CHAIR OF

COMPUTATIONAL

INTELLIGENCE

• Christoph Steup
• Alexander Dockhorn
• Sebastian Mai
• Michael Preuß
• Sanaz Mostaghim

Language: 

The course will be held in English or German depending on the participant's preferred language.

Participants:

All students of bachelor and master curriculums of the faculty are eligible to visit the course. The course can be taken as Digital Engineering Project, Inter-Disciplinary Project or Team-Project. The actual type of course depends on the needs of the students and the available Topics. It is mandatory for a participant to have background knowledge in at least one of the following topics:

• Robotics
• Programming in C/C++ , Java, Javascript or Python
• Theorie and Algorithms of Swarm Intellligence
• Communications and Networks
• Development of Embedded Applications  in C
• Robot Operating System (ROS)
• Linux Server Admisitration
• Deep Learning / Machine Learning / Reinforcement Learning
• Computational Intelligence in Games
• Optimization
• Game Development
• Evolutionary Algorithms
• Procedural Content Generation

First Meeting: 

The first meeting is mandatory for at least one team member per group and will be used to present available projects and give the opportunity to form teams for collaboratively working on them during the semester.  In case you cannot participate, please contact the respective contact person of the topics listed below.

The meeting will take place on the 08.04.2022 at 13:00 in G29.

Organization:

The course will be taken in groups of 3-5 Students per topic. You can already come as a group but the the final group setup will be chosen by use depending on your background. The individual topics are not fully fixed, extensions, and modifications are possible depending on the skills and interests of participating students. This will be discussed in the first meeting and the tasks and requirements will be fixed in a design document.

We will have personalized meetings with the working groups. The frequency of the meetings depends on the progress of the group. The Swarmlab will be available as a coworking space during the semester.

The result of each project is a working demonstration with commented source code and a written documentation indicating the general concept and a description on "How to to start the demo".

Available Topics:

Teamproject: Extending the CIcker

The aim of this project is to extend our real-world game AI platform, CIcker. The current state of the CIcker is not yet robust enough for long-term play against Humans and AI vs AI. Additionally, automatic ball feeding and unstuck behaviors are necessary to enable autonomous training for reinforcement learning. This project will solve exactly these issues:

Tasks

• Ruggedize / Robustify hardware for long-term operation
• Develop, implement, test  and evaluate an automatic ball feeding mechanism
• Develop, implement, test  and evaluate an automatic unstuck behavior

Prerequisits and capabilities (recommended)

• ROS
• CAD-Design
• 3D-Printing
• Microcontroller Programming

Contact: steup@ovgu.de

Scientific Project: Humanitarian logistics via UAVs

This scientific project is in the field of disaster management. The project consists of determining exploration strategies to evaluate the accessibility of victims by road using a fleet of drones. The evaluation of the known road network can be addressed as a real-time drone routing problem aimed at minimizing the required route length to determine the accessibility of all victim locations by road. This research project is offered in collaboration with the Chair of Computational Intelligence, Deutsches Zentrum für Luft- und Raumfahrt (DLR), and the Chair of Operations Management.

Tasks

Prepare a scientific paper presenting the literature and state-of-the-art concerning (1) the use of drones in humanitarian actions and (2) heuristic algorithms for decision-making in emergency situations. Additionally, the scientific paper should present the characterization of the heuristic algorithms: information required (Data), decision variables, decision-making criteria (optimization criteria), and main constraints of the problem. The constraints have to consider the technical limitations of drones and drone use regulations. Evaluate the designed algorithms in self-chosen example scenarios

Organization

We will have personalized meetings with the working groups. The frequency of the meetings depends on the progress of the group. The scientific project has a maximum capacity
of 6 students.

Contact: Dr. Christoph Steup or Dr. Lorena Silvana Reyes Rubiano

Software-/Teamprojects: Competitive AI in Games

Every year the IEEE conference on games offers multiple competitions on AI in games for various levels of experience. In this project you can approach one of the more challenging competitions, which each pose their special challenges on your and the AI's skills. All of these competitions are tackling open research questions, which, depending on the competition you choose to approach, will be the focus of your project. The participation in the following competitions will be considered:

• Bot Bowl 4: https://github.com/njustesen/botbowl
  • Machine learning agents have struggled with the vast state and action space of this highly strategic turn-based tabletop game. Your focus will be the identification and effective usage of key skills to reduce the search time in a search-based agent. Solutions using reinforcement learning may be considered as well, but will require extensive amounts of training and therefore make use of our cluster.
• Dota 2 5v5 AI Competition: https://games.mau.se/research/the-dota2-5v5-ai-competition/
  • Dota 2 is one of the most prominent games of the MOBA genre. In this work, a whole team of agents needs to be programmed to destroy the opponent's ancient as fast as possible. The game requires to react fast and cooperate well. Training capabilities will be limited due to the way in which the framework works and the length of a game in general. Therefore, idenfitying key skills to be trained and used during the matches will be a special challenge.
• Tabletop Games Competition - Pandemic: http://www.tabletopgames.ai/
  • The cooperative board-game pandemic asks agents to work in a team of four players, of which each of them fills a different role. Pandemic is based on the premise that four diseases have broken out in the world, each threatening to wipe out a region. Through the combined effort of all the players, the goal is to discover all four cures before any of several game-losing conditions are reached. In this project, different types of cooperation shall be explored.
• Keke AI Competition: http://keke-ai-competition.com/
  • The concept of the game, which is better known as "Baba is you", demands the agent to modify and adapt the rules of the game. During the game the agent pushes blocks to change the rules of the game and find interest ways in solving clever puzzles. The high action space makes this game a challenging task for modern AI solutions. A recent paper on sokoban has shown interesting performance improvements. Making use of similar techniques and adapting them to this use-case will be a main-task in this work.

Contact: alexander.dockhorn@ovgu.de

Teamproject: Optimizing and Learning from Tool-assisted Speedrun

Tool-assisted Speedruns (TAS) is generally defined as speedrunning an emulated, slowed down (to allow precise inputs to be done with ease), and splicing (taking one good speedrun that had a mistake of otherwise and replacing parts with a different speedrun) with the goal of creating a theoretically perfect playthrough. At the same time they provide an invaluable training set which that links the state of the game to the user's inputs. In this project is the goal to either further optimize a TAS or to use it as a baseline for training an AI agent with similar capabilities.

Ressources:

• Automating Speedrun Routing: Overview and Vision: https://arxiv.org/pdf/2106.01182.pdf

Contact: alexander.dockhorn@ovgu.de

Software-/Teamproject: Learning a Playable Generative Model of a Game

In previous work, a neural network has been shown to mimic a game environment (Pac-Man) so closely, that the game can be nearly perfectly simulated by the neural network. Other works have shown that similar techniques can be used to simulate much more complex games (GAN Theft Auto). The project shall explore the limits of this approach and demonstrate a working prototype for playing a game in the internal model of a deep neural network.

Ressources:

• Learning to Simulate Dynamic Environments with GameGAN: https://nv-tlabs.github.io/gameGAN/
• GAN Theft Auto: https://github.com/Sentdex/GANTheftAuto

Contact: alexander.dockhorn@ovgu.de

Software-/Teamproject: Understanding the Hidden Activations of a Game-playing Neural Network

Many works have demonstrated that deep learning can be a promising tool for designing AI agents in games. However, the black-box nature often makes it hard to understand the agent's behavior during evaluation. In aim of this project is to relate a neural network's activations to the task that is currently performed for identifying which neurons correlate with the different aspects of the task. Similar techniques have been used in the field of image classification to understand the features a neuron react to.

Ressources:

• Feature Visualization: https://distill.pub/2017/feature-visualization/

Contact: alexander.dockhorn@ovgu.de

Additional Topics on request