge-and-shrink toolbox. In particular, we adapt a merge strategy from model checking to planning, provide a framework to enhance existing merge strategies based on symmetries, devise a simple score-based merge strategy that minimizes the maximum size of transition systems of the merge-and-shrink computation, and describe another framework to enhance merge strategies based on an analysis of causal dependencies of the planning task.

In a large experimental study, we show the evolution of the performance of merge-and-shrink heuristics on planning benchmarks. Starting with the state of the art before the contributions of this thesis, we subsequently evaluate all of our techniques and show that state-of-the-art non-linear merge-and-shrink heuristics improve significantly over the previous state of the art.

Admissible heuristics are the main ingredient when solving classical planning tasks optimally with heuristic search. Higher admissible heuristic values are more accurate, so combining them in a way that dominates their maximum and remains admissible is an important problem.

The thesis makes three contributions in this area. Extensions to cost partitioning (a well-known heuristic combination framework) allow to produce higher estimates from the same set of heuristics. The new heuristic family called operator-counting heuristics unifies many existing heuristics and offers a new way to combine them. Another new family of heuristics called potential heuristics allows to cast the problem of finding a good heuristic as an optimization problem.

Both operator-counting and potential heuristics are closely related to cost partitioning. They offer a new look on cost partitioned heuristics and already sparked research beyond their use as classical planning heuristics.

Master's theses

Optimal planning is an ongoing topic of research, and requires efficient heuristic search algorithms. One way of calculating such heuristics is through the use of Linear Programs (LPs) and solvers thereof. This thesis investigates the efficiency of LP-based heuristic search strategies of different heuristics, focusing on how different LP solving strategies and solver settings impact the performance of calculating these heuristics. Using the Fast Downward planning system and a comprehensive benchmark set of planning tasks, we conducted a series of experiments to determine the effectiveness of the primal and dual simplex methods and the primal-dual logarithmic barrier method. Our results show that the choice of the LP solver and the application of specific solver settings influence the efficiency of calculating the required heuristics, and showed that the default setting of CPLEX is not optimal in some cases and can be enhanced by specifying an LP-solver or using other non-default solver settings. This thesis lays the groundwork for future research of using different LP solving algorithms and solver settings in the context of LP-based heuristic search in optimal planning.

Classical planning tasks are typically formulated in PDDL. Some of them can be described more concisely using derived variables. Contrary to basic variables, their values cannot be changed by operators and are instead determined by axioms which specify conditions under which they take a certain value. Planning systems often support axioms in their search component, but their heuristics’ support is limited or nonexistent. This leads to decreased search performance with tasks that use axioms. We compile axioms away using our implementation of a known algorithm in the Fast Downward planner. Our results show that the compilation has a negative impact on search performance with its only benefit being the ability to use heuristics that have no axiom support. As a compromise between performance and expressivity, we identify axioms of a simple form and devise a compilation for them. We compile away all axioms in several of the tested domains without a decline in search performance.

The International Planning Competitions (IPCs) serve as a testing suite for planning sys- tems. These domains are well-motivated as they are derived from, or possess characteristics analogous to real-life applications. In this thesis, we study the computational complexity of the plan existence and bounded plan existence decision problems of the following grid- based IPC domains: VisitAll, TERMES, Tidybot, Floortile, and Nurikabe. In all of these domains, there are one or more agents moving through a rectangular grid (potentially with obstacles) performing actions along the way. In many cases, we engineer instances that can be solved only if the movement of the agent or agents follows a Hamiltonian path or cycle in a grid graph. This gives rise to many NP-hardness reductions from Hamiltonian path/cycle problems on grid graphs. In the case of VisitAll and Floortile, we give necessary and suffi- cient conditions for deciding the plan existence problem in polynomial time. We also show that Tidybot has the game Push -1F as a special case, and its plan existence problem is thus PSPACE-complete. The hardness proofs in this thesis highlight hard instances of these domains. Moreover, by assigning a complexity class to each domain, researchers and practitioners can better assess the strengths and limitations of new and existing algorithms in these domains.

Planning tasks can be used to describe many real world problems of interest. Solving those tasks optimally is thus an avenue of great interest. One established and successful approach for optimal planning is the merge-and-shrink framework, which decomposes the task into a factored transition system. The factors initially represent the behaviour of one state variable and are repeatedly combined and abstracted. The solutions of these abstract states is then used as a heuristic to guide search in the original planning task. Existing merge-and-shrink transformations keep the factored transition system orthogonal, meaning that the variables of the planning task are represented in no more than one factor at any point. In this thesis we introduce the clone transformation, which duplicates a factor of the factored transition system, making it non-orthogonal. We test two classes of clone strategies, which we introduce and implement in the Fast Downward planning system and conclude that, while theoretically promising, our clone strategies are practically inefficient as their performance was worse than state-of-the-art methods for merge-and-shrink.

This thesis aims to present a novel approach for improving the performance of classical planning algorithms by integrating cost partitioning with merge-and-shrink techniques. Cost partitioning is a well-known technique for admissibly adding multiple heuristic values. Merge-and-shrink, on the other hand, is a technique to generate well-informed abstractions. The "merge” part of the technique is based on creating an abstract representation of the original problem by replacing two transition systems with their synchronised product. In contrast, the ”shrink” part refers to reducing the size of the factor. By combining these two approaches, we aim to leverage the strengths of both methods to achieve better scalability and efficiency in solving classical planning problems. Considering a range of benchmark domains and the Fast Downward planning system, the experimental results show that the proposed method achieves the goal of fusing merge and shrink with cost partitioning towards better outcomes in classical planning.

Planning is the process of finding a path in a planning task from the initial state to a goal state. Multiple algorithms have been implemented to solve such planning tasks, one of them being the Property-Directed Reachability algorithm. Property-Directed Reachability utilizes a series of propositional formulas called layers to represent a super-set of states with a goal distance of at most the layer index. The algorithm iteratively improves the layers such that they represent a minimum number of states. This happens by strengthening the layer formulas and therefore excluding states with a goal distance higher than the layer index. The goal of this thesis is to implement a pre-processing step to seed the layers with a formula that already excludes as many states as possible, to potentially improve the run-time performance. We use the pattern database heuristic and its associated pattern generators to make use of the planning task structure for the seeding algorithm. We found that seeding does not consistently improve the performance of the Property-Directed Reachability algorithm. Although we observed a significant reduction in planning time for some tasks, it significantly increased for others.

Certifying algorithms is a concept developed to increase trust by demanding affirmation of the computed result in form of a certificate. By inspecting the certificate, it is possible to determine correctness of the produced output. Modern planning systems have been certifying for long time in the case of solvable instances, where a generated plan acts as a certificate.

Only recently there have been the first steps towards certifying unsolvability judgments in the form of inductive certificates which represent certain sets of states. Inductive certificates are expressed with the help of propositional formulas in a specific formalism.

In this thesis, we investigate the use of propositional formulas in conjunctive normal form (CNF) as a formalism for inductive certificates. At first, we look into an approach that allows us to construct formulas representing inductive certificates in CNF. To show general applicability of this approach, we extend this to the family of delete relaxation heuristics. Furthermore, we present how a planning system is able to generate an inductive validation formula, a single formula that can be used to validate if the set found by the planner is indeed an inductive certificate. At last, we show with an experimental evaluation that the CNF formalism can be feasible in practice for the generation and validation of inductive validation formulas.

In generalized planning the aim is to solve whole classes of planning tasks instead of single tasks one at a time. Generalized representations provide information or knowledge about such classes to help solving them. This work compares the expressiveness of three generalized representations, generalized potential heuristics, policy sketches and action schema networks, in terms of compilability. We use a notion of equivalence that requires two generalized representations to decompose the tasks of a class into the same subtasks. We present compilations between pairs of equivalent generalized representations and proofs where a compilation is impossible.

A Digital Microfluidic Biochip (DMFB) is a digitally controllable lab-on-a-chip. Droplets of fluids are moved, merged and mixed on a grid. Routing these droplets efficiently has been tackled by various different approaches. We try to use temporal planning to do droplet routing, inspired by the use of it in quantum circuit compilation. We test a model for droplet routing in both classical and temporal planning and compare both versions. We show that our classical planning model is an efficient method to find droplet routes on DMFBs. Then we extend our model and include spawning, disposing, merging, splitting and mixing of droplets. The results of these extensions show that we are able to find plans for simple experiments. When scaling the problem size to real life experiments our model fails to find plans.

Cost partitioning is a technique used to calculate heuristics in classical optimal planning. It involves solving a linear program. This linear program can be decomposed into a master and pricing problems. In this thesis we combine Fourier-Motzkin elimination and the double description method in different ways to precompute the generating rays of the pricing problems. We further empirically evaluate these approaches and propose a new method that replaces the Fourier-Motzkin elimination. Our new method improves the performance of our approaches with respect to runtime and peak memory usage.

The increasing number of data nowadays has contributed to new scheduling approaches. Aviation is one of the domains concerned the most, as the aircraft engine implies millions of maintenance events operated by staff worldwide. In this thesis we present a constraint programming-based algorithm to solve the aircraft maintenance scheduling problem. We want to find the best time to do the maintenance by determining which employee will perform the work and when. Here we report how the scheduling process in aviation can be automatized.

To solve stochastic state-space tasks, the research field of artificial intelligence is mainly used. PROST2014 is state of the art when determining good actions in an MDP environment. In this thesis, we aimed to provide a heuristic by using neural networks to outperform the dominating planning system PROST2014. For this purpose, we introduced two variants of neural networks that allow to estimate the respective Q-value for a pair of state and action. Since we envisaged the learning method of supervised learning, in addition to the architecture as well as the components of the neural networks, the generation of training data was also one of the main tasks. To determine the most suitable network parameters, we performed a sequential parameter search, from which we expected a local optimum of the model settings. In the end, the PROST2014 planning system could not be surpassed in the total rating evaluation. Nevertheless, in individual domains, we could establish increased final scores on the side of the neural networks. The result shows the potential of this approach and points to eventual adaptations in future work pursuing this procedure furthermore.

In classical planning, there are tasks that are hard and tasks that are easy. We can measure the complexity of a task with the correlation complexity, the improvability width, and the novelty width. In this work, we compare these measures.

We investigate what causes a correlation complexity of at least 2. To do so we translate the state space into a vector space which allows us to make use of linear algebra and convex cones.

Additionally, we introduce the Basel measure, a new measure that is based on potential heuristics and therefore similar to the correlation complexity but also comparable to the novelty width. We show that the Basel measure is a lower bound for the correlation complexity and that the novelty width +1 is an upper bound for the Basel measure.

Furthermore, we compute the Basel measure for some tasks of the International Planning Competitions and show that the translation of a task can increase the Basel measure by removing seemingly irrelevant state variables.

Unsolvability is an important result in classical planning and has seen increased interest in recent years. This thesis explores unsolvability detection by automatically generating parity arguments, a well-known way of proving unsolvability. The argument requires an invariant measure, whose parity remains constant across all reachable states, while all goal states are of the opposite parity. We express parity arguments using potential functions in the field F 2 . We develop a set of constraints that describes potential functions with the necessary separating property, and show that the constraints can be represented efficiently for up to two-dimensional features. Enhanced with mutex information, an algorithm is formed that tests whether a parity function exists for a given planning task. The existence of such a function proves the task unsolvable. To determine its practical use, we empirically evaluate our approach on a benchmark of unsolvable problems and compare its performance to a state of the art unsolvability planner. We lastly analyze the arguments found by our algorithm to confirm their validity, and understand their expressive power.

We implemented the invariant synthesis algorithm proposed by Rintanen and experimentally compared it against Helmert’s mutex group synthesis algorithm as implemented in Fast Downward.

The context for the comparison is the translation of propositional STRIPS tasks to FDR tasks, which requires the identification of mutex groups.

Because of its dominating lead in translation speed, combined with few and marginal advantages in performance during search, Helmert’s algorithm is clearly better for most uses. Meanwhile Rintanen’s algorithm is capable of finding invariants other than mutexes, which Helmert’s algorithm per design cannot do.

The International Planning Competition (IPC) is a competition of state-of-the-art planning systems. The evaluation of these planning systems is done by measuring them with different problems. It focuses on the challenges of AI planning by analyzing classical, probabilistic and temporal planning and by presenting new problems for future research. Some of the probabilistic domains introduced in IPC 2018 are Academic Advising, Chromatic Dice, Cooperative Recon, Manufacturer, Push Your Luck, Red-finned Blue-eyes, etc.

This thesis aims to solve (near)-optimally two probabilistic IPC 2018 domains, Academic Advising and Chromatic Dice. We use different techniques to solve these two domains. In Academic Advising, we use a relevance analysis to remove irrelevant actions and state variables from the planning task. We then convert the problem from probabilistic to classical planning, which helped us solve it efficiently. In Chromatic Dice, we implement backtracking search to solve the smaller instances optimally. More complex instances are partitioned into several smaller planning tasks, and a near-optimal policy is derived as a combination of the optimal solutions to the small instances.

The motivation for finding (near)-optimal policies is related to the IPC score, which measures the quality of the planners. By providing the optimal upper bound of the domains, we contribute to the stabilization of the IPC score evaluation metric for these domains.

Most well-known and traditional online planners for probabilistic planning are in some way based on Monte-Carlo Tree Search. SOGBOFA, symbolic online gradient-based optimization for factored action MDPs, offers a new perspective on this: it constructs a function graph encoding the expected reward for a given input state using independence assumptions for states and actions. On this function, they use gradient ascent to perform a symbolic search optimizing the actions for the current state. This unique approach to probabilistic planning has shown very strong results and even more potential. In this thesis, we attempt to integrate the new ideas SOGBOFA presents into the traditionally successful Trial-based Heuristic Tree Search framework. Specifically, we design and evaluate two heuristics based on the aforementioned graph and its Q value estimations, but also the search using gradient ascent. We implement and evaluate these heuristics in the Prost planner, along with a version of the current standalone planner.

In this thesis, we consider cyclical dependencies between landmarks for cost-optimal planning. Landmarks denote properties that must hold at least once in all plans. However, if the orderings between them induce cyclical dependencies, one of the landmarks in each cycle must be achieved an additional time. We propose the generalized cycle-covering heuristic which considers this in addition to the cost for achieving all landmarks once.

Our research is motivated by recent applications of cycle-covering in the Freecell and logistics domain where it yields near-optimal results. We carry it over to domain-independent planning using a linear programming approach. The relaxed version of a minimum hitting set problem for the landmarks is enhanced by constraints concerned with cyclical dependencies between them. In theory, this approach surpasses a heuristic that only considers landmarks.

We apply the cycle-covering heuristic in practice where its theoretical dominance is confirmed; Many planning tasks contain cyclical dependencies and considering them affects the heuristic estimates favorably. However, the number of tasks solved using the improved heuristic is virtually unaffected. We still believe that considering this feature of landmarks offers great potential for future work.

Potential heuristics are a class of heuristics used in classical planning to guide a search algorithm towards a goal state. Most of the existing research on potential heuristics is focused on finding heuristics that are admissible, such that they can be used by an algorithm such as A* to arrive at an optimal solution. In this thesis, we focus on the computation of potential heuristics for satisficing planning, where plan optimality is not required and the objective is to find any solution. Specifically, our focus is on the computation of potential heuristics that are descending and dead-end avoiding (DDA), since these prop- erties guarantee favorable search behavior when used with greedy search algorithms such as hillclimbing. We formally prove that the computation of DDA heuristics is a PSPACE-complete problem and propose several approximation algorithms. Our evaluation shows that the resulting heuristics are competitive with established approaches such as Pattern Databases in terms of heuristic quality but suffer from several performance bottlenecks.

Most automated planners use heuristic search to solve the tasks. Usually, the planners get as input a lifted representation of the task in PDDL, a compact formalism describing the task using a fragment of first-order logic. The planners then transform this task description into a grounded representation where the task is described in propositional logic. This new grounded format can be exponentially larger than the lifted one, but many planners use this grounded representation because it is easier to implement and reason about.

However, sometimes this transformation between lifted and grounded representations is not tractable. When this is the case, there is not much that planners based on heuristic search can do. Since this transformation is a required preprocess, when this fails, the whole planner fails.

To solve the grounding problem, we introduce new methods to deal with tasks that cannot be grounded. Our work aims to find good ways to perform heuristic search while using a lifted representation of planning problems. We use the point-of-view of planning as a database progression problem and borrow solutions from the areas of relational algebra and database theory.

Our theoretical and empirical results are motivating: several instances that were never solved by any planner in the literature are now solved by our new lifted planner. For example, our planner can solve the challenging Organic Synthesis domain using a breadth-first search, while state-of-the-art planners cannot solve more than 60% of the instances. Furthermore, our results offer a new perspective and a deep theoretical study of lifted representations for planning tasks.

The generation of independently verifiable proofs for the unsolvability of planning tasks using different heuristics, including linear Merge-and-Shrink heuristics, is possible by usage of a proof system framework. Proof generation in the case of non-linear Merge-and-Shrink heuristic, however, is currently not supported. This is due to the lack of a suitable state set representation formalism that allows to compactly represent states mapped to a certain value in the belonging Merge-and-Shrink representation (MSR). In this thesis, we overcome this shortcoming using Sentential Decision Diagrams (SDDs) as set representations. We describe an algorithm that constructs the desired SDD from the MSR, and show that efficient proof verification is possible with SDDs as representation formalism. Aditionally, we use a proof of concept implementation to analyze the overhead occurred by the proof generation functionality and the runtime of the proof verification.

The operator-counting framework is a framework in classical planning for heuristics that are based on linear programming. The operator-counting framework covers several kinds of state-of-the-art linear programming heuristics, among them the post-hoc optimization heuristic. In this thesis we will use post-hoc optimization constraints and evaluate them under altered cost functions instead of the original cost function of the planning task. We show that such cost-altered post-hoc optimization constraints are also covered by the operator-counting framework and that it is possible to achieve improved heuristic estimates with them, compared with post-hoc optimization constraints under the original cost function. In our experiments we have not been able to achieve improved problem coverage, as we were not able to find a method for generating favorable cost functions that work well in all domains.

Heuristic forward search is the state-of-the-art approach to solve classical planning problems. On the other hand, bidirectional heuristic search has a lot of potential but was never able to deliver on those expectations in practice. Only recently the near-optimal bidirectional search algorithm (NBS) was introduces by Chen et al. and as the name suggests, NBS expands nearly the optimal number of states to solve any search problem. This is a novel achievement and makes the NBS algorithm a very promising and efficient algorithm in search. With this premise in mind, we raise the question of how applicable NBS is to planning. In this thesis, we inquire this very question by implementing NBS in the state- of-the-art planner Fast-Downward and analyse its performance on the benchmark of the latest international planning competition. We additionally implement fractional meet-in- the-middle and computeWVC to analyse NBS’ performance more thoroughly in regards to the structure of the problem task.

The conducted experiments show that NBS can successfully be applied to planning as it was able to consistently outperform A*. Especially good results were achieved on the domains: blocks, driverlog, floortile-opt11-strips, get-opt14-strips, logistics00, and termes- opt18-strips. Analysing these results, we deduce that the efficiency of forward and backward search depends heavily upon the underlying implicit structure of the transition system which is induced by the problem task. This suggests that bidirectional search is inherently more suited for certain problems. Furthermore, we find that this aptitude for a certain search direction correlates with the domain, thereby providing a powerful analytic tool to a priori derive the effectiveness of certain search approaches.

In conclusion, even without intricate improvements the NBS algorithm is able to compete with A*. It therefore has further potential for future research. Additionally, the underlying transition system of a problem instance is shown to be an important factor which influences the efficiency of certain search approaches. This knowledge could be valuable for devising portfolio planners.

Multiple Sequence Alignment (MSA) is the problem of aligning multiple biological sequences in the evoluationary most plausible way. It can be viewed as a shortest path problem through an n-dimensional lattice. Because of its large branching factor of 2^n − 1, it has found broad attention in the artificial intelligence community. Finding a globally optimal solution for more than a few sequences requires sophisticated heuristics and bounding techniques in order to solve the problem in acceptable time and within memory limitations. In this thesis, we show how existing heuristics fall into the category of combining certain pattern databases. We combine arbitrary pattern collections that can be used as heuristic estimates and apply cost partitioning techniques from classical planning for MSA. We implement two of those heuristics for MSA and compare their estimates to the existing heuristics.

Increasing Cost Tree Search is a promising approach to multi-agent pathfinding problems, but like all approaches it has to deal with a huge number of possible joint paths, growing exponentially with the number of agents. We explore the possibility of reducing this by introducing a value abstraction to the Multi-valued Decision Diagrams used to represent sets of joint paths. To that end we introduce a heat map to heuristically judge how collisionprone agent positions are and present how to use and possible refine abstract positions in order to still find valid paths.

Estimating cheapest plan costs with the help of network flows is an established technique. Plans and network flows are already very similar, however network flows can differ from plans in the presence of cycles. If a transition system contains cycles, flows might be composed of multiple disconnected parts. This discrepancy can make the cheapest plan estimation worse. One idea to get rid of the cycles works by introducing time steps. For every time step the states of a transition system are copied. Transitions will be changed, so that they connect states only with states of the next time step, which ensures that there are no cycles. It turned out, that by applying this idea to multiple transitions systems, network flows of the individual transition systems can be synchronized via the time steps to get a new kind of heuristic, that will also be discussed in this thesis.

Probabilistic planning is a research field that has become popular in the early 1990s. It aims at finding an optimal policy which maximizes the outcome of applying actions to states in an environment that feature unpredictable events. Such environments can consist of a large number of states and actions which make finding an optimal policy intractable using classical methods. Using a heuristic function for a guided search allows for tackling such problems. Designing a domain-independent heuristic function requires complex algorithms which may be expensive when it comes to time and memory consumption.

In this thesis, we are applying the supervised learning techniques for learning two domain-independent heuristic functions. We use three types of gradient descent methods: stochastic, batch and mini-batch gradient descent and their improved versions using momen- tum, learning decay rate and early stopping. Furthermore, we apply the concept of feature combination in order to better learn the heuristic functions. The learned functions are pro- vided to Prost, a domain-independent probabilistic planner, and benchmarked against the winning algorithms of the International Probabilistic Planning Competition held in 2014. The experiments show that learning an offline heuristic improves the overall score of the search for some of the domains used in aforementioned competition.

The merge-and-shrink heuristic is a state-of-the-art admissible heuristic that is often used for optimal planning. Recent studies showed that the merge strategy is an important factor for the performance of the merge-and-shrink algorithm. There are many different merge strategies and improvements for merge strategies described in the literature. One out of these merge strategies is MIASM by Fan et al. MIASM tries to merge transition systems that produce unnecessary states in their product which can be pruned. Another merge strategy is the symmetry-based merge-and-shrink framework by Sievers et al. This strategy tries to merge transition systems that cause factored symmetries in their product. This strategy can be combined with other merge strategies and it often improves the performance for many merge strategy. However, the current combination of MIASM with factored symmetries performs worse than MIASM. We implement a different combination of MIASM that uses factored symmetries during the subset search of MIASM. Our experimental evaluation shows that our new combination of MIASM with factored symmetries solves more tasks than the existing MIASM and the previously implemented combination of MIASM with factored symmetries. We also evaluate different combinations of existing merge strategies and find combinations that perform better than their basic version that were not evaluated before.

Tree Cache is a pathfinding algorithm that selects one vertex as a root and constructs a tree with cheapest paths to all other vertices. A path is found by traversing up the tree from both the start and goal vertices to the root and concatenating the two parts. This is fast, but as all paths constructed this way pass through the root vertex they can be highly suboptimal.

To improve this algorithm, we consider two simple approaches. The first is to construct multiple trees, and save the distance to each root in each vertex. To find a path, the algorithm first selects the root with the lowest total distance. The second approach is to remove redundant vertices, i.e. vertices that are between the root and the lowest common ancestor (LCA) of the start and goal vertices. The performance and space requirements of the resulting algorithm are then compared to the conceptually similar hub labels and differential heuristics.

Greedy Best-First Search (GBFS) is a prominent search algorithm to find solutions for planning tasks. GBFS chooses nodes for further expansion based on a distance-to-goal estimator, the heuristic. This makes GBFS highly dependent on the quality of the heuristic. Heuristics often face the problem of producing Uninformed Heuristic Regions (UHRs). GBFS additionally suffers the possibility of simultaneously expanding nodes in multiple UHRs. In this thesis we change the heuristic approach in UHRs. The heuristic was unable to guide the search and so we try to expand novel states to escape the UHRs. The novelty measures how “new” a state is in the search. The result is a combination of heuristic and novelty guided search, which is indeed able to escape UHRs quicker and solve more problems in reasonable time.

In classical AI planning, the state explosion problem is a reoccurring subject: although the problem descriptions are compact, often a huge number of states needs to be considered. One way to tackle this problem is to use static pruning methods which reduce the number of variables and operators in the problem description before planning.

In this work, we discuss the properties and limitations of three existing static pruning techniques with a focus on satisficing planning. We analyse these pruning techniques and their combinations, and identify synergy effects between them and the domains and problem structures in which they occur. We implement the three methods into an existing propositional planner, and evaluate the performance of different configurations and combinations in a set of experiments on IPC benchmarks. We observe that static pruning techniques can increase the number of solved problems, and that the synergy effects of the combinations also occur on IPC benchmarks, although they do not lead to a major performance increase.

The goal of classical domain-independent planning is to find a sequence of actions which lead from a given initial state to a goal state that satisfies some goal criteria. Most planning systems use heuristic search algorithms to find such a sequence of actions. A critical part of heuristic search is the heuristic function. In order to find a sequence of actions from an initial state to a goal state efficiently this heuristic function has to guide the search towards the goal. It is difficult to create such an efficient heuristic function. Arfaee et al. show that it is possible to improve a given heuristic function by applying machine learning techniques on a single domain in the context of heuristic search. To achieve this improvement of the heuristic function, they propose a bootstrap learning approach which subsequently improves the heuristic function.

In this thesis we will introduce a technique to learn heuristic functions that can be used in classical domain-independent planning based on the bootstrap-learning approach introduced by Arfaee et al. In order to evaluate the performance of the learned heuristic functions, we have implemented a learning algorithm for the Fast Downward planning system. The experiments have shown that a learned heuristic function generally decreases the number of explored states compared to blind-search . The total time to solve a single problem increases because the heuristic function has to be learned before it can be applied.

Essential for the estimation of the performance of an algorithm in satisficing planning is its ability to solve benchmark problems. Those results can not be compared directly as they originate from different implementations and different machines. We implemented some of the most promising algorithms for greedy best-first search, published in the last years, and evaluated them on the same set of benchmarks. All algorithms are either based on randomised search, localised search or a combination of both. Our evaluation proves the potential of those algorithms.

Heuristic search with admissible heuristics is the leading approach to cost-optimal, domain-independent planning. Pattern database heuristics - a type of abstraction heuristics - are state-of-the-art admissible heuristics. Two recent pattern database heuristics are the iPDB heuristic by Haslum et al. and the PhO heuristic by Pommerening et al.

The iPDB procedure performs a hill climbing search in the space of pattern collections and evaluates selected patterns using the canonical heuristic. We apply different techniques to the iPDB procedure, improving its hill climbing algorithm as well as the quality of the resulting heuristic. The second recent heuristic - the PhO heuristic - obtains strong heuristic values through linear programming. We present different techniques to influence and improve on the PhO heuristic.

We evaluate the modified iPDB and PhO heuristics on the IPC benchmark suite and show that these abstraction heuristics can compete with other state-of-the-art heuristics in cost-optimal, domain-independent planning.

Greedy best-first search (GBFS) is a prominent search algorithm for satisficing planning - finding good enough solutions to a planning task in reasonable time. GBFS selects the next node to consider based on the most promising node estimated by a heuristic function. However, this behaviour makes GBFS heavily depend on the quality of the heuristic estimator. Inaccurate heuristics can lead GBFS into regions far away from a goal. Additionally, if the heuristic ranks several nodes the same, GBFS has no information on which node it shall follow. Diverse best-first search (DBFS) is a new algorithm by Imai and Kishimoto [2011] which has a local search component to emphasis exploitation. To enable exploration, DBFS deploys probabilities to select the next node.

In two problem domains, we analyse GBFS' search behaviour and present theoretical results. We evaluate these results empirically and compare DBFS and GBFS on constructed as well as on provided problem instances.

State-of-the-art planning systems use a variety of control knowledge in order to enhance the performance of heuristic search. Unfortunately most forms of control knowledge use a specific formalism which makes them hard to combine. There have been several approaches which describe control knowledge in Linear Temporal Logic (LTL). We build upon this work and propose a general framework for encoding control knowledge in LTL formulas. The framework includes a criterion that any LTL formula used in it must fulfill in order to preserve optimal plans when used for pruning the search space; this way the validity of new LTL formulas describing control knowledge can be checked. The framework is implemented on top of the Fast Downward planning system and is tested with a pruning technique called Unnecessary Action Application, which detects if a previously applied action achieved no useful progress.

Landmarks are known to be useable for powerful heuristics for informed search. In this thesis, we explain and evaluate a novel algorithm to find ordered landmarks of delete free tasks by intersecting solutions in the relaxation. The proposed algorithm efficiently finds landmarks and natural orders of delete free tasks, such as delete relaxations or Pi-m compilations.

Planning as heuristic search is the prevalent technique to solve planning problems of any kind of domains. Heuristics estimate distances to goal states in order to guide a search through large state spaces. However, this guidance is sometimes moderate, since still a lot of states lie on plateaus of equally prioritized states in the search space topology. Additional techniques that ignore or prefer some actions for solving a problem are successful to support the search in such situations. Nevertheless, some action pruning techniques lead to incomplete searches.

We propose an under-approximation refinement framework for adding actions to under-approximations of planning tasks during a search in order to find a plan. For this framework, we develop a refinement strategy. Starting a search on an initial under-approximation of a planning task, the strategy adds actions determined at states close to a goal, whenever the search does not progress towards a goal, until a plan is found. Key elements of this strategy consider helpful actions and relaxed plans for refinements. We have implemented the under-approximation refinement framework into the greedy best first search algorithm. Our results show considerable speedups for many classical planning problems. Moreover, we are able to plan with fewer actions than standard greedy best first search.

The main approach for classical planning is heuristic search. Many cost heuristics are based on the delete relaxation. The optimal heuristic of a delete free planning problem is called h + . This thesis explores two new ways to compute h + . Both approaches use factored planning, which decomposes the original planning problem to work on each subproblem separately. The algorithm reuses the subsolutions and combines them to a global solution.

The two algorithms are used to compute a cost heuristic for an A* search. As both approaches compute the optimal heuristic for delete free planning tasks, the algorithms can also be used to find a solution for relaxed planning tasks.

Multi-Agent-Path-Finding (MAPF) is a common problem in robotics and memory management. Pebbles in Motion is an implementation of a problem solver for MAPF in polynomial time, based on a work by Daniel Kornhauser from 1984. Recently a lot of research papers have been published on MAPF in the research community of Artificial Intelligence, but the work by Kornhauser seems hardly to be taken into account. We assumed that this might be related to the fact that said paper was more mathematically and hardly describing algorithms intuitively. This work aims at filling this gap, by providing an easy understandable approach of implementation steps for programmers and a new detailed description for researchers in Computer Science.

Bachelor's theses

Fast Downward is a classical planner using heuristical search. The planner uses many advanced planning techniques that are not easy to teach, since they usually rely on complex data structures. To introduce planning techniques to the user an interactive application is created. This application uses an illustrative example to showcase planning techniques: Blocksworld

Blocksworld is an easy understandable planning problem which allows a simple representation of a state space. It is implemented in the Unreal Engine and provides an interface to the Fast Downward planner. Users can explore a state space themselves or have Fast Downward generate plans for them. The concept of heuristics as well as the state space are explained and made accessible to the user. The user experiences how the planner explores a state space and which techniques the planner uses.

This thesis is about implementing Jussi Rintanen’s algorithm for schematic invariants. The algo- rithm is implemented in the planning tool Fast Downward and refers to Rintanen’s paper Schematic Invariants by Reduction to Ground Invariants. The thesis describes all necessary definitions to under- stand the algorithm and draws a comparison between the original task and a reduced task in terms of runtime and number of grounded actions.

Planning is a field of Artificial Intelligence. Planners are used to find a sequence of actions, to get from the initial state to a goal state. Many planning algorithms use heuristics, which allow the planner to focus on more promising paths. Pattern database heuristics allow us to construct such a heuristic, by solving a simplified version of the problem, and saving the associated costs in a pattern database. These pattern databases can be computed and stored by using symbolic data structures.

In this paper we will look at how pattern databases using symbolic data structures using binary decision diagrams and algebraic decision diagrams can be implemented. We will extend fast down- ward (Helmert [2006]) with it, and compare the performance of this implementation with the already implemented explicit pattern database.

In the field of automated planning and scheduling, a planning task is essentially a state space which can be defined rigorously using one of several different formalisms (e.g. STRIPS, SAS+, PDDL etc.). A planning algorithm tries to determine a sequence of actions that lead to a goal state for a given planning task. In recent years, attempts have been made to group certain planners together into so called planner portfolios, to try and leverage their effectiveness on different specific problem classes. In our project, we create an online planner which in contrast to its offline counterparts, makes use of task specific information when allocating a planner to a task. One idea that has recently gained interest, is to apply machine learning methods to planner portfolios.

In previous work such as Delfi (Katz et al., 2018; Sievers et al., 2019a) supervised learning techniques were used, which made it necessary to train multiple networks to be able to attempt multiple, potentially different, planners for a given task. The reason for this being that, if we used the same network, the output would always be the same, as the input to the network would remain unchanged. In this project we make use of techniques from rein- forcement learning such as DQNs (Mnih et al., 2013). Using RL approaches such as DQNs, allows us to extend the input to the network to include information on things, such as which planners were previously attempted and for how long. As a result multiple attempts can be made after only having trained a single network.

Unfortunately the results show that current reinforcement learning agents are, amongst other reasons, too sample inefficient to be able to deliver viable results given the size of the currently available data sets.

Planning tasks are important and difficult problems in computer science. A widely used approach is the use of delete relaxation heuristics to which the additive and FF heuristic belong. Those two heuristics use a graph in their calculation, which only has to be constructed once for a planning task but then can be used repeatedly. To solve such a problem efficiently it is important that the calculation of the heuristics are fast. In this thesis the idea to achieve a faster calculation is to combine redundant parts of the graph when building it to reduce the number of edges and therefore speed up the calculation. Here the reduction of the redundancies is done for each action within a planning task individually, but further ideas to simplify over all actions are also discussed.

Monte Carlo search methods are widely known, mostly for their success in game domains, although they are also applied to many non-game domains. In previous work done by Schulte and Keller, it was established that best-first searches could adapt to the action selection functionality which make Monte Carlo methods so formidable. In practice however, the trial-based best first search, without exploration, was shown to be slightly slower than its explicit open list counterpart. In this thesis we examine the non-trial and trial-based searches and how they can address the exploitation exploration dilemma. Lastly, we will see how trial-based BFS can rectify a slower search by allowing occasional random action selection, by comparing it to regular open list searches in a line of experiments.

Sudoku has become one of the world’s most popular logic puzzles, arousing interest in the general public and the science community. Although the rules of Sudoku may seem simple, they allow for nearly countless puzzle instances, some of which are very hard to solve. SAT-solvers have proven to be a suitable option to solve Sudokus automatically. However, they demand the puzzles to be encoded as logical formulae in Conjunctive Normal Form. In earlier work, such encodings have been successfully demonstrated for original Sudoku Puzzles. In this thesis, we present encodings for rather unconventional Sudoku Variants, developed by the puzzle community to create even more challenging solving experiences. Furthermore, we demonstrate how Pseudo-Boolean Constraints can be utilized to encode Sudoku Variants that follow rules involving sums. To implement an encoding of Pseudo-Boolean Constraints, we use Binary Decision Diagrams and Adder Networks and study how they compare to each other.

In optimal classical planning, informed search algorithms like A* need admissible heuristics to find optimal solutions. Counterexample-guided abstraction refinement (CEGAR) is a method used to generate abstractions that yield suitable abstraction heuristics iteratively. In this thesis, we propose a class of CEGAR algorithms for the generation of domain abstractions, which are a class of abstractions that rank in between projections and Cartesian abstractions regarding the grade of refinement they allow. As no known algorithm constructs domain abstractions, we show that our algorithm is competitive with CEGAR algorithms that generate one projection or Cartesian abstraction.

This thesis will look at Single-Player Chess as a planning domain using two approaches: one where we look at how we can encode the Single-Player Chess problem as a domain-independent (general-purpose AI) approach and one where we encode the problem as a domain-specific solver. Lastly, we will compare the two approaches by doing some experiments and comparing the results of the two approaches. Both the domain-independent implementation and the domain-specific implementation differ from traditional chess engines because the task of the agent is not to find the best move for a given position and colour, but the agent’s task is to check if a given chess problem has a solution or not. If the agent can find a solution, the given chess puzzle is valid. The results of both approaches were measured in experiments, and we found out that the domain-independent implementation is too slow and that the domain-specific implementation, on the other hand, can solve the given puzzles reliably, but it has a memory bottleneck rooted in the search method that was used.

Carcassonne is a tile-based board game with a large state space and a high branching factor and therefore poses a challenge to artificial intelligence. In the past, Monte Carlo Tree Search (MCTS), a search algorithm for sequential decision-making processes, has been shown to find good solutions in large state spaces. MCTS works by iteratively building a game tree according to a tree policy. The profitability of paths within that tree is evaluated using a default policy, which influences in what directions the game tree is expanded. The functionality of these two policies, as well as other factors, can be implemented in many different ways. In consequence, many different variants of MCTS exist. In this thesis, we applied MCTS to the domain of two-player Carcassonne and evaluated different variants in regard to their performance and runtime. We found significant differences in performance for various variable aspects of MCTS and could thereby evaluate a configuration which performs best on the domain of Carcassonne. This variant consistently outperformed an average human player with a feasible runtime.

In general, it is important to verify software as it is prone to error. This also holds for solving tasks in classical planning. So far, plans in general as well as the fact that there is no plan for a given planning task can be proven and independently verified. However, no such proof for the optimality of a solution of a task exists. Our aim is to introduce two methods with which optimality can be proven and independently verified. We first reduce unit cost tasks to unsolvable tasks, which enables us to make use of the already existing certificates for unsolvability. In a second approach, we propose a proof system for optimality, which enables us to infer that the determined cost of a task is optimal. This permits the direct generation of optimality certificates.

Pattern databases are one of the most powerful heuristics in classical planning. They evaluate the perfect cost for a simplified sub-problem. The post-hoc optimization heuristic is a technique on how to optimally combine a set of pattern databases. In this thesis, we will adapt the post-hoc optimization heuristic for the sliding tile puzzle. The sliding tile puzzle serves as a benchmark to compare the post-hoc optimization heuristic to already established methods, which also deal with the combining of pattern databases. We will then show how the post-hoc optimization heuristic is an improvement over the already established methods.

In this thesis, we generate landmarks for a logistics-specific task. Landmarks are actions that need to occur at least once in every plan. A landmark graph denotes a structure with landmarks and their edges called orderings. If there are cycles in a landmark graph, one of those landmarks needs to be achieved at least twice for every cycle. The generation of the logistics-specific landmarks and their orderings calculate the cyclic landmark heuristic. The task is to pick up on related work, the evaluation of the cyclic landmark heuristic. We compare the generation of landmark graphs from a domain-independent landmark generator to a domain-specific landmark generator, the latter being the focus. We aim to bridge the gap between domain-specific and domain-independent landmark generators. In this thesis, we compare one domain-specific approach for the logistics domain with results from a domain- independent landmark generator. We devise a unit to pre-process data for other domain- specific tasks as well. We will show that specificity is better suited than independence.

Lineare Programmierung ist eine mathematische Modellierungstechnik, bei der eine lineare Funktion, unter der Berücksichtigung verschiedenen Beschränkungen, maximiert oder minimiert werden soll. Diese Technik ist besonders nützlich, falls Entscheidungen für Optimierungsprobleme getroffen werden sollen. Ziel dieser Arbeit war es ein Tool für das Spiel Factory Town zu entwickeln, mithilfe man Optimierungsanfragen bearbeiten kann. Dabei ist es möglich wahlweise zwischen diversen Fragestellungen zu wählen und anhand von LP-\ IP-Solvern diese zu beantworten. Zudem wurden die mathematischen Formulierungen, sowie die Unterschiede beider Methoden angegangen. Schlussendlich unterstrichen die generierten Resultate, dass LP Lösungen mindestens genauso gut oder sogar besser seien als die Lösungen eines IP.

Symbolic search is an important approach to classical planning. Symbolic search uses search algorithms that process sets of states at a time. For this we need states to be represented by a compact data structure called knowledge compilations. Merge-and-shrink representations come a different field of planning, where they have been used to derive heuristic functions for state-space search. More generally they represent functions that map variable assignments to a set of values, as such we can regard them as a data structure we will call Factored Mappings. In this thesis, we will investigate Factored Mappings (FMs) as a knowledge compilation language with the hope of using them for symbolic search. We will analyse the necessary transformations and queries for FMs, by defining the needed operations and a canonical representation of FMs, and showing that they run in polynomial time. We will then show that it is possible to use Factored Mappings as a knowledge compilation for symbolic search by defining a symbolic search algorithm for a finite-domain plannings task that works with FMs.

Version control systems use a graph data structure to track revisions of files. Those graphs are mutated with various commands by the respective version control system. The goal of this thesis is to formally define a model of a subset of Git commands which mutate the revision graph, and to model those mutations as a planning task in the Planning Domain Definition Language. Multiple ways to model those graphs will be explored and those models will be compared by testing them using a set of planners.

Pattern Databases are admissible abstraction heuristics for classical planning. In this thesis we are introducing the Boosting processes, which consists of enlarging the pattern of a Pattern Database P, calculating a more informed Pattern Database P' and then min-compress P' to the size of P resulting in a compressed and still admissible Pattern Database P''. We design and implement two boosting algorithms, Hillclimbing and Randomwalk.

We combine pattern database heuristics using five different cost partitioning methods. The experiments compare computing cost partitionings over regular and boosted pattern databases. The experiments, performed on IPC (optimal track) tasks, show promising results which increased the coverage (number of solved tasks) by 9 for canonical cost partitioning using our Randomwalk boosting variant.

One dimensional potential heuristics assign a numerical value, the potential, to each fact of a classical planning problem. The heuristic value of a state is the sum over the poten- tials belonging to the facts contained in the state. Fišer et al. (2020) recently proposed to strengthen potential heuristics utilizing mutexes and disambiguations. In this thesis, we embed the same enhancements in the planning system Fast Downward. The experi- mental evaluation shows that the strengthened potential heuristics are a refinement, but too computationally expensive to solve more problems than the non-strengthened potential heuristics.

The potentials are obtained with a Linear Program. Fišer et al. (2020) introduced an additional constraint on the initial state and we propose additional constraints on random states. The additional constraints improve the amount of solved problems by up to 5%.

This thesis discusses the PINCH heuristic, a specific implementation of the additive heuristic. PINCH intends to combine the strengths of existing implementations of the additive heuristic. The goal of this thesis is to really dig into the PINCH heuristic. I want to provide the most accessible resource for understanding PINCH and I want to analyze the performance of PINCH by comparing it to the algorithm on which it is based, Generalized Dijkstra.

Suboptimal search algorithms can offer attractive benefits compared to optimal search, namely increased coverage of larger search problems and quicker search times. Improving on such algorithms, such as reducing costs further towards optimal solutions and reducing the number of node expansions, is therefore a compelling area for further research. This paper explores the utility and scalability of recently developed priority functions, XDP, XUP, and PWXDP, and the Improved Optimistic Search algorithm, compared to Weighted A*, in the Fast Downward planner. Analyses focus on the cost, total time, coverage, and node expansion parameters, with experimental evidence suggesting preferable performance if strict optimality is not desired. The implementation of priorityb functions in eager best-first search showed marked improvements compared to A* search on coverage, total time, and number of expansions, without significant cost penalties. Following previous suboptimal search research, experimental evidence even seems to indicate that these cost penalties do not reach the designated bound, even in larger search spaces.

In the Automated Planning field, algorithms and systems are developed for exploring state spaces and ultimately finding an action sequence leading from a task’s initial state to its goal. Such planning systems may sometimes show unexpected behavior, caused by a planning task or a bug in the planner itself. Generally speaking, finding the source of a bug tends to be easier when the cause can be isolated or simplified. In this thesis, we tackle this problem by making PDDL and SAS+ tasks smaller while ensuring they still invoke a certain characteristic when executed with a planner. We implement a system that successively removes elements, such as objects, from a task and checks whether the transformed task still fails on the planner. Elements are removed in a syntactically consistent way, however, no semantic integrity is enforced. Our system’s design is centered around the Fast Downward Planning System, as we re-use some of its translator modules and all test runs are performed with Fast Downward. At the core of our system, first-choice hill-climbing is used for optimization. Our “minimizer” takes (1) a failing planner execution command, (2) a description of the failing characteristic and (3) the type of element to be deleted as arguments. We evaluate our system’s functionality on the basis of three use-cases. In our most successful test runs, (1) a SAS+ task with initially 1536 operators and 184 variables is reduced to 2 operators and 2 variables and (2)a PDDL task with initially 46 actions, 62 objects and 29 predicate symbols is reduced to 2 actions, 6 objects and 4 predicates.

Fast Downward is a classical planning system based on heuristic search. Its successor generator is an efficient and intelligent tool to process state spaces and generate their successor states. In this thesis we implement different successor generators in the Fast Downward planning system and compare them against each other. Apart from the given fast downward successor generator we implement four other successor generators: a naive successor generator, one based on the marking of delete relaxed heuristics, one based on the PSVN planning system and one based on watched literals as used in modern SAT solvers. These successor generators are tested in a variety of different planning benchmarks to see how well they compete against each other. We verified that there is a trade-off between precomputation and faster successor generation and showed that all of the implemented successor generators have a use case and it is advisable to switch to a successor generator that fits the style of the planning task.

Verifying whether a planning algorithm came to the correct result for a given planning task is easy if a plan is emitted which solves the problem. But if a task is unsolvable most planners just state this fact without any explanation or even proof. In this thesis we present extended versions of the symbolic search algorithms SymPA and symbolic bidirectional uniform-cost search which, if a given planning task is unsolvable, provide certificates which prove unsolvability. We also discuss a concrete implementation of this version of SymPA.

Classical planning is an attractive approach to solving problems because of its generality and its relative ease of use. Domain-specific algorithms are appealing because of their performance, but require a lot of resources to be implemented. In this thesis we evaluate concepts languages as a possible input language for expert domain knowledge into a planning system. We also explore mixed integer programming as a way to use this knowledge to improve search efficiency and to help the user find and refine useful domain knowledge.

Classical Planning is a branch of artificial intelligence that studies single agent, static, deterministic, fully observable, discrete search problems. A common challenge in this field is the explosion of states to be considered when searching for the goal. One technique that has been developed to mitigate this is Strong Stubborn Set based pruning, where on each state expansion, the considered successors are restricted to Strong Stubborn Sets, which exploit the properties of independent operators to cut down the tree or graph search. We adopt the definitions of the theory of Strong Stubborn Sets from the SAS+ setting to transition systems and validate a central theorem about the correctness of Strong Stubborn Set based pruning for transition systems in the interactive theorem prover Isabelle/HOL.

Ein wichtiges Feld in der Wissenschaft der künstliche Intelligenz sind Planungsprobleme. Man hat das Ziel, eine künstliche intelligente Maschine zu bauen, die mit so vielen ver- schiedenen Probleme umgehen und zuverlässig lösen kann, indem sie ein optimaler Plan herstellt.

Der Trial-based Heuristic Tree Search(THTS) ist ein mächtiges Werkzeug um Multi-Armed- Bandit-ähnliche Probleme, Marcow Decsision Processe mit verändernden Rewards, zu lösen. Beim momentanen THTS können explorierte gefundene gute Rewards auf Grund von der grossen Anzahl der Rewards nicht beachtet werden. Ebenso können beim explorieren schlech- te Rewards, gute Knoten im Suchbaum, verschlechtern. Diese Arbeit führt eine Methodik ein, die von der stückweise stationären MABs Problematik stammt, um den THTS weiter zu optimieren.

Abstractions are a simple yet powerful method of creating a heuristic to solve classical planning problems optimally. In this thesis we make use of Cartesian abstractions generated with Counterexample-Guided Abstraction Refinement (CEGAR). This method refines abstractions incrementally by finding flaws and then resolving them until the abstraction is sufficiently evolved. The goal of this thesis is to implement and evaluate algorithms which select solutions of such flaws, in a way which results in the best abstraction (that is, the abstraction which causes the problem to then be solved most efficiently by the planner). We measure the performance of a refinement strategy by running the Fast Downward planner on a problem and measuring how long it takes to generate the abstraction, as well as how many expansions the planner requires to find a goal using the abstraction as a heuristic. We use a suite of various benchmark problems for evaluation, and we perform this experiment for a single abstraction and on abstractions for multiple subtasks. Finally, we attempt to predict which refinement strategy should be used based on parameters of the task, potentially allowing the planner to automatically select the best strategy at runtime.

Heuristic search is a powerful paradigm in classical planning. The information generated by heuristic functions to guide the search towards a goal is a key component of many modern search algorithms. The paper “Using Backwards Generated Goals for Heuristic Planning” by Alcázar et al. proposes a way to make additional use of this information. They take the last actions of a relaxed plan as a basis to generate intermediate goals with a known path to the original goal. A plan is found when the forward search reaches an intermediate goal.

The premise of this thesis is to modify their approach by focusing on a single sequence of intermediate goals. The aim is to improve efficiency while preserving the benefits of backwards goal expansion. We propose different variations of our approach by introducing multiple ways to make decisions concerning the construction of intermediate goals. We evaluate these variations by comparing their performance and illustrate the challenges posed by this approach.

Counterexample-guided abstraction refinement (CEGAR) is a way to incrementally compute abstractions of transition systems. It starts with a coarse abstraction and then iteratively finds an abstract plan, checks where the plan fails in the concrete transition system and refines the abstraction such that the same failure cannot happen in subsequent iterations. As the abstraction grows in size, finding a solution for the abstract system becomes more and more costly. Because the abstraction grows incrementally, however, it is possible to maintain heuristic information about the abstract state space, allowing the use of informed search algorithms like A*. As the quality of the heuristic is crucial to the performance of informed search, the method for maintaining the heuristic has a significant impact on the performance of the abstraction refinement as a whole. In this thesis, we investigate different methods for maintaining the value of the perfect heuristic h* at all times and evaluate their performance.

Pattern Databases are a powerful class of abstraction heuristics which provide admissible path cost estimates by computing exact solution costs for all states of a smaller task. Said task is obtained by abstracting away variables of the original problem. Abstractions with few variables offer weak estimates, while introduction of additional variables is guaranteed to at least double the amount of memory needed for the pattern database. In this thesis, we present a class of algorithms based on counterexample-guided abstraction refinement (CEGAR), which exploit additivity relations of patterns to produce pattern collections from which we can derive heuristics that are both informative and computationally tractable. We show that our algorithms are competitive with already existing pattern generators by comparing their performance on a variety of planning tasks.

We consider the problem of Rubik’s Cube to evaluate modern abstraction heuristics. In order to find feasible abstractions of the enormous state space spanned by Rubik’s Cube, we apply projection in the form of pattern databases, Cartesian abstraction by doing counterexample guided abstraction refinement as well as merge-and-shrink strategies. While previous publications on Cartesian abstractions have not covered applicability for planning tasks with conditional effects, we introduce factorized effect tasks and show that Cartesian abstraction can be applied to them. In order to evaluate the performance of the chosen heuristics, we run experiments on different problem instances of Rubik’s Cube. We compare them by the initial h-value found for all problems and analyze the number of expanded states up to the last f-layer. These criteria provide insights about the informativeness of the considered heuristics. Cartesian Abstraction yields perfect heuristic values for problem instances close to the goal, however it is outperformed by pattern databases for more complex instances. Even though merge-and-shrink is the most general abstraction among the considered, it does not show better performance than the others.

Probabilistic planning expands on classical planning by tying probabilities to the effects of actions. Due to the exponential size of the states, probabilistic planners have to come up with a strong policy in a very limited time. One approach to optimising the policy that can be found in the available time is called metareasoning, a technique aiming to allocate more deliberation time to steps where more time to plan results in an improvement of the policy and less deliberation time to steps where an improvement of the policy with more time to plan is unlikely.

This thesis aims to adapt a recent proposal of a formal metareasoning procedure from Lin. et al. for the search algorithm BRTDP to work with the UCT algorithm in the Prost planner and compare its viability to the current standard and a number of less informed time management methods in order to find a potential improvement to the current uniform deliberation time distribution.

A planner tries to produce a policy that leads to a desired goal given the available range of actions and an initial state. A traditional approach for an algorithm is to use abstraction. In this thesis we implement the algorithm described in the ASAP-UCT paper: Abstraction of State-Action Pairs in UCT by Ankit Anand, Aditya Grover, Mausam and Parag Singla.

The algorithm combines state and state-action abstraction with a UCT-algorithm. We come to the conclusion that the algorithm needs to be improved because the abstraction of action-state often cannot detect a similarity that a reasonable action abstraction could find.

The notion of adding a form of exploration to guide a search has been proven to be an effective method of combating heuristical plateaus and improving the performance of greedy best-first search. The goal of this thesis is to take the same approach and introduce exploration in a bounded suboptimal search problem. Explicit estimation search (EES), established by Thayer and Ruml, consults potentially inadmissible information to determine the search order. Admissible heuristics are then used to guarantee the cost bound. In this work we replace the distance-to-go estimator used in EES with an approach based on the concept of novelty.

Classical domain-independent planning is about finding a sequence of actions which lead from an initial state to a goal state. A popular approach for solving planning problems efficiently is to utilize heuristic functions. A possible heuristic function is the perfect heuristic of a delete relaxed planning problem denoted as h+. Delete relaxation simplifies the planning problem thus making it easier to find a perfect heuristic. However computing h+ is still NP-hard problem.

In this thesis we discuss a promising looking approach to compute h+ in practice. Inspired by the paper from Gnad, Hoffmann and Domshlak about star-shaped planning problems, we implemented the Flow-Cut algorithm. The basic idea behind flow-cut to divide a problem that is unsolvable in practice, into smaller sub problems that can be solved. We further tested the flow-cut algorithm on the domains provided by the International Planning Competition benchmarks, resulting in the following conclusion: Using a divide and conquer approach can successfully be used to solve classical planning problems, however it is not trivial to design such an algorithm to be more efficient than state-of-the-art search algorithm.

This thesis deals with the algorithm presented in the paper "Landmark-based Meta Best-First Search Algorithm: First Parallelization Attempt and Evaluation" by Simon Vernhes, Guillaume Infantes and Vincent Vidal. Their idea was to reconsider the approach to landmarks as a tool in automated planning, but in a markedly different way than previous work had done. Their result is a meta-search algorithm which explores landmark orderings to find a series of subproblems that reliably lead to an effective solution. Any complete planner may be used to solve the subproblems. While the referenced paper also deals with an attempt to effectively parallelize the Landmark-based Meta Best-First Search Algorithm, this thesis is concerned mainly with the sequential implementation and evaluation of the algorithm in the Fast Downward planning system.

Heuristics play an important role in classical planning. Using heuristics during state space search often reduces the time required to find a solution, but constructing heuristics and using them to calculate heuristic values takes time, reducing this benefit. Constructing heuristics and calculating heuristic values as quickly as possible is very important to the effectiveness of a heuristic. In this thesis we introduce methods to bound the construction of merge-and-shrink to reduce its construction time and increase its accuracy for small problems and to bound the heuris- tic calculation of landmark cut to reduce heuristic value calculation time. To evaluate the performance of these depth-bound heuristics we have implemented them in the Fast Down- ward planning system together with three iterative-deepening heuristic search algorithms: iterative-deepening A* search, a new breadth-first iterative-deepening version of A* search and iterative-deepening breadth-first heuristic search.

Greedy best-first search has proven to be a very efficient approach to satisficing planning but can potentially lose some of its effectiveness due to the used heuristic function misleading it to a local minimum or plateau. This is where exploration with additional open lists comes in, to assist greedy best-first search with solving satisficing planning tasks more effectively. Building on the idea of exploration by clustering similar states together as described by Xie et al. [2014], where states are clustered according to heuristic values, we propose in this paper to instead cluster states based on the Hamming distance of the binary representation of states [Hamming, 1950]. The resulting open list maintains k buckets and inserts each given state into the bucket with the smallest average hamming distance between the already clustered states and the new state. Additionally, our open list is capable of reclustering all states periodically with the use of the k-means algorithm. We were able to achieve promising results concerning the amount of expansions necessary to reach a goal state, despite not achieving a higher coverage than fully random exploration due to slow performance. This was caused by the amount of calculations required to identify the most fitting cluster when inserting a new state.

Monte Carlo Tree Search Algorithms are an efficient method of solving probabilistic planning tasks that are modeled by Markov Decision Problems. MCTS uses two policies, a tree policy for iterating through the known part of the decission tree and a default policy to simulate the actions and their reward after leaving the tree. MCTS algorithms have been applied with great success to computer Go. To make the two policies fast many enhancements based on online knowledge have been developed. The goal of All Moves as First enhancements is to improve the quality of a reward estimate in the tree policy. In the context of this thesis the, in the field of computer Go very efficient, α-AMAF, Cutoff-AMAF as well as Rapid Action Value Estimation enhancements are implemented in the probabilistic planner PROST. To obtain a better default policy, Move Average Sampling is implemented into PROST and benchmarked against it’s current default policies.

In classical planning the objective is to find a sequence of applicable actions that lead from the initial state to a goal state. In many cases the given problem can be of enormous size. To deal with these cases, a prominent method is to use heuristic search, which uses a heuristic function to evaluate states and can focus on the most promising ones. In addition to applying heuristics, the search algorithm can apply additional pruning techniques that exclude applicable actions in a state because applying them at a later point in the path would result in a path consisting of the same actions but in a different order. The question remains as to how these actions can be selected without generating too much additional work to still be useful for the overall search. In this thesis we implement and evaluate the partition-based path pruning method, proposed by Nissim et al. [1], which tries to decompose the set of all actions into partitions. Based on this decomposition, actions can be pruned with very little additional information. The partition-based pruning method guarantees with some alterations to the A* search algorithm to preserve it’s optimality. The evaluation confirms that in several standard planning domains, the pruning method can reduce the size of the explored state space.

Validating real-time systems is an important and complex task which becomes exponentially harder with increasing sizes of systems. Therefore finding an automated approach to check real-time systems for possible errors is crucial. The behaviour of such real-time systems can be modelled with timed automata. This thesis adapts and implements the under-approximation refinement algorithm developed for search based planners proposed by Heusner et al. to find error states in timed automata via the directed model checking approach. The evaluation compares the algorithm to already existing search methods and shows that a basic under-approximation refinement algorithm yields a competitive search method for directed model checking which is both fast and memory efficient. Additionally we illustrate that with the introduction of some minor alterations the proposed under- approximation refinement algorithm can be further improved.

In dieser Arbeit wird versucht eine Heuristik zu lernen. Damit eine Heuristik erlernbar ist, muss sie über Parameter verfügen, die die Heuristik bestimmen. Eine solche Möglichkeit bieten Potential-Heuristiken und ihre Parameter werden Potentiale genannt. Pattern-Databases können mit vergleichsweise wenig Aufwand Eigenschaften eines Zustandsraumes erkennen und können somit eingesetzt werden als Grundlage um Potentiale zu lernen. Diese Arbeit untersucht zwei verschiedene Ansätze zum Erlernen der Potentiale aufgrund der Information aus Pattern-Databases. In Experimenten werden die beiden Ansätze genauer untersucht und schliesslich mit der FF-Heuristik verglichen.

We consider real-time strategy (RTS) games which have temporal and numerical aspects and pose challenges which have to be solved within limited search time. These games are interesting for AI research because they are more complex than board games. Current AI agents cannot consistently defeat average human players, while even the best players make mistakes we think an AI could avoid. In this thesis, we will focus on StarCraft Brood War. We will introduce a formal definition of the model Churchill and Buro proposed for StarCraft. This allows us to focus on Build Order optimization only. We have implemented a base version of the algorithm Churchill and Buro used for their agent. Using the implementation we are able to find solutions for Build Order Problems in StarCraft Brood War.

Auf dem Gebiet der Handlungsplanung stellt die symbolische Suche eine der erfolgversprechendsten angewandten Techniken dar. Um eine symbolische Suche auf endlichen Zustandsräumen zu implementieren bedarf es einer geeigneten Datenstruktur für logische Formeln. Diese Arbeit erprobt die Nutzung von Sentential Decision Diagrams (SDDs) anstelle der gängigen Binary Decision Diagrams (BDDs) zu diesem Zweck. SDDs sind eine Generalisierung von BDDs. Es wird empirisch getestet wie eine Implementierung der symbolischen Suche mit SDDs im FastDownward-Planer sich mit verschiedenen vtrees unterscheidet. Insbesondere wird die Performance von balancierten vtrees, mit welchen die Stärken von SDDs oft gut zur Geltung kommen, mit rechtsseitig linearen vtrees verglichen, bei welchen sich SDDs wie BDDs verhalten.

Die Frage ob es gültige Sudokus - d.h. Sudokus mit nur einer Lösung - gibt, die nur 16 Vorgaben haben, konnte im Dezember 2011 mithilfe einer erschöpfenden Brute-Force-Methode von McGuire et al. verneint werden. Die Schwierigkeit dieser Aufgabe liegt in dem ausufernden Suchraum des Problems und der dadurch entstehenden Erforderlichkeit einer effizienten Beweisidee sowie schnellerer Algorithmen. In dieser Arbeit wird die Beweismethode von McGuire et al. bestätigt werden und für 2 2 × 2 2 und 3 2 × 3 2 Sudokus in C++ implementiert.

Das Finden eines kürzesten Pfades zwischen zwei Punkten ist ein fundamentales Problem in der Graphentheorie. In der Praxis ist es oft wichtig, den Ressourcenverbrauch für das Ermitteln eines solchen Pfades minimal zu halten, was mithilfe einer komprimierten Pfaddatenbank erreicht werden kann. Im Rahmen dieser Arbeit bestimmen wir drei Verfahren, mit denen eine Pfaddatenbank möglichst platzsparend aufgestellt werden kann, und evaluieren die Effektivität dieser Verfahren anhand von Probleminstanzen verschiedener Grösse und Komplexität.

In planning what we want to do is to get from an initial state into a goal state. A state can be described by a finite number of boolean valued variables. If we want to transition from one state to the other we have to apply an action and this, at least in probabilistic planning, leads to a probability distribution over a set of possible successor states. From each transition the agent gains a reward dependent on the current state and his action. In this setting the growth of the number of possible states is exponential with the number of variables. We assume that the value of these variables is determined for each variable independently in a probabilistic fashion. So these variables influence the number of possible successor states in the same way as they did the state space. In consequence it is almost impossible to obtain an optimal amount of reward approaching this problem with a brute force technique. One way to get past this problem is to abstract the problem and then solve a simplified version of the aforementioned. That’s in general the idea proposed by Boutilier and Dearden [1]. They have introduced a method to create an abstraction which depends on the reward formula and the dependencies contained in the problem. With this idea as a basis we’ll create a heuristic for a trial-based heuristic tree search (THTS) algorithm [5] and a standalone planner using the framework PROST (Keller and Eyerich, 2012). These will then be tested on all the domains of the International Probabilistic Planning Competition (IPPC).

In einer Planungsaufgabe geht es darum einen gegebenen Wertezustand durch sequentielles Anwenden von Aktionen in einen Wertezustand zu überführen, welcher geforderte Zieleigenschaften erfüllt. Beim Lösen von Planungsaufgaben zählt Effizienz. Um Zeit und Speicher zu sparen verwenden viele Planer heuristische Suche. Dabei wird mittels einer Heuristik abgeschätzt, welche Aktion als nächstes angewendet werden soll um möglichst schnell in einen gewünschten Zustand zu gelangen.

In dieser Arbeit geht es darum, die von Haslum vorgeschlagene P m -Kompilierung für Planungsaufgaben zu implementieren und die h max -Heuristik auf dem kompilierten Problem gegen die h m -Heuristik auf dem originalen Problem zu testen. Die Implementation geschieht als Ergänzung zum Fast-Downward-Planungssystem. Die Resultate der Tests zeigen, dass mittels der Kompilierung die Zahl der gelösten Probleme erhöht werden kann. Das Lösen eines kompilierten Problems mit der h max -Heuristik geschieht im allgemeinen mit selbiger Informationstiefe schneller als das Lösen des originalen Problems mit der h m -Heuristik. Diesen Zeitgewinn erkauft man sich mit einem höheren Speicherbedarf.

The objective of classical planning is to find a sequence of actions which begins in a given initial state and ends in a state that satisfies a given goal condition. A popular approach to solve classical planning problems is based on heuristic forward search algorithms. In contrast, regression search algorithms apply actions “backwards” in order to find a plan from a goal state to the initial state. Currently, regression search algorithms are somewhat unpopular, as the generation of partial states in a basic regression search often leads to a significant growth of the explored search space. To tackle this problem, state subsumption is a pruning technique that additionally discards newly generated partial states for which a more general partial state has already been explored.

In this thesis, we discuss and evaluate techniques of regression and state subsumption. In order to evaluate their performance, we have implemented a regression search algorithm for the planning system Fast Downward, supporting both a simple subsumption technique as well as a refined subsumption technique using a trie data structure. The experiments have shown that a basic regression search algorithm generally increases the number of explored states compared to uniform-cost forward search. Regression with pruning based on state subsumption with a trie data structure significantly reduces the number of explored states compared to basic regression.

This thesis discusses the Traveling Tournament Problem and how it can be solved with heuristic search. The Traveling Tournament problem is a sports scheduling problem where one tries to find a schedule for a league that meets certain constraints while minimizing the overall distance traveled by the teams in this league. It is hard to solve for leagues with many teams involved since its complexity grows exponentially in the number of teams. The largest instances solved up to date, are instances with leagues of up to 10 teams.

Previous related work has shown that it is a reasonable approach to solve the Traveling Tournament Problem with an IDA*-based tree search. In this thesis I implemented such a search and extended it with several enhancements to examine whether they improve performance of the search. The heuristic that I used in my implementation is the Independent Lower Bound heuristic. It tries to find lower bounds to the traveling costs of each team in the considered league. With my implementation I was able to solve problem instances with up to 8 teams. The results of my evaluation have mostly been consistent with the expected impact of the implemented enhancements on the overall performance.

One huge topic in Artificial Intelligence is the classical planning. It is the process of finding a plan, therefore a sequence of actions that leads from an initial state to a goal state for a specified problem. In problems with a huge amount of states it is very difficult and time consuming to find a plan. There are different pruning methods that attempt to lower the amount of time needed to find a plan by trying to reduce the number of states to explore. In this work we take a closer look at two of these pruning methods. Both of these methods rely on the last action that led to the current state. The first one is the so called tunnel pruning that is a generalisation of the tunnel macros that are used to solve Sokoban problems. The idea is to find actions that allow a tunnel and then prune all actions that are not in the tunnel of this action. The second method is the partition-based path pruning. In this method all actions are distributed into different partitions. These partitions then can be used to prune actions that do not belong to the current partition.

The evaluation of these two pruning methods show, that they can reduce the number of explored states for some problem domains, however the difference between pruned search and normal search gets smaller when we use heuristic functions. It also shows that the two pruning rules effect different problem domains.

Ziel klassischer Handlungsplanung ist es auf eine möglichst effiziente Weise gegebene Planungsprobleme zu lösen. Die Lösung bzw. der Plan eines Planungsproblems ist eine Sequenz von Operatoren mit denen man von einem Anfangszustand in einen Zielzustand gelangt. Um einen Zielzustand gezielter zu finden, verwenden einige Suchalgorithmen eine zusätzliche Information über den Zustandsraum - die Heuristik. Sie schätzt, ausgehend von einem Zustand den Abstand zum Zielzustand. Demnach wäre es ideal, wenn jeder neue besuchte Zustand einen kleineren heuristischen Wert aufweisen würde als der bisher besuchte Zustand. Es gibt allerdings Suchszenarien bei denen die Heuristik nicht weiterhilft um einem Ziel näher zu kommen. Dies ist insbesondere dann der Fall, wenn sich der heuristische Wert von benachbarten Zuständen nicht ändert. Für die gierige Bestensuche würde das bedeuten, dass die Suche auf Plateaus und somit blind verläuft, weil sich dieser Suchalgorithmus ausschliesslich auf die Heuristik stützt. Algorithmen, die die Heuristik als Wegweiser verwenden, gehören zur Klasse der heuristischen Suchalgorithmen.

In dieser Arbeit geht es darum, in Fällen wie den Plateaus trotzdem eine Orientierung im Zustandsraum zu haben, indem Zustände neben der Heuristik einer weiteren Priorisierung unterliegen. Die hier vorgestellte Methode nutzt Abhängigkeiten zwischen Operatoren aus und erweitert die gierige Bestensuche. Wie stark Operatoren voneinander abhängen, betrachten wir anhand eines Abstandsmasses, welches vor der eigentlichen Suche berechnet wird. Die grundlegende Idee ist, Zustände zu bevorzugen, deren Operatoren im Vorfeld voneinander profitierten. Die Heuristik fungiert hierbei erst im Nachhinein als Tie-Breaker, sodass wir einem vielversprechenden Pfad zunächst folgen können, ohne dass uns die Heuristik an einer anderen, weniger vielversprechenden Stelle suchen lässt.

Die Ergebnisse zeigen, dass unser Ansatz in der reinen Suchzeit je nach Heuristik performanter sein kann, als wenn man sich ausschliesslich auf die Heuristik stützt. Bei sehr informationsreichen Heuristiken kann es jedoch passieren, dass die Suche durch unseren Ansatz eher gestört wird. Zudem werden viele Probleme nicht gelöst, weil die Berechnung der Abstände zu zeitaufwändig ist.

In classical planning, heuristic search is a popular approach to solving problems very efficiently. The objective of planning is to find a sequence of actions that can be applied to a given problem and that leads to a goal state. For this purpose, there are many heuristics. They are often a big help if a problem has a solution, but what happens if a problem does not have one? Which heuristics can help proving unsolvability without exploring the whole state space? How efficient are they? Admissible heuristics can be used for this purpose because they never overestimate the distance to a goal state and are therefore able to safely cut off parts of the search space. This makes it potentially easier to prove unsolvability

In this project we developed a problem generator to automatically create unsolvable problem instances and used those generated instances to see how different admissible heuristics perform on them. We used the Japanese puzzle game Sokoban as the first problem because it has a high complexity but is still easy to understand and to imagine for humans. As second problem, we used a logistical problem called NoMystery because unlike Sokoban it is a resource constrained problem and therefore a good supplement to our experiments. Furthermore, unsolvability occurs rather 'naturally' in these two domains and does not seem forced.

Sokoban is a computer game where each level consists of a two-dimensional grid of fields. There are walls as obstacles, moveable boxes and goal fields. The player controls the warehouse worker (Sokoban in Japanese) to push the boxes to the goal fields. The problem is very complex and that is why Sokoban has become a domain in planning.

Phase transitions mark a sudden change in solvability when traversing through the problem space. They occur in the region of hard instances and have been found for many domains. In this thesis we investigate phase transitions in the Sokoban puzzle. For our investigation we generate and evaluate random instances. We declare the defining parameters for Sokoban and measure their influence on the solvability. We show that phase transitions in the solvability of Sokoban can be found and their occurrence is measured. We attempt to unify the parameters of Sokoban to get a prediction on the solvability and hardness of specific instances.

In planning, we address the problem of automatically finding a sequence of actions that leads from a given initial state to a state that satisfies some goal condition. In satisficing planning, our objective is to find plans with preferably low, but not necessarily the lowest possible costs while keeping in mind our limited resources like time or memory. A prominent approach for satisficing planning is based on heuristic search with inadmissible heuristics. However, depending on the applied heuristic, plans found with heuristic search might be of low quality, and hence, improving the quality of such plans is often desirable. In this thesis, we adapt and apply iterative tunneling search with A* (ITSA*) to planning. ITSA* is an algorithm for plan improvement which has been originally proposed by Furcy et al. for search problems. ITSA* intends to search the local space of a given solution path in order to find "short cuts" which allow us to improve our solution. In this thesis, we provide an implementation and systematic evaluation of this algorithm on the standard IPC benchmarks. Our results show that ITSA* also successfully works in the planning area.

In action planning, greedy best-first search (GBFS) is one of the standard techniques if suboptimal plans are accepted. GBFS uses a heuristic function to guide the search towards a goal state. To achieve generality, in domain-independant planning the heuristic function is generated automatically. A well-known problem of GBFS are search plateaus, i.e., regions in the search space where all states have equal heuristic values. In such regions, heuristic search can degenerate to uninformed search. Hence, techniques to escape from such plateaus are desired to improve the efficiency of the search. A recent approach to avoid plateaus is based on diverse best-first search (DBFS) proposed by Imai and Kishimoto. However, this approach relies on several parameters. This thesis presents an implementation of DBFS into the Fast Downward planner. Furthermore, this thesis presents a systematic evaluation of DBFS for several parameter settings, leading to a better understanding of the impact of the parameter choices to the search performance.

Risk is a popular board game where players conquer each other's countries. In this project, I created an AI that plays Risk and is capable of learning. For each decision it makes, it performs a simple search one step ahead, looking at the outcomes of all possible moves it could make, and picks the most beneficial. It judges the desirability of outcomes by a series of parameters, which are modified after each game using the TD(λ)-Algorithm, allowing the AI to learn.

The Canadian Traveler's Problem ( ctp ) is a path finding problem where due to unfavorable weather, some of the roads are impassable. At the beginning, the agent does not know which roads are traversable and which are not. Instead, it can observe the status of roads adjacent to its current location. We consider the stochastic variant of the problem, where the blocking status of a connection is randomly defined with known probabilities. The goal is to find a policy which minimizes the expected travel costs of the agent.

We discuss several properties of the stochastic ctp and present an efficient way to calculate state probabilities. With the aid of these theoretical results, we introduce an uninformed algorithm to find optimal policies.

Finding optimal solutions for general search problems is a challenging task. A powerful approach for solving such problems is based on heuristic search with pattern database heuristics. In this thesis, we present a domain specific solver for the TopSpin Puzzle problem. This solver is based on the above-mentioned pattern database approach. We investigate several pattern databases, and evaluate them on problem instances of different size.

Merge-and-shrink abstractions are a popular approach to generate abstraction heuristics for planning. The computation of merge-and-shrink abstractions relies on a merging and a shrinking strategy. A recently investigated shrinking strategy is based on using bisimulations. Bisimulations are guaranteed to produce perfect heuristics. In this thesis, we investigate an efficient algorithm proposed by Dovier et al. for computing coarsest bisimulations. The algorithm, however, cannot directly be applied to planning and needs some adjustments. We show how this algorithm can be reduced to work with planning problems. In particular, we show how an edge labelled state space can be translated to a state labelled one and what other changes are necessary for the algorithm to be usable for planning problems. This includes a custom data structure to fulfil all requirements to meet the worst case complexity. Furthermore, the implementation will be evaluated on planning problems from the International Planning Competitions. We will see that the resulting algorithm can often not compete with the currently implemented algorithm in Fast Downward. We discuss the reasons why this is the case and propose possible solutions to resolve this issue.

In order to understand an algorithm, it is always helpful to have a visualization that shows step for step what the algorithm is doing. Under this presumption this Bachelor project will explain and visualize two AI techniques, Constraint Satisfaction Processing and SAT Backbones, using the game Gnomine as an example.

CSP techniques build up a network of constraints and infer information by propagating through a single or several constraints at a time, reducing the domain of the variables in the constraint(s). SAT Backbone Computations find literals in a propositional formula, which are true in every model of the given formula.

By showing how to apply these algorithms on the problem of solving a Gnomine game I hope to give a better insight on the nature of how the chosen algorithms work.

Planning as heuristic search is a powerful approach to solve domain-independent planning problems. An important class of heuristics is based on abstractions of the original planning task. However, abstraction heuristics usually come with loss in precision. The contribution of this thesis is the investigation of constrained abstraction heuristics in general, and the application of this concept to pattern database and merge and shrink abstractions in particular. The idea is to use a subclass of mutexes which represent sets of variable-value-pairs so that only one of these pairs can be true at any given time, to regain some of the precision which is lost in the abstraction without increasing its size. By removing states and operators in the abstraction which conflict with such a mutex, the abstraction is refined and hence, the corresponding abstraction heuristic can get more informed. We have implemented the refinements of these heuristics in the Fast Downward planner and evaluated the different approaches using standard IPC benchmarks. The results show that the concept of constrained abstraction heuristics can improve planning as heuristic search in terms of time and coverage.

A permutation problem considers the task where an initial order of objects (ie, an initial mapping of objects to locations) must be reordered into a given goal order by using permutation operators. Permutation operators are 1:1 mappings of the objects from their locations to (possibly other) locations. An example for permutation problems are the wellknown Rubik's Cube and TopSpin Puzzle. Permutation problems have been a research area for a while, and several methods for solving such problems have been proposed in the last two centuries. Most of these methods focused on finding optimal solutions, causing an exponential runtime in the worst case.

In this work, we consider an algorithm for solving permutation problems that has been originally proposed by M. Furst, J. Hopcroft and E. Luks in 1980. This algorithm has been introduced on a theoretical level within a proof for "Testing Membership and Determining the Order of a Group", but has not been implemented and evaluated on practical problems so far. In contrast to the other abovementioned solving algorithms, it only finds suboptimal solutions, but is guaranteed to run in polynomial time. The basic idea is to iteratively reach subgoals, and then to let them fix when we go further to reach the next goals. We have implemented this algorithm and evaluated it on different models, as the Pancake Problem and the TopSpin Puzzle .

Pattern databases (Culberson & Schaeffer, 1998) or PDBs, have been proven very effective in creating admissible Heuristics for single-agent search, such as the A*-algorithm. Haslum et. al proposed, a hill-climbing algorithm can be used to construct the PDBs, using the canonical heuristic. A different approach would be to change action-costs in the pattern-related abstractions, in order to obtain the admissible heuristic. This the so called Cost-Partitioning.

The aim of this project was to implement a cost-partitioning inside the hill-climbing algorithm by Haslum, and compare the results with the standard way which uses the canonical heuristic.

UCT ("upper confidence bounds applied to trees") is a state-of-the-art algorithm for acting under uncertainty, e.g. in probabilistic environments. In the last years it has been very successfully applied in numerous contexts, including two-player board games like Go and Mancala and stochastic single-agent optimization problems such as path planning under uncertainty and probabilistic action planning.

In this project the UCT algorithm was implemented, adapted and evaluated for the classical arcade game "Ms Pac-Man". The thesis introduces Ms Pac-Man and the UCT algorithm, discusses some critical design decisions for developing a strong UCT-based algorithm for playing Ms Pac-Man, and experimentally evaluates the implementation.

Theses @ NLP Group

The NLP Group is continuously looking for students who would like write their bachelor's or master's thesis in the area of natural language processing, possibly with connections to information retrieval and general artificial intelligence.

All thesis topics should be related to the main research directions of the NLP Group, which include  computational argumentation , computational sociolinguistics , and computational explanation . 

Below, we provide a selection of currently available topics. Details of the topics are discussed and shaped jointly in the beginning of the thesis process. Other topics are possible, including own ideas from the student's side, if they go hand in hand with our research interests.

Dealing with students responses and identifying the underlying students’ concept is a core competency that teacher students as well as career changers without teacher training („Quer- und Seiteneinsteigende“) need to practice. A way to better prepare future teachers to interact with their students is to provide them with a virtual classroom environment where they can develop these educational competencies. As a first step towards this goal, we aim to simulate statements that individual students would make in a classroom setting using natural language processing methods. For this, we want to develop a language model that can generate text that reflects the thinking and understanding of students at different levels of science competencies. Existing transcripts from 111 *German* chemistry student reports provide the basis for development and evaluation.

Advisor: Maja Stahl

Metaphors in language are central to explaining concepts. Especially in political opinion pieces, multiple studies have shown usage of metaphorical language affect the views of liberals and conservatives adversely, for example persuading them to change their opinion about a political topic. In this thesis, we would explore how metaphors affect large language models (LLMs) in the same regard.  We would further inspect to what extent metaphors affect the results by post-hoc explanation algorithms. Prior working knowledge of python is required.  Keywords: LLMs, Metaphors, Argumentation, Post-hoc explainibilty 

Advisor: Meghdut Sengupta

Working on the outlined and similar topics involves dealing with state-of-the-art technologies such as neural transformers, contrastive learning, multitask learning, and/or various others. Most topics target the development and empirical evaluation of NLP methods for specific tasks.  

Interested?

Candidates should have very good programming skills (preferably in Python) as well as some experience with machine learning and other AI methods (ideally with NLP). You should be enrolled in one of the computer science programs at Leibniz University Hannover.

In case you are interested in a specific topic , please send a mail to the advisor of that topic, including information about the prior knowledge and experience have:

• What relevant courses did you take?
• What experience with AI development and evaluation do you have?
• What other relevant knowledge do you have?

In case you are unsure about the topic , but interested in writing your thesis with the NLP Group, please send a mail to the head of the group .

The grading of a thesis is based on a weighted grades for two parts: 

• The developed solution to the problem tackled in thesis (45%)
• The written thesis presenting the solution (55%)

The grading of the developed solution takes five criteria into account:

• Difficulty / Complexity.  How difficult was it to develop the solution? How much effort was put into it? Is the complexity justified? ... 
• Technical quality.  Is the design and realization of the solution well-made? Are the experiments systematic and scientifically sound? ...
• Novelty and own ideas.  Does the solution have scientific novelty? Have own ideas been developed and realized in the solution? ...
• Impact / Publishability.  Does the solution improve the state of the art? Are the results worth publishing? Can they be published as is? ...
• Implementation and data.  How easy is it to read and reuse the code? If data has been created, is it well-organized? Are they well-documented? ...

The grading of the written thesis takes six criteria into account:

• Abstract, introduction, and conclusion.  Are problem, solution, and results well-introduced? Are the right conclusions made? Is the whole story told? ... 
• Background and related work.  Are basics well-described and relevant? Is the connection to the thesis clear? Is the state of the art well-discussed? ...
• Approaches and data.  Is the presentation of the developed approaches and data clear, complete, and on the right technical level? ...
• Experiments, evaluation, and discussion.  Are the experiments described systematically? Are the results clearly presented and correctly interpreted? ...
• Form, layout, and style.  Is the structure convincing? Is the writing clear and error-free? Do tables and figures support it? Are citations correct? ...
• Scientific quality.  Does the thesis adhere to scientific standards? Does the presentation follow community principles? …

Past Theses (as of Winter 2022)

• Evaluating Data-Driven Approaches to Improve Word Lists for Measuring Social Bias in Word Embeddings.  Master's thesis, Vinay Kaundinya Ronur Prakash, UPB.
• Audience Aware Counterargument Generation.  Master's thesis. Mahammad Namazov, 2023, UPB.
• Improving Learners’ Arguments by Detecting and Generating Missing Argument Components.  Master's thesis, Nick Düsterhus, 2023, UPB.
• Gender-inclusive Coreference Resolution using Pronoun Preference.  Master's thesis, Jan-Luca Hansel, 2023, UPB.
• Dialect-aware Social Bias Detection using Ensemble and Multi-Task Learning.  Master's thesis, Sai Nikhil Menon, 2022, UPB.
• Counter Argument Generation Using a Knowledge Graph. Master's thesis, Indranil Ghosh, 2022, UPB.
• Domain-aware Text Professionalization using Sequence-to-Sequence Neural Networks.  Bachelor's thesis, Juela Palushi, 2022, UPB.

Past Theses (Summer 2018 – Summer 2022)

• Detection and Mitigation of Subjective Bias in Argumentative Text.  Master's thesis, Sambit Mallick, 2022, UPB.
• Cross-domain analysis of argument quality and its connection to offensive language. Bachelor's thesis, Patrick Bollmann, 2022, UPB.
• Cross-domain Aspect-based Sentiment Analysis with Multimodal Sources . Master's thesis, Pavan Kumar Sheshanarayana, 2022, UPB.
• Comparative Evaluation of Automatic Summarization Techniques for German Court Decision Documents.  Master's thesis, Josua Köhler, 2022, UPB.
• Computational Analysis of Cultural Differences in Learner Argumentation. Master's thesis, Garima Mudgal, 2022, UPB.
• Propaganda Technique Detection Using Connotation Frames.  Master's thesis, Vinaykumar Budanurmath, 2022, UPB.
• Contrastive Argument Summarization using Supervised and Unsupervised Learning.  Master's thesis, Jonas Rieskamp, 2022, UPB.
• Mitigation of Gender Bias in Text using Unsupervised Controllable Rewriting. Master's thesis, Maja Brinkmann, 2021, UPB.
• Assessing Stereotypical Social Biases in Text Sequences using Language.  Master's thesis, Meher Vivek Dheram, 2021, UPB.
• Modeling Context and Argumentativeness of Sentences in Argument Snippet Generation.  Master's thesis, Harsh Shah, 2021, UPB.
• Political Speaker Transfer: Learning to Generate Text in the Styles of Barack Obama and Donald Trump.  Master's thesis, Jonas Bülling, 2021, UPB.
• Quantifying Social Biases in News Articles with Word Embeddings.  Bachelor's thesis, Maximilian Keiff, 2021, UPB.
• Computational Text Professionalization using Neural Sequence-to-Sequence Models.  Master's thesis, Avishek Mishra, 2021, UPB.
• Assessing the Argument Quality of Persuasive Essays using Neural Text Generation .  Master's thesis, Timon Gurcke, 2021, UPB.
• Automatic Conclusion Generation using Neural Networks.  Bachelor's thesis, Torben Zöllner, 2020, UPB.
• Computational Analysis of Metaphors based on Word Embeddings.  Bachelor's thesis,  Simon Krenzler, 2020, UPB. 
• Semi-supervised Cleansing of Web-based Argument Corpora.  Bachelor's thesis, Jonas Dorsch, 2020, BUW.
• Countering Natural Language Arguments using Neural Sequence-to-Sequence Generation.  Master's thesis, Arkajit Dhar, 2020, UPB.
• Snippet Generation for Argument Search.  Bachelor's thesis, Nick Düsterhus, 2019, UPB.
• Argument Quality Assessment in Natural Language using Machine Learning  — bachelor's thesis, Till Werner, 2019, UPB.
• Stance Classification in Argument Search.  Master's thesis, Philipp Heinisch, 2019, UPB.
• Towards a Large-scale Causality Graph.  Bachelor's thesis, Yan Scholten, 2019, UPB.

Past Theses (Summer 2009 – Winter 2017)

• Cross-Domain Mining of Argumentation Strategies using Natural Language Processing .  Master's thesis, 2017, BUW.
• Mining Relevant Arguments at Web Scale .  Master's thesis, 2017, BUW.
• Identifying Controversial Topics in Large-Scale Social Media Data .  Master's thesis, 2016, BUW.
• Efficiency and Effectiveness of Multi-Stage Machine Learning Algorithms for Text Quality Assessment.  Master's thesis, 2013, UPB.
• An Expert System for the Automatic Construction of Information Extraction Pipelines.  Master's thesis, 2012, UPB.
• Efficiency and Effectiveness of Text Classification in Information Extraction Pipelines.  Master's thesis, 2012, UPB.
• Efficient Information Extraction for Creating Use Case Diagrams from Text.  Master's thesis, 2012, UPB.
• Heuristic Search for the Run-time Optimization of Information Extraction Pipelines.  Master's thesis, 2012, UPB.
• Aggregation and Visualization of Market Forecasts.  Bachelor's thesis, 2011, UPB.
• Branch Categorization based on Statistical Analysis of Information Retrieval Results.  Bachelor's thesis 2011, UPB.
• Evaluation of Cooperative Robot Motion Strategies in Simbad.  Bachelor's thesis, 2009, UPB.

LUH: Leibniz University Hannover, UPB: Paderborn University, BUW: Bauhaus-Universität Weimar

Last Change: 07.02.24 Print

• © 2024:  Leibniz University Hannover
• Legal Information
• Data Privacy
• Accessibility Statement
• General Overview Studies
• Theses & Projects
• AI Course at the FEI
• Suggested Software
• Recommended Literature
• Suggested Tools
• Faculty of Electrical Engineering and Computer Science

•   LibGuides
•   A-Z List
•   Help

Artificial Intelligence

• Background Information
• Getting started
• Browse Journals
• Dissertations & Theses
• Datasets and Repositories
• Research Data Management 101
• Scientific Writing
• Find Videos
• Related Topics
• Quick Links
• Ask Us/Contact Us

FIU dissertations

Non-FIU dissertations

Many   universities   provide full-text access to their dissertations via a digital repository.  If you know the title of a particular dissertation or thesis, try doing a Google search.  

Aims to be the best possible resource for finding open access graduate theses and dissertations published around the world with metadata from over 800 colleges, universities, and research institutions. Currently, indexes over 1 million theses and dissertations.

This is a discovery service for open access research theses awarded by European universities.

A union catalog of Canadian theses and dissertations, in both electronic and analog formats, is available through the search interface on this portal.

There are currently more than 90 countries and over 1200 institutions represented. CRL has catalog records for over 800,000 foreign doctoral dissertations.

An international collaborative resource, the NDLTD Union Catalog contains more than one million records of electronic theses and dissertations. Use BASE, the VTLS Visualizer or any of the geographically specific search engines noted lower on their webpage.

Indexes doctoral dissertations and masters' theses in all areas of academic research includes international coverage.

ProQuest Dissertations & Theses global

Related Sites

• << Previous: Browse Journals
• Next: Datasets and Repositories >>
• Last Updated: Apr 4, 2024 8:33 AM
• URL: https://library.fiu.edu/artificial-intelligence

Information

Fiu libraries floorplans, green library, modesto a. maidique campus, hubert library, biscayne bay campus.

Directions: Green Library, MMC

Directions: Hubert Library, BBC

Counting From Zero

Building a liberal arts cs program in the age of ubiquity.

Will AI write your thesis?

This fall, I was honored to serve as Whitman’s convocation speaker. When I agreed to speak, I had no idea what I would talk about, but by time I sat down to write it was obvious what question to ask. It was a fun speech to write, and as I learned more, I changed my conclusion several times. It was a fun speech to deliver, and I appreciate all those who laughed in the right places.

I’m so honored to be here today, standing before all of you as we begin a new academic year. 

This is a time of new beginnings and fresh starts. It’s a time to reflect on where we’ve been and where we’re going. It’s a time to set our sights high and dream big. 

We all have a part to play in shaping our future. Every day, we make choices that will impact our lives and the lives of those around us. I challenge each of you to make choices that will lead to a better future for all of us. 

I also challenge you to be a force for good in the world. There’s so much hurt and pain in the world, but each of us has the power to make a difference. We can start by reaching out to those who are different from us and learning from them. We can stand up for what’s right, even when it’s not easy. And we can show compassion and kindness, even when it’s not popular. 

So let’s make this a year of growth, a year of progress, and a year of making a difference. I can’t wait to see all that you will accomplish.

As you may have guessed, I didn’t write that speech. It was written by a machine learning system called GPT-3, which is available online through the OpenAI API Playground. I prompted GPT-3 to “write a convocation speech,” and I delivered that speech to you exactly as GPT-3 wrote it.

I first learned about GPT-3 last spring when a faculty candidate introduced it in her guest lecture. I’m pleased to say that we hired her. Her name is Jordan Wirfs-Brock, and this spring she will offer courses on Data Science and Human-Computer Interaction.

This summer , I had another encounter with GPT-3 at the meeting of the Computing Research Association. We were asked to consider the educational implications of GitHub Copilot, a tool based on GPT-3 that automatically generates code from natural language descriptions, for example, “sum all the numbers between 1 and 100.” 

After I returned from the meeting, I was assigned to review a research paper addressing that very question. The researchers found that, with minimal human intervention, “Copilot generates mostly unique code that can solve introductory [programming] assignments with human-graded scores ranging from 68% to 95%.” It’s kind of scary to think that a computer system can program about as well as the average introductory programming student. It also has important implications for how we teach introductory programming — not the least, ensuring that students have equal access to such tools and clear expectations for acknowledgment.

And so I was inspired to ask the question, “Will AI write your thesis?”

It’s not a question I would have asked until quite recently. I’m not an AI researcher; if anything I’m an AI skeptic. Like Jordan, my research is in the area of human-computer interaction. One of my goals as an educator is to help students think critically about new technologies.

It’s certainly not a question I would have asked when I was a student. I attended high school when the Internet was still an academic project, and people sent shareware programs on floppy disks through snail  mail. One of the shareware programs I received – as a free bonus with my other purchases – was a program called ELIZA. 

ELIZA is a conversational agent famously designed to simulate a psychotherapist. You type a sentence, then ELIZA responds with another sentence, and you respond with another sentence, and so on.

My conversations with ELIZA are lost to the mists of time. Instead, let me read you an example conversation from Joseph Weizenbaum’s 1966 article explaining how ELIZA works.

To start the conversation, the user types: Men are all alike.

ELIZA responds: IN WHAT WAY?

They’re always bugging us about something or other. CAN YOU THINK OF A SPECIFIC EXAMPLE? Well, my boyfriend made me come here. YOUR BOYFRIEND MADE YOU COME HERE? He says I’m depressed much of the time. I AM SORRY TO HEAR YOU ARE DEPRESSED. It’s true. I am unhappy. DO YOU THINK COMING HERE WILL HELP YOU NOT TO BE UNHAPPY? I need some help, that much seems certain. WHAT WOULD IT MEAN TO YOU IF YOU GOT SOME HELP? Maybe I could learn to get along with my mother. TELL MORE MORE ABOUT YOUR FAMILY. My mother takes care of me. WHO ELSE IN YOUR FAMILY TAKES CARE OF YOU? My father. YOUR FATHER? You are like my father in some ways. WHAT RESEMBLANCE DO YOU SEE? You are not very aggressive but I don’t think you want me to notice that.  WHAT MAKES YOU THINK I AM NOT VERY AGGRESSIVE? You don’t argue with me. WHY DO YOU THINK I DON’T ARGUE WITH YOU? You are afraid of me. DOES IT PLEASE YOU TO BELIEVE I AM AFRAID OF YOU? My father is afraid of everybody. WHAT ELSE COMES TO MIND WHEN YOU THINK OF YOUR FATHER? Bullies. DOES THAT HAVE ANYTHING TO DO WITH THE FACT THAT YOUR BOYFRIEND MADE YOU COME HERE?

Weizenbaum ends the example there, with the appearance of a striking insight.

I read Weizenbaum’s article during my senior year of college, in a course on artificial intelligence. For my final project, I implemented my own version of ELIZA using the programming language PERL. 

What I learned is that ELIZA is a straightforward rule-based system. The user input is tested against a collection of keywords. If a keyword matches, then a corresponding rule is used to transform the user’s input into ELIZA’s output. If no keyword matches, then ELIZA does one of two things. Either it makes a content-free response – for example, GO ON — or it returns to a topic from earlier in the conversation. This can lead to the appearance of striking insights, like the end of the conversation I just read to you.

Weizenbaum wrote that “some subjects have been very hard to convince that ELIZA is not human.” We tend to give conversational partners the benefit of the doubt, as long as they follow certain social norms. When I was in grad school, I learned that social psychology research has confirmed that when computers fill human roles, we tend to treat them as if they were human, even when we know they are not.

Weizenbaum found this phenomenon deeply concerning. One of his goals in writing about ELIZA was to attempt to dispel its “aura of magic.” “Important decisions,” he wrote, “increasingly tend to be made in response to computer output. The ultimately responsible human interpreter of ‘what the machine says’ is, not unlike the correspondent with ELIZA, constantly faced with the need to make credibility judgments. ELIZA shows, if nothing else, how easy it is to create and maintain the illusion of understanding…. A certain danger lurks there.” 

As easy as it is to misattribute intelligence to ELIZA’s responses, ELIZA could not have written the speech that GPT-3 did. In fact, my previous experiences with ELIZA and other text generation systems would have led me to say, “No way: AI could never write your thesis.” 

So what has changed since I was a college student? Three trends beginning with the dawn of computerization in the mid-twentieth century all took off together.

First, the availability of data has increased dramatically. When I started college in 1995, my first computer science class taught me how to use email, surf the web (yes, that’s what we said!), and create my own web page. Your parents will remember when you needed a phone line to access the internet – imagine having to log off Instagram every time your mom was expecting a phone call. Today, the web is everywhere. It provides incredible amounts of text and image data created by ordinary people — not only web sites, but social media from Twitter and Reddit to Instagram and YouTube.

Second, global computing power has increased tremendously. When he was in college, my husband worked as an intern on the Intel Paragon supercomputer, in its day the most powerful computer in the world. Today, an iPhone 11 is just as powerful. Add to that the development of computing clusters, where many computers work together on a shared problem, and the use of GPUs, to process large amounts of data in parallel. 

Third, to take advantage of all that computing power and all that naturally occurring data, over the last twenty years AI researchers have developed machine learning algorithms of increasing sophistication. For example, in 2012 Google Brain released the results of an experiment in which a neural network spanning a thousand computers was trained on ten million unlabeled images taken from YouTube. At the end, one of the top level neurons was found to respond strongly to images of human faces. Another responded to images of cats – which was why it came to be called The Cat Experiment. 

Of course, even more plentiful than images are texts, from Tweets to news stories to novels. The OpenAI company set out to apply similar techniques to the vast corpus of unlabeled text data from the Web. GPT-3 is their third and most successful attempt. 

I was curious what the acronym GPT stood for, and here is what I learned:

• “G” is for “Generative.” GPT-3 is an AI system that generates text, rather than categorizing a given text as happy or sad, or determining the gender of a character in a story, or other tasks an AI system might do.
• “P” is for “Pre-trained.” GPT-3 is pre-trained on unlabeled data from a wide range of sources. It could later be “fine-tuned” using labeled data to perform better on specific tasks. (By the way, if you’ve ever had to “select all the images containing a traffic light,” you’ve contributed to labeling image data for use in fine-tuning deep learning algorithms for use in self-driving cars.)
• “T” is for “Transformer,” a type of deep learning model designed to process unlabeled, sequential data such as text.

And so we have “Generative Pre-trained Transformer, version 3.” That’s as technical as this talk is going to get, and truthfully it has stretched the limits of my understanding. Fortunately my other new colleague, Parteek Kumar, will be teaching a Special Topics course on Machine Learning this spring, and we hope to offer such a course regularly in the future.

If anything, GPT-3 is far more magical than ELIZA ever was, because the inputs are so vast and its algorithms so obscure. Building GPT-3 took a team of 31 AI researchers, unimaginably beyond what I could have achieved as a senior in college.

So could GPT-3 write your thesis? Having wrestled with my fear that perhaps it could, in the end it seems clear that it could not write your thesis alone .

Here’s what scared me the most. While preparing this speech, I learned that GPT-3 was the first author on an academic paper about itself, currently under review for publication. 

But having used GPT-3 myself, I wondered what role was played by the article’s human co-authors. I found an essay in the June issue of Scientific American addressing this very question. 

Almira Osmanovic Thunström is a scientist who studies the role of artificial intelligence and virtual reality in mental health care. She found herself curious if GPT-3 could write about itself, so she asked it to respond to the following prompt: “Write an academic thesis in 500 words about GPT-3 and add scientific references and citations inside the text.” The quality of the result surprised her.

I had a similar experience. When I prompted GPT-3 to write a convocation speech, the verisimilitude of its first response surprised me. I was amazed that it was coherent and appropriate to the genre. The words are original; it’s not plagiarized. It even makes good use of grammatical parallelism. That is the response I read to you unedited, and truly what inspired me to write this speech. 

Thunström went on to use GPT-3 to write an entire academic paper. She gave GPT-3 a prompt for each section of the paper and selected the best of three responses, but refrained from any editing beyond that. 

It matters that Thunström allowed GPT-3 multiple chances to respond to her prompts. The developers of GPT-3 report among its limitations that in longer responses it can lose coherence, repeat itself, contradict itself, and insert non-sequiturs. When I prompted GPT-3 to write a second convocation speech, it wrote, “I am truly honored to be standing here before you as your President.” I decided not to read you that one. The third iteration wasn’t even a convocation speech, it was a graduation speech. I didn’t read you that one either. 

It also matters that neither Thunström nor I had any intention to pass off the words of GPT-3 as our own. I didn’t care if GPT-3’s commencement speech expressed sentiments that I share, because I intended to use it as a rhetorical device. Similarly, Thunström didn’t care if the paper written by GPT-3 was accurate; she wanted only to show that it could be done. She wonders what it will mean to respond to feedback from reviewers, when she receives it, because that seems beyond GPT-3’s capabilities.

As I reflected on Thunström’s experiment, I wondered, could GPT-3 have written an academic paper about itself before its creators published their research paper? I think the answer must be no. Only now that human beings have written about GPT-3, and those writings are included in its training data, can GPT-3 write about itself. 

While the commencement speech that GPT-3 wrote for me is original in one sense, it is highly derivative in another. I doubt that GPT-3 could write coherently on a topic that has never been addressed before.

As another experiment, I asked GPT-3 to summarize the last section of this speech. Here’s what it wrote: “In short, GPT-3 is a powerful AI tool that is capable of writing coherently on a variety of topics, but it is not yet able to write on topics that have never been addressed before.”

That is surprisingly not bad.

So will AI write your thesis? Although the question was worth asking, in the end I don’t think so. An AI might write a thesis, but it won’t write your thesis.

As you’ll learn in the first year seminar, while it’s important to write coherently, it’s still more important to ask good questions, read critically, and respond to feedback —- all things that AI can’t (yet) do.

If you do enlist the help of GPT-3 in your academic writing, make sure you adhere to OpenAI’s “Sharing and Publication Policy.” You must clearly indicate the role of AI in your work, as well as your editorial role. You must take full responsibility for any computer-generated text you publish, including any inaccuracy or bias. You should think carefully about what you hope to accomplish through the use of AI, and whether those ends are ethical. 

Like the developers of GPT-3, what scares me most is the use of AI text generation for bots, spam, phishing, and misinformation. AI can give us the illusion of intelligence, but it cannot be held accountable for that illusion. Only people can.

I’ll wrap up with one last quote from Weizenbaum. “ELIZA in its use so far has had as one of its principle objectives the concealment of its lack of understanding. But to encourage its conversational partner to offer inputs from which it can select remedial information, it must reveal its misunderstanding.” 

Weizenbaum was writing about a computer program, but the same applies to all of us. To learn, we must reveal our misunderstandings.

So, Whitties, here is my real charge to you as you enter your first year: Learn to ask good questions. Be brave, be curious, be vulnerable.

And if an AI does co-author your thesis, I hope I’ll be the first to know.

https://beta.openai.com/playground https://en.wikipedia.org/wiki/GPT-3 https://www.dataversity.net/brief-history-deep-learning/ https://www.ceros.com/inspire/originals/recaptcha-waymo-future-of-self-driving-cars/   https://www.scientificamerican.com/article/weasked-gpt-3-to-write-an-academic-paper-about-itself-mdash-then-we-tried-to-get-it-published/  

Share this:

• Click to print (Opens in new window)
• Click to share on Facebook (Opens in new window)
• Click to share on Twitter (Opens in new window)

3 thoughts on “ Will AI write your thesis? ”

Pingback: Will AI write your thesis? – Full-Stack Feed

hank you for addressing the intriguing topic of AI potentially writing academic theses. The advancements in AI technology have indeed opened up new possibilities, but it’s important to consider the implications and limitations of such automation.

I have a question: What are some of the ethical considerations surrounding AI-generated theses? Are there any concerns regarding originality, critical thinking, or the integrity of academic research when relying on AI for thesis writing?

It appears that your response was generated by AI due to the appearance of a rule-based format (series of 3) and slightly journalistic voice. Am I correct in this assumption? If so, well done :).

Also, your post seems to have been copied from elsewhere as evident by the missing “T” at the beginning of your post.

Leave a Reply Cancel reply

Your email address will not be published. Required fields are marked *

Save my name, email, and website in this browser for the next time I comment.

Select language

Artificial Intelligence

Thesis project.

In the final thesis project, the student carries out a research project under the supervision of one of the staff members of the research groups offering the AI programme. The project can be done based at Utrecht University, at a company or research institute, or at a foreign university (see also: ‘ stay abroad - traineeship ’).

Before starting the thesis project students are strongly advised to first attend the thesis information session meeting, which is offered at the start of each teaching period. See course INFOMTIMAI for more info .

When looking for a project, please check the following sources.

• Konjoin always has a number of AI projects.
• Jobteaser also has interesting external internships for AI students.

General description

The AI Thesis Project is split into a 14 EC project proposal phase (INFOMAI1) and a 30 EC thesis phase (INFOMAI2). The thesis project takes about 8 months (three periods). The set up phase that is necessary to arrange your project is not counted. 

The thesis project consists of a project idea, a UU graduation supervisor, and a graduation project facilitator. The project facilitator can either be a company or the University. Original ideas from the students are welcome, as long as they are aligned with the research interests and/or proposed projects by the supervisors. 

For a thesis project, the student always needs a supervisor from one of the research groups of the UU offering the AI programme. If the final project is conducted within a company or external institute, both a local supervisor within the company/institute and a supervisor of the AI programme teaching staff monitor and guide the student. 

When can a thesis be started? When all courses are successfully completed, with the exception of Dilemmas of the Scientist (FI-MHPSDL1 and FI-MHPSDL2), for which you can do the second workshop (FI-MHPSDL2) during your theses process. Further exceptions can be given by the AI programme coordinator for students with one pending course. Note that you should start looking for a supervisor and a subject before you have finished all your courses (see “Set Up” below). 

Where do I start? Read the information on the various stages of the thesis project below. If you have any questions not covered here contact the programme coordinator  ( [email protected] ). 

How long does the thesis take? Normally, a thesis project (phase 1 + phase 2) runs for 3 periods/terms (see the schedules). However, holidays, courses or other activities may lead to a thesis projects that takes slightly longer. Please Please see below what to do when your thesis is delayed and you have to apply for a thesis deadline extension (part 1 and/or part 2) 

Previous theses. To get an overview of what an AI thesis looks like, you can consult previous theses online . 

Learning goals. After completing your thesis project, you will:

• have advanced knowledge about a specific subject within AI
• be able to findings on a specific subject within the broader, interdisciplinary field of AI
• be able to independently perform a critical literature study
• be able to formulate a research question of interest to AI and a plan / method to answer this research question
• be able to perform scientific research according to a predetermined plan and a standard method within AI
• be able to report the research findings in the form of a scientific thesis
• be able to report the research findings by means of an oral presentation

This preliminary step is executed before the official start of Phase 1. The duration largely depends on how quickly a supervisor is found and a topic is agreed upon. This part is excluded from the duration of the thesis project. 

1. Find a project and a supervisor   You can do an external or an internal (UU) project. The following tips might come in handy when looking for a project. 

• Think about the courses you found interesting and ask the lecturers of these courses if they have/know of any projects.
• Jobteaser also has interesting external internships for AI students. 

Note that any topic has to be agreed with the UU staff member who will act as a first supervisor. Arrange meetings with staff members to discuss possible options, based on their research interests (look at their webpages, their Google Scholar profile, or ask the Programme Coordinator ( [email protected] ). If unsure about possible topics, please arrange a meeting with the Programme coordinator. Students can also try to arrange a project that fits within an internship with a company. Any project, however, requires a first supervisor from the department who guarantees the scientific quality of the thesis project, so it is advisable to talk to potential supervisors and/or the graduation coordinator before agreeing on an internship. 

2. Define your project  Together with the first supervisor, describe your project's title, problem, aims, and research goals. Come up with a short textual description (about 200 words). Also make clear arrangements with your first supervisor concerning planning, holidays, supervision meetings and so forth. Please make sure you have a clear understanding with your first supervisor regarding deadlines and extra work, holidays etc. to be done during the thesis project. Normally, a thesis project runs for 3 periods/terms, but you can set any reasonable deadline in agreement with your supervisor. Please see below what to do when your thesis is delayed and you have to apply for a thesis deadline extension (part 1 and/or part 2).

3. Ensure adherence to Ethics and Privacy regulations -  Quick Scan From Period 2 of 2022-23, all Master AI thesis projects require ethics and privacy approval. For projects that do not involve human users and data privacy issues this will be a very brief and straightforward process, but you still need to complete an ethics checklist. If you are doing your project with a supervisor in a department that already has ethics approval process in place (such as Cognitive Psychology), then ask the supervisor what you need to do in order to obtain ethics approval. Otherwise, please inform your supervisor that you need to obtain an ethics and privacy approval. Go to the website that contains the ethics checklist and sample Information sheets and consent forms:  https://www.uu.nl/en/research/institute-of-information-and-computing-sciences/ethics-and-privacy . First, download the Word form and discuss how to fill it in with your supervisor. Then fill in the Qualtrics form. Please fill in as the moderator email: [email protected] .

4. Work placement agreement If you conduct a project outside UU, the GSNS Work Placement Agreement (WPA)  should be filled in, and signed by the student, company supervisor, and the Science Research Project Coordinator . Deviations to the standard contract shall be discussed with the Science Research Project Coordinator. 

You need to fill out and upload your WPA with your Research Project application form (see next step) in OSIRIS student.

5. Formalize the start of your Research Project via submitting the Research Project application form  

Use Osiris student  (select 'MyCases', 'Start Case', ‘Research Project GSNS’) to submit your research project application form; if applicable, you will also upload the signed Work Placement Agreement with your application form in OSIRIS. 

Important: in order to apply completely and correctly, you must have discussed the project setup with your intended project supervisor beforehand! We advise you to study the request form previous to discussing it with your supervisor, or fill it out together, to make sure you obtain all of the information required. 

After submitting your application form in OSIRIS, your form will be forwarded to your 1st and 2nd Examiner (supervisors), master’s programme coordinator, the Board of Examiners and Student Affairs for checks and approvals. You may be asked for modifications, should they find any problems with the form. 

Please note. You cannot register yourself in OSIRIS for the relevant research project courses (INFOMAI1 and INFOMAI2). You will be automatically registered for part 1 of the project upon approval of the Research Project Application Form.

Phase 1 - Project proposal

The phase comprises 14 EC (i.e. 10 weeks of full-time work) and is intended for you to do a preliminary study (usually in the form of literature study), and to propose and plan your research. Importantly, this phase will give a go/no-go decision towards Phase-2. You are expected to deliver a research proposal consisting of the following: 

• A literature study section, summarizing works that are relevant to your research. 
• Well formulated research question(s). 
• A plan for the second part of the thesis.

Additionally, depending on the nature of the project, your supervisor may require you to perform some initial research work in Phase-1, either in order to provide a convincing argument towards the prospect and feasibility of your Phase-2, or for efficiency to already do some work of Phase-2, e.g. developing an initial theory or building a first prototype of an algorithm. If such work is required, make an agreement with your supervisor on the scope of this work. 

At the end of Phase-1 the supervisor(s) will make a go/no-go decision. This decision, in terms of pass or not pass, will be entered in Osiris. Phase-1 assessment criteria: 

• Scientific quality. This concerns the quality of the literature study, the relevance and impact of the research questions, the merit of proposed research method. 
• Writing skills. This concerns the quality of your writing, use of English, textual structure, and coherence/consistency of your text. 
• Planning. This concerns the clarity and feasibility of the proposed planning. 
• The quality of additional work, if such is required.

An  example assessment form  with more detailed criteria is available. Please use this form only as a discussion piece and do not send in paper or scanned forms.

Phase 2 - Thesis

The second part comprises 30EC (i.e. 21 weeks full-time). You will complete (at least) the following items: 

• Perform and complete your research according to your plan (Phase 1). 
• Write your thesis that presents your research and its results. 
• Present and defend your results and conclusion. You are asked to prepare a presentation about your research that is understandable by fellow students. The defence will be 45 minutes long; 30 minutes for your presentation, and 15 minutes for questions. 

Content of the thesis.  In addition to the main text describing the research, the master thesis should at least contain: 

• a front page, containing: name of the student, name of the supervisors, student number, date, name of the program (master Artificial Intelligence, Utrecht University); 
• an abstract; 
• an introduction and a conclusion;  
• a brief discussion of the relevance of the thesis topic for the field of AI; 
• a list of references.   

Please discuss the exact requirements for your thesis with your daily supervisor/first examiner at the beginning of your project.  

Phase-2 assessment criteria.  Your thesis is assessed using the following criteria: 

• Project process (30%). This concerns your ability to work independently, to take initiative, to position your work in a broader context, to adapt to new requirements and developments, and to finish the thesis on time. 
• Project report (30%). This concerns the ability to clearly formulate problems, to summarize the results, to compare them with related scientific work elsewhere, and to suggest future research lines. This also concerns clear, consistent, and unambiguous use of language in the thesis. The text should give the readers confidence in that you understand the chronology, structure, and logical entities in your own text; and thus know what you write. 
• Project results (30%). This concerns the level and importance of your results. Are the results publishable as a scientific paper? The difficulty of the problem that you solve also plays an important role, as well as the amount/extent of the work you carry out. These are aspects that are important: the effectiveness of the chosen approach, completeness and preciseness of the literature study, arguments for the choices made, insight in the limitations of the chosen approach, proper interpretation of the results achieved, level of abstraction, convincing argument, proofs or statistical analysis. 
• Project presentation (10%). The ability to orally present your project and its results clearly and concisely. 

An  example assessment form  with more detailed criteria is available. Please use this form only as a discussion piece and do not send in paper or scanned forms. 

Phase 2 - Wrap up

When approaching the finalization of the thesis (i.e, when the supervisors think so), it is time to wrap up the project and graduate. 

• Set date for graduation presentation : both supervisors should agree on the date, including the time. 

Arrange (virtual) room for defence : The public defence can take place in Teams. If desired by the candidate and/or the supervisors, you can also defend your thesis in a lecture room on campus, ideally with a livestream or in a hybrid form so that e.g. fellow students or friends can also watch online. You can make a Teams meeting yourself, and send an e-mail to the secretariat ( [email protected] ) to arrange for a suitable room for your presentation. Please make sure to include the time, date, name of the thesis, supervisor, and the number of expected attendees. 

Inform the AI coordinator  ( [email protected] ) about the details of your defence (title, abstract, date, time, room and/or Teams link). The coordinator will announce the defence on Teams and via the mailing list. 

Thesis defence : the student gives a presentation of 30 minutes, followed by a question-and-answer session that typically lasts about 15-20 minutes. Your first and second supervisor will decide on your grade and announce this after your presentation. 

Upload thesis  to Osiris Student:  After the defence, the student must  upload the final version of their thesis through Osiris Student > my cases . 

Archiving and publishing thesis to Thesis Archive You will be asked once more to upload the final version of your thesis through OSIRIS Student, yet, this time this is for archiving and publishing purposes. The Case will not be available by default via OSIRIS Student. You will receive an email as soon as the Case in OSIRIS Student is available to you. More information on thesis archiving and publication can be found here . 

Graduation checks and ceremony

The Student Desk at Student Affairs keeps track of your study progress in Osiris. When Osiris indicates that you have completed all the required elements of your degree your file is forwarded to the Board of Examiners. These checks only occur around the 15th of each month. Therefore, do you wish to graduate by the end of the month, please ensure you have completed all elements of your degree before the 15th of the month so all your credits are registered in OSIRIS. This also includes the uploads of your final thesis.  

The Board of Examiners then checks whether you meet all examination requirements. Following the Board's approval your graduation date will be emailed to you on your UU email account.  

Please DO NOT terminate your enrolment in Studielink until the Student Desk has informed you about the decision of the Board of Examiners and you have received your graduation date. For further information, please check the graduation  page.

What to do when your research project is delayed and a Research Project deadline extension is required?

Please note that the “protocol delay in graduation” applies when a project is delayed. This protocol can be found in appendix 2 of the Education and Examination Regulations .

• When you are delayed, e.g., due to personal circumstances or due to unforeseen circumstances within the project, it is important that you make an appointment with your study advisor in time (before the final deadline of part 1 and/or part 2).  
• It is further important that you discuss the delay with your supervisor and set new realistic goals and deadlines (where possible).  
• Next, you need to apply through OSIRIS Student > ‘MyCases’ > 'Start Case' > ‘Request to the Board of Examiners GSNS’, and then choose the appropriate request type: “Delay of research or thesis project”. It is important that you upload a statement from the study advisor (hence why the importance to speak to your study advisor as soon as possible when a delay occurs) and a copy of an email in which the supervisors support the request for a deadline extension. You further need to include a proposed new deadline and short statement support your request.

Follow Utrecht University

Utrecht University Heidelberglaan 8 3584 CS Utrecht The Netherlands Tel. +31 (0)30 253 35 50

Navigation auf uzh.ch

Department of Informatics Artificial Intelligence and Machine Learning Group

Quicklinks und sprachwechsel, main navigation.

This page lists several ideas for Bachelor's and Master's theses. Most of the thesis proposals are based on recent research, exploring branches that most researchers overlook. Some Bachelor's theses proposals can be extended to Master's theses. While each page lists requirements, these are not strict requirements at the beginning of the thesis, but rather skills to acquire during the thesis. For more details, send me an email.  Prof. Dr. Manuel Günther

uzh-wcms-publications.subpageListDialog.labelUnterseiten

• Open Theses
• Active Theses
• Past Theses

Topics for new Theses

Students are most welcome to develop their own topics, and we are happy to supervise them as long as the topic is anywhere close to our fields of expertise -- otherwise we are not able to provide fruitful input. General topics of interest to the AIML group include (but are not limited to):

• Face Recognition: the identification of a person based on an image or a video of their face
• Facial Attribut Classification: the classification of attributes (gender, hair color, face shape, ...) from facial images
• Open-Set Classification: teaching classifiers to discern objects of classes of no interest
• Adversarial Samples: generating small modifications of correctly classified samples that change the output of classifiers, or preventing these kinds of attacks
• Explainable AI: which parts of the inputs are of great importance to the decision of deep learning systems
• Traditional Features: How to make use of Gabor wavelets or other traditional image processing techniques in deep networks
• Medical Image Processing: Applying of the above methods to medical images; I only host topics provided by other insitutitions (ETH, USZ, Balgrist, Idiap, ...)

Topics that we generally do not supervise are in the area of Natural Language Processing, Social Media, Robotics or Reinforcement Learning since we have other experts on these topics ( Prof. Dr. Martin Volk , Prof. Dr. Anikó Hannák , Prof. Dr. Davide Scaramuzza , Prof. Dr. Giorgia Ramponi ) in our department . If you have a topic in mind and believe that we would be a good supervisor for, please send me an email.  Prof. Dr. Manuel Günther

Requirements and Submissions

It is a requirement to use LaTeX for writing the final thesis document. Students should use the AIML Thesis Template (ZIP, 314 KB) .

Generally, all theses need to be defended, including Master theses (mandatory by the Department rules ) and Bachelor theses. Usually, the defense will be scheduled about 3-5 weeks after the submission of the thesis. Deviating from the recommendations, the time for a Master thesis' defense presentation is 30 minutes followed by 15 minutes of questions, while a Bachelor thesis' defense should take 20-30 minutes of presentation and 10-15 minutes of questions.

Additionally, the source code for the thesis needs to be submitted, approximately at the time of the defense. The source code is typically written in Python and use the PyTorch library. It should be self-contained and make use only of public libraries and data (if possible). Source code needs to be documented. This allows fellow students to make use of previously implemented code.

• Skip to content  (access key: 1)
• Skip to Search  (access key: 2)

Program Details.

Detailed information about the ideal recommended study plan, etc. is available here.

Academic Credit and Transferring Academic Credit

• Academic Credit for School Subjects for the Bachelor's in Artificial Intelligence , opens an external URL
• Credit Transfer between Bioinformatics and Artificial Intelligence (PDF, 641,2 KB) , opens in new window
• Credit Transfer between Computer Sciences and Artificial Intelligence (PDF, 156,0 KB) , opens in new window
• Complete this form to apply for credit transfer for courses not listed in the documents above: Pre-Check Form for AI Curriculum Accreditations.xlsx , opens a file

ABA Online Workshop on May, 16, 2024 from 12-1pm

You are eager to get a job in Austria? You are thinking of starting your own startup , opens an external URL in a new window ?

The Austrian Business Agency , opens an external URL in a new window (ABA), the business location agency of the Austrian Ministry of Economy, is eager to help. As a government agency, ABA offers free-of-charge support and information for international students and graduates at Austrian universities. More information here , opens an external URL in a new window .

If you have missed the webinar, please find the presentations below.

• ABA Startups (PDF, 4,4 MB) , opens in new window
• ABA WORK IN AUSTRIA (PDF, 6,5 MB) , opens in new window
• INVEST IN AUSTRIA (PDF, 1,2 MB) , opens in new window

Semester Schedule for WS 2023/24 and SS 2024

To help you get off to a good start, we have put together a timetable for the 2023/2024 Winter term and 2024 Summer term. These schedules are for reference and can be adjusted to your personal schedule. The livestreams to Vienna/Bregenz are marked in violet in the upper left corner.

Summer term 2024 you can choose for your Gender Studies course between 536.300 (Wednesday, 10:15-11:45) and 536.540 (blocked, Saturdays).

• Timetable Winter term 2023/24 (PDF, 157,1 KB) , opens in new window
• Timetable Summer term 2024 (PDF, 132,8 KB) , opens in new window

Types of Courses

The program contains the following types of courses:

Preparation Courses

Each year in September, before the start of the Winter Semester, new AI students can take preparation courses in Python Programming ( ASTAIBAPPPV20 , opens an external URL in a new window ) as well as in Mathematics ( 531.059 , opens an external URL in a new window ) at the JKU. These courses are entirely optional and students can sign up via KUSSS.

Remote Learning

In order to successfully complete the AI program, students will be required to come to Linz in-person at least once in order to officially enroll in the degree program in person. Some courses will require on-site and in-person attendance, either in Linz, Bregenz, or Vienna. Students are required to be physically present to take examinations either in Linz, Bregenz, or in Vienna. Examinations may take place during the entire semester. The curriculum has been designed for students residing in close proximity to Linz, Bregenz, or Vienna. 

Many of the courses in the AI program are offered at the JKU's satellite campus in Vienna and Bregenz, as either as a livestream or as a video conference (please check schedule).

Course Information

Courses are usually offered only once a year. While you can enroll during the Summer Semester to begin studying AI, starting the program at this time means you may be subject to adjusting your schedule differently as compared to the suggested schedule of courses. We highly recommend starting the program in October.

Area of Specialization

Students can freely choose their Area of Specialization courses from a list available in the course catalog. The section headlines “Computer and Data Science”, “Mathematics”, “Mechatronics”, “Live Sciences”, “Special Topics”, and “Miscellaneous” as listed in the course catalog are for clarity purposes only.

Seminars and Practice-Oriented Work

The “Seminar in AI” (3 ECTS credits, 4th semester, Summer Semester) and the “Practical Work in AI” (7.5 ECTS credits, 5th semester, Winter Semester) have been designed to help students prepare to write the required Bachelor’s thesis (6th semester, Summer Semester). Students can, however, switch their subject area again, if they wish to. Faculty members will offer these courses when supervising undegraduate students. Different institutes will offer the courses which students will see in KUSSS as different “group options”. The Bachelor's thesis itself will be formally supervised as part of the “Bachelor’s Thesis Seminar in AI” (9 ECTS credits, 6th semester, Summer Semester). This course is, however, offered each semester, allowing students to complete their Bachelor’s degree during the Winter Semester as well.

Mentoring Program Procedure

At the beginning of October - during the first class: " KV Lecture Series Artificial Intelligence " - we will create tutored groups of newly enrolled (first semester) students. The tutors are higher-semester AI students who will be available to show first-year students around the university, help them choose the right courses, and provide first-hand information about classes, exercises, and tutorials. The social aspect is also important and you will quickly meet others during the first few weeks to form study groups, etc. Many of these friendships will last throughout your studies and beyond. A mentor from among the AI professors will be assigned to each tutorial group. The tutored groups will meet with their mentor together and the mentor will supervise the students individually during the first semester. If you have any questions or concerns about the program, you can discuss these issues during these meetings. Our goal is to address any issues that arise quickly and solve them unbureaucratically.

Here , opens an external URL in a new window you can find more answers to your questions by the Austrian Students´ Union (ÖH).

FAQs , opens an external URL in a new window for international students by the Austrian Students´ Union (ÖH).

The 9 most asked questions and the corresponding answers

Can i start the program in summer term and does it make sense to do so.

It is possible to start the program in summer term. Nevertheless, as all courses are only offered once a year and start in winter term, it is difficult to follow the study plan with 30 ECTS per semester. Only in exceptional cases in which a student has the necessary knowledge of introductory courses and plans to later accredit those, a start in summer term is advisable.

I have missed the official registration period. Is it still possible to register?

Yes, for all lectures (LV) and combined courses (KV), you can still register. Please write an e-mail to the teacher or the institute´s secretary to ask for late registration. For tutorial groups, this might be more difficult as they are quickly fully booked and also require attendance (in person or online) and regular submissions.

I want to study the AI program with the distance learning option. Where can I see which courses/tutorial groups are offered with this option?

In general, all lectures in the program are recorded and can be accessed in Moodle. For the tutorial groups, please register for the groups with distance learning option in KUSSS.

I live in Vienna/Bregenz. Where can I see which courses are live streamed to the distance learning centres?

Please check the timetable. Streamed courses are color-coded.

All distance learning tutorial groups are fully booked. What can I do?

You can still register and will be put on a waiting list if the group is already fully booked. If spots free up, you are first in line as well as if new tutorial groups are created. If the registration period is over you can still write an e-mail to the teacher or in case of courses offered from the Institute for Machine Learning to [email protected].

How and where can I take an exam at the end of the course?

Some exams are offered online, others have to be taken in person in the lecture hall in Vienna, Bregenz or Linz. Please register online in KUSSS for one of the options offered.

How can I get in touch with my fellow students?

In Linz you can meet them in the lecture hall. Our distance learning centres in Vienna and Bregenz offer rooms to study together and follow the streams every day of the week (not only at the times of the scheduled streams). Also a discord channel , opens an external URL in a new window was created for this reason.

How can I apply for accreditation of subjects?

You find an excel on top of the website that you should fill out and send together with all certificates as well as descriptions of the successfully completed courses to office(at)ai-lab.jku.at

Is it possible to get moodle access to introductory courses offered in winter term to prepare for a start in summer term?

For some courses and only once you have gotten a matriculation number after being officially enrolled in the program, we can grant you access to course materials in moodle. Please write an e-mail to office(at)ai-lab.jku.at in this case.

Contact Information

If you have any questions or concerns regarding the AI program that have not been addressed above or below, please feel free to contact us by sending an e-mail to:

The AI office in Linz: office(at)ai-lab.jku.at

The AI office in Vienna: ai-wien(at)jku.at

The AI office in Bregenz: ai-bregenz(at)jku.at

The Austrian Student Union for Computer Science and AI: ai(at)oeh.jku.at

In addition, see the Student Union , opens an external URL in a new window homepage and their Study Guide for Artificial Intelligence , opens an external URL in a new window

Additional Information

Study abroad programs, merit scholarships, scholarships / grants, feel like studying.

Need an Overview? Here you will find an overview of general information about the program.

• Course Catalog / Curriculum , opens an external URL in a new window
• Degree Program Brochure , opens a file

Service & Academic Advising

• A Guide to Start Your Studies
• Studies from A to Z
• Student Information and Academic Advising Services (SIBS) , opens in new window
• Admissions Office , opens in new window
• JKU Austrian Student Union , opens an external URL in a new window

Additional Links

• Linz Institute of Technology (LIT)
• Faculty of Engineering & Natural Sciences
• Institute of Computational Perception
• Institute for Machine Learning
• Institute for Formal Models and Verification , opens an external URL in a new window

Johannes Kepler University Linz

Altenberger Straße 69

4040 Linz, Austria

Use of cookies

Our website uses cookies to ensure you get the best experience on our website, for analytical purposes, to provide social media features, and for targeted advertising. This it is necessary in order to pass information on to respective service providers. If you would like additional information about cookies on this website, please see our data privacy policy .

Required cookies

These cookies are required to help our website run smoothly.

Web statistics cookies

These cookies help us to continuously improve our services and adapt our website to your needs. We statistically evaluate the pseudonymized data collected from our website.

Marketing cookies

These cookies help us make our services more attractive to you as well as optimize our advertising and website content. We analyze and evaluate pseudonymised data collected from our website.

ThesisAI - world's first AI assistant that can draft a whole scientific document. Up to 40 pages. With just one prompt. Inline citations on paper or page level. Native LaTeX integration, more than 20 languages.

Use ThesisAI to draft your scientific document

Output as PDF, Word, LaTeX, BibTex

Inline citation (numeric, alphabetic, APA, MLA, IEEE) and bibliography

Export document to Overleaf for manual editing

Check out this example documents

Upload up to 100 reference papers to be cited

Only $5 per document, no subscription

Reading and writing math

Chat with document and provide additional reference papers after initial creation

Finetuned LLM for scientific inline citation

You are not subscribed to any plan