Reinforcement Learning pro hraní karetních her

Abstract

In this thesis we will analyze different reinforcement learning methods in multi-player card games, characterized by action cards, higher randomness and complexity. We will divide them into three groups - game theory based, TD learning (DQN) and Monte Carlo methods (DMC). We will discuss applicability of game theory in this environment, implement DQN and DMC to represent the last two groups and analyze their behaviour. We will show that of the three groups, Monte Carlo methods are the only ones practically applicable for this type of environment.