allensmile / rl

torch_rl : A starting Reinforcement Learning PyTorch Library

Torch_rl is a (small) Sequential Learning (and Reinforcement Learning) library for PyTorch that can be used to:

• learn policies over openAI Gym environments
• learn policies with DQN or Policy Gradient techniques (more to come)
• Model complex environements (not only reward-based environment)

Note that the library will evolve during the next months.

Gerating Documentation

Go to the docs directory then: PYTHONPATH=.. make html. The generated HTML file are under _build.html/index.html

Key packages

• the torch_rl.core package contains core classes. A classical RL environment is modeled using a triplet:

  • World: describes the physics of the world
  • Sensor: describes a (partial) view over the world.
  • Task: describes the task to solve. It can basically be defined through a Reward function

• the Env class is used to cast a triplet (World,Sensor,Task) to an openAI Gym Environment

• the torch_rl.core.sensors propose some basic tools for building high-level sensors from other sensors

• the torch_rl.core.spaces propose complex action and observation spaces

• the torch_rl.policies package contains policies definitions

  • Policy is the classical (non-learning) definition of a policy.

• the torch_rl.learners contains learning algorithms

  • DeepQN
  • PolicyGradient
  • BatchPolicyGradient (where multiple trajectories are sampled simultaneously)
  • RecurrentPolicyGradient
  • (...more to come...)

• the tutorials directory contains simple examples