Artificial Intelligence (Summer Semester, jointly with Professor Jörg Hoffmann)
This course offers a general introduction into the field of Artificial Intelligence (AI), its history, key assumptions, paradigms, concepts, and fundamental methods. Students learn to master and apply techniques developed in the fields of intelligent agents, search algorithms and game playing, knowledge representation, logical reasoning and deduction, planning, constraint reasoning, machine learning, and reasoning under uncertainty. We explain how and when these methods work and for which problem classes they are suitable to successfully build AI systems. With the knowledge acquired in this course, students are well-prepared to successfully attend the many special lectures on AI offered at UdS or to write their Bachelor or Master thesis in the field of AI.
Latest publicly available lecture materials
Architectural Thinking for Intelligent Systems (Winter Semester)
This course teaches established and successful methods of modern architectural thinking, which are applied by software architects for the systematic planning, conception, and evaluation of complex software (and hardware) architectures. Examples are taken from the context of intelligent systems, where AI technology needs to be integrated into often complex hardware and software environments. Starting from an initially vague understanding of the system to be built, we systematically refine our understanding by asking the right questions, developing a vision, applying architectural styles and pattern, minimizing risks, and evaluating the outcome of architectural decisions.
Latest publicly available lecture materials
Bachelor and Master Thesis Projects
The main focus of the AI chair is on optimization problems solved using AI CP/SAT solvers or reinforcement learning. We are especially interested in the intelligibility of models and in methods to make modeling optimization problems easier. Furthermore, we are interested in reference architectures for intelligent systems. Apart from these areas, we are also open to your suggestions for topics.
Below you find our current suggestions for thesis projects for the summer term 2021. For each topic, we give a profile showing you if the work is more on the conceptual or practical level, or a combination of both. You also see which topics are available for Bachelor/Master theses.
Please apply with a short CV, your transcript of records, and by relating either to one of the topics below or to one of my current research projects. Describe own ideas in a short paragraph.
Reinforcement Learning for Flow-Shop and Job-Shop Scheduling Problems
Modern production systems move towards customer-individual manufacturing, which causes a high volatility in supply chain management and requires to apply flexible optimization methods. Using digital twins, you explore whether robust and efficient manufacturing operations can be learned using a reinforcement learning approach. Finally, you will compare your results to the results of current state of the art constraint programming solvers such as IBM Cplex CP and Google OR-Tools. Profile: conceptual/practical, Bachelor/Master
Constraint Patterns for Flow Shop and Traveling Salesperson Problems
Modeling an optimization problem such that a modern constraint solver can easily scale to large instances is an art. Some techniques exist such as developing a dual model or adding redundant constraints, but the techniques cannot be easily applied to an existing model in a fully automatic fashion. You explore the existing set of techniques and work on collecting patterns for Flow Shop and Traveling Sales Person problems. You apply these patterns to constraint solvers and evaluate the quality of resulting models based on the occurrence and interplay of several patterns. Profile: conceptual/practical, Bachelor/Master
Learning Patterns from Event Logs
Constraint programming solvers, such as IBM Cplex CP or Google OR-Tools, are considered one of the main paradigms for solving combinatorial optimization problems. However, to use these solvers the problem must be explicitly modelled by an expert, which represents a bottleneck in the more widespread use of these powerful techniques. Learning constraints from examples or event logs, so-called Constraint Acquisition, is a powerful method for simplifying the modeling process. Recently, we proposed a novel Constraint Acquisition system which makes use of patterns, commonly reoccurring sets of constraints, for Flow Shop problems. To enhance the potential of the system, further patterns need to be identified and explicitly and efficiently implemented. This project concerns the identification of patterns, the investigation of efficiently evaluable conditions for the occurrence of the pattern and the implementation of the patterns. Profile: conceptual/practical, Master
A Blackboard for the Multi-Car Elevator Control System
Multi-car elevator systems
consist of several autonomous cabins that transport passengers in a system of transportation shafts of flexible layout. A complex state representation serves as central component of the multi-car elevator system, which holds information about the cabins, e.g., their positions, door states, currently planned stops for each cabin, and building information like the number of floors. The dynamic information in this state representation highly depends on the current calls by passengers and strategies to answer them. In this thesis, you will explore how this complex state representation can be implemented using the Blackboard
architectural style. This style maps control strategies and passenger calls to knowledge sources, which jointly work on the blackboard under the supervision of a controller. Profile: practical, Master/Bachelor
Software Cost Estimation for AI Systems with COCOMO 2
One of the fundamental problems of planning a software project is to estimate the costs of the project. One of the first models to address this problem is the COCOMO model (constructive cost model). COCOMO 2
was published by B. Boehm in 1995 and is widely used in the industry. The task of this thesis is to explore which cost estimation models are currently used in the industry, compare these models, and explore if the COCOMO 2 model (or other models) can be applied to AI systems. Profile: conceptual, Master/Bachelor
Recognizing Hard Problem Instances
Whereas NP-complete problems are considered “hard”, there are many instances of NP-complete problems, which are easy to solve, i.e., they can be solved in a short time. It is conjectured that all NP-complete problems have at least one parameter describing how constrained a problem instance is and that the hard to solve instances are around a critical value of this parameter, see “Where the really hard problems are” by Peter Cheeseman, Bob Kanefsky and William M. Taylor. In this project, you work with candidate parameters and determine their influence on the difficulty of an instance. You run experiments with constraint solvers such as IBM Cplex CP and Google OR-Tools to measure the influence of a parameter and you also conduct theoretical analyses to confirm a parameter or to find alternative parameter candidates. Profile: practical, Master/Bachelor
Classification of Traveling Salesman Problems
Given a set of cities and distances between each pair of cities, the Traveling Salesperson problem is the problem of finding the shortest route visiting each city exactly once. It is an NP-hard problem and one of the most famous problems in combinatorial optimization. Today, many variants of this famous problem have been introduced in the literature. This project conducts a thorough literature review of recent TSP literature with the aim of creating a good overview of all relevant variants. Profile: conceptual, Master
Construction Site Optimization
This project concerns the optimization of processes on large construction sites. In particular, the routes of lorries, which transport materials from source places (where diggers load the lorries) to target places (where the material is needed), need to be optimized. A collaborating research group recognizes actions from the lorries based on sensor data and translates this data to event logs. The aim of this project is to model the construction site problem with the IBM Cplex or Google OR-tools solvers based on the data given in the event logs.
Machine Learning for Construction Site Optimization
This project concerns the optimization of processes on large construction sites. In particular, the routes of lorries, which transport materials from source places (where diggers load the lorries) to target places (where the material is needed), need to be optimized. A collaborating research group recognizes actions from the lorries based on sensor data and translates this data to event logs. The aim of this project is to use Unsupervised Machine Learning or Reinforcement Learning techniques to analyze the event log data and to optimize the schedules of the lorries. Profile: conceptual/practical, Bachelor/Master
Generation of Test Data for Constraint Optimization Problems
Given an application domain and an optimization problem, one usually tests optimization algorithms on a set of instances to determine how well the algorithm scales. Very often, the amount of test instances is limited, and one is interested to systematically generate instances of varying size and difficulty. In this project, you will generate test data by applying methods, which systematically vary parameters influencing the size and difficulty of an instance. You run experiments and test constraint solvers such as IBM Cplex CP and Google OR-Tools or reinforcement learners on the generated test data. Profile: practical, Bachelor
AI Systems as Teaching Assistants for the AI Lecture
With all lectures moved online last Summer, we have produced a wide variety of digital resources to support the AI lecture at UdS. The AI lecture alone has over 1500 pages of slides. We currently develop Chatbot solutions, which make it easier for students to access the vast amount of content, work more effectively on exercises, and actively test their knowledge. In this project, you build on prior work and you harden the existing systems such that they can be used in production in the upcoming semesters. As technology candidates we currently work with Google DialogFlow and Alexa AWS services. You must have passed the AI lecture at UdS with good results to apply for this project. Profile: practical, Bachelor/Master
AI in the Movies – Myth and Reality
There is a long history of very good movies that investigate challenges around AI and that make predictions about AI systems. You analyze these movies and contrast them with the state-of-the-art in AI and with predictions made currently on ai.sciencebets.org. Your task is to create an entertaining and insightful report and presentation on the subject. Profile: practical, Bachelor
Artistic Visualizations of AI Search Algorithms
Deep learning has inspired artists to visualize the processing of information, which happens in the various layers of a neural network, see for example. We are looking for visualizations that bring out the beauty of reasoning in stochastic search algorithms such as UCT or inside constraint and SAT solvers. If you have an artistic streak and if you have passed the AI lecture with good results, this may be a project for you. Profile: practical, Bachelor/Master