Deep Reinforcement Learning for Game Bots [multiple projects: must be specialised]

Yuzhi Wei


Supervised by Frank C Langbein; Moderated by Juan Hernandez Vega

AI bots

in games generally perform quite poorly compared to human performance, without employing tricks to improve their perceived performance. The aim of this project is to investigate reinforcement learning approaches to train human-level game bots, utilising the same or at least similar information than a human gets from the game. You are free to choose the game you wish to try this on and the particular reinforcement learning approach you wish to test (see Doom/VizDoom, OpenAI/DeepMind examples, others...). This is a difficult project, requiring advanced maths and programming skills, and a strong understanding of reinforcement learning with deep learning. You will also need to have access to suitable hardware, in particular a GPU (or more, depending on approach), where we can only provide quite limited resources via university. When chosen the game make sure it is suitable for doing this without needed complex setups (even if one option would also be to explore a setup that could play a game with a video feed, and a USB connection for inputs with limited requirements on the actual AI; discuss with me, if you are interested). Which game to choose must be discussed with me and decided during selection.

Final Report (08/10/2023) [Zip Archive]

Publication Form