minirl

基于Numpy的深度强化学习

---

这个 repo 包含一组从头开始使用 numpy 实现的用于强化学习的应用程序和算法。包括的算法q-learning、基于深度神经网络的REINFORCE、Actor-Critic和ppo等。

项目结构

.
├── core
    ├── bandit.py               # EpsilonGreedy/UCB/LinUCB/ThompsonSampling algorithm
    ├── smab.py                 # stochastic Multi-Armed Bandit (sMAB)
    ├── cmab.py                 # contextual Multi-Armed Bandit (cMAB) based on Thompson Sampling
    ├── onlineCluster.py        # online k-means using Lloyd's algorithm
    ├── pg.py                   # REINFORCE algorithm
    ├── deep_q_learning.py      # Deep Neural Network based Q-learning
    ├── ac.py                   # Actor-Critic algorithm
    ├── ppo.py                  # Proximal Policy Optimization
    ├── DynaQ.py                # Dyna-Q algorithm
    ├── DynaQ_plus.py           # Time-based model for planning in Dyna-Q+
├── preprocessing
    ├── feature_transformer.py  # OneHotEncoder/TargetEncoder
    ├── scaler.py               # StandardScaler/MinMaxScaler/MaxAbsScaler
    ├── stats.py                # runningReward
├── common                      
    ├── net.py                  # Common-deep-network
    ├── optim.py                # Optimizer        
└── README.md

技术架构