Algorithms¶
Deep-Q Learning¶
Deep-Q Learning (DQN) [paper] is a basic reinforcement learning (RL) algorithm. We wrap DQN as an example to show how RL algorithms can be connected to the environments. In the DQN agent, the following classes are implemented:
DQNAgent: The agent class that interacts with the environment.Normalizer: The responsibility of this class is to keep a running mean and std. The Normalizer will first preprocess the state before feeding the state into the model.Memory: A memory buffer that manages the storing and sampling of transitions.Estimator: The neural network that is used to make predictions.
NFSP¶
Neural Fictitious Self-Play (NFSP) [paper] end-to-end approach to solve card games with deep reinforcement learning. NFSP has an inner RL agent and a supervised agent that is trained based on the data generated by the RL agent. In the toolkit, we use DQN as RL agent.
CFR¶
Counterfactual Regret Minimization (CFR) [paper] is a regret minimizaiton method for solving imperfect information games.
DeepCFR¶
Deep Counterfactual Regret Minimization (DeepCFR) [paper] is a state-of-the-art framework for solving imperfect-information games. We wrap DeepCFR as an example to show how state-of-the-art framework can be connected to the environments. In the DeepCFR, the following classes are implemented:
DeepCFR: The DeepCFR class that interacts with the environment.Fixed Size Ring Buffer: A memory buffer that manages the storing and sampling of transitions.