Deep Reinforcement Learning Approaches for the Game of Briscola

Reinforcement learning is increasingly becoming one of the most interesting areas of research in recent years. It is a machine learning approach that aims to design autonomous agents capable of learning from interaction with the envi- ronment, similar to how a human does. This peculiarity makes it particularly suitable for sequential decision making problems such as games. Indeed games are a perfect testing ground for reinforcement learning agents, due to a con- trolled environment, challenging tasks and a clear objective. Recent advances in deep learning allowed reinforcement learning algorithms to exceed human level performance in multiple games, the most notorious example being AlphaGo. In this thesis work we will apply deep reinforcement learning methods to Briscola, one of the most popular card games in Italy. After formalizing the two-player Briscola as a RL problem, we will apply two algorithms: Deep Q-learning and Proximal Policy Optimization. The agents will be trained against a random agent and an agent with predefined moves. The win rate will be used as a performance measure to compare the final results.