evgenii-nikishin/pg_pytorch

Implementation of Policy Gradient methods for Reinforcement Learning

A library with implementations of popular Policy Gradient methods for Reinforcement Learning

Includes the following algorithms:

• A2C
• TRPO
• ACKTR

The code is organized as follows:

• src/agent.py: implementations of Agent.
• src/env_wrappers.py: wrapper for multiple parallel environments
• src/kfac.py: implementation of K-FAC optimizer, compatible with torch.optim.Optimizer
• src/networks.py: neural network architectures of actor and critic for different environments
• src/trpo.py: implementation of TRPO-optimizer routines
• src/utils.py: utils for models and optimizers

TODO:

• effective vectorization with n-step returns
• PPO
• A3C

---

Code is developed and supported by:

• Evgenii Nikishin nikishin-evg (nikishin.evg@gmail.com)
• Iurii Kemaev hbq1 (y.kemaev@gmail.com)
• Maxim Kuznetsov binom16 (binom16@gmail.com)

---

Inherited from https://github.com/nikishin-evg/acktr_pytorch