Deep Reinforcement Learning for Incomplete Information Card Games (BlackJack and Poker)

Application of RL and DRL algorithms to strategic decision-making in Blackjack and Poker, exploring uncertainty and incomplete information.

This thesis explores reinforcement learning (RL) and deep reinforcement learning (DRL) techniques in Blackjack and Poker, two games of incomplete information that require complex strategy under uncertainty.

Methods include Deep Q-Networks (DQN), Counterfactual Regret Minimization (CFR), Deep Monte Carlo (DMC), and Neural Fictitious Self-Play (NFSP), evaluated for their ability to handle stochasticity, hidden information, and strategic adaptation.

Key results:
- DQN learns structured Blackjack strategies effectively
- CFR and NFSP excel in Poker, adapting to incomplete information
- DMC achieves robust convergence in high-dimensional state spaces
- DRL methods demonstrate potential for superhuman performance in competitive card games

Read full-text on ResearchGate

Deep reinforcement learning applied to Blackjack and Poker