PhD Defence Thomas (Tom) Hubertus Judith Pepels

Keywords: Monte-Carlo Tree Search, Game AI, Regret Minimization, Real-Time Decision Making 
 

"Monte-Carlo Tree Search is Work in Progress"

By learning from games, computers make better choices. Computers need to make smart choices, and Tom Pepels investigates how they can do so. One method is the Monte Carlo Tree Search (MCTS) that allows computers to make smart choices in games by trying out different moves at random and learning from the results.  

Tom’s main goal is to make MCTS more flexible and useful in many different game situations from slow, turn-based board games to fast, real-time video games. His research explores new ideas to make MCTS stronger. One idea is to mix MCTS with another method called minimax, which helps in two-player games where one player wins and the other loses. It also adds smart ways to guess future results more accurately. Another improvement is using better reward systems to help the computer understand which moves are good, by giving extra points for better choices. He also presents a new version of MCTS called Hybrid MCTS, which balances trying out new moves with sticking to known good ones. Lastly, it shows how to adjust MCTS for real-time games like Ms. Pac-Man by using deeper planning, smarter move testing, long-term goals, and reusing past search efforts. Tests show that these changes help MCTS perform better in many games. This work shows how MCTS can be adapted to different game styles and why it’s important to adjust it depending on the game. The research contributes to a deeper understanding of MCTS and its potential for broader application in all kinds of complex decision-making scenarios. 

Click here for the live stream.