tianshou

Tianshou(天授) is a reinforcement learning platform. The following image illustrate its architecture.

agent

    Examples

    Self-play Framework

core

Model

    DQN, Policy-Value Network of AlphaGo Zero, PPO-specific, TROP-specific

Algorithm

Loss design

    Actor-Critic (Variations), DQN (Variations), DDPG, TRPO, PPO

Optimization method

    SGD, ADAM, TRPO, natural gradient, etc.

Planning

    MCTS

data

    Training style - Monte Carlo or Temporal Difference

    Reward Reshaping/ Advantage Estimation Function

    Importance weight

    Multithread Read/Write

environment

    DQN repeat frames etc.

simulator

    Go, Othello/Reversi, Warzone

TODO

Search based method parallel.