tw-policies is a simple library/tool for Reinforcement Learning, with a focus in the policy gradient class of algorithms. Notably, it contains numpy-only implementations of the policy gradient for discrete and guassian linear function approximations. It contains several premade algorithm implementations using this framework, such as the REINFORCE algorithm and several Actor-Critic algorithms like A2C and PPO.
You can find the networks and inference code for the policies and state-value evaluation code in the two main modules models.policy which contains the most essential component: the function approximations for different kinds of policies. and models.value. You can find premade algorithm implementations in the agents. Some feed-forward neural networks are available in networks.
- test PPO implementation
- Support offline training
- Multi-Headed network outputs
- Goal-conditioned training
- Shared weights feature extraction for Actor-Critic
- automated tuning?