Work

Applying AlphaZero Algorithm to an Imperfect Information Card Game

Year: 2026
Type: dissertation
Abstract: Reinforcement learning is a machine learning paradigm where the learning system does not require labelled training data, it learns by experimentation and environment feedback. Google DeepMind’s AlphaGo demonstrated that reinforcement learning can beat a human champion in Go. It used recorded human expert games for supervised learning. DeepMind’s AlphaZero was trained tabula rasa, without human knowledge. The method used self-play with Monte Carlo Tree Search to improve the suggestions from network. AlphaZero's differentiating property compared to AlphaGo is bootstrapping from games played without domain understanding. This thesis studies the possibility to apply AlphaZero-inspired approach to learn to play games of imperfect information. An important part of the approach is to evaluate model goodness without human intelligence. In this work I created an evaluation system using classic minimax, fixed strategy baselines and games against older versions of the model. Results were promising and surprisingly clear. Firstly, the system was shown to learn and to beat trivial strategies and to beat older versions of the model. Secondly, it was shown and measured how much the imperfect information affected the learning potential. It was shown that 73% learning from the absolute maximum was reached and only 7% of the learning potential was left undone. The rest 20% is considered to be behind so called "information barrier", caused by the imperfect information. Learning that 20% would require more advanced methods and those are discussed in the thesis. Thirdly, a methodological approach for decomposition of learning progress was created. Fourth, the created learning and evaluation system is adaptable to other games or to measuring the effect of different amount of hidden information. The covered theoretical background includes reinforcement learning and AlphaZero, probability-related search methods, Markov Decision Process and exploration-exploitation trade-off. Further study, applications and approach limitations were considered, showing interesting directions for future work.
Author: Krappe, Sami
Institution: University of Turku
Language: English
Open Access status: green
OpenAlex ID: W7220817904
Cites: 24
Cited by:
Related to: 10