/policy_gradient

Primary LanguagePythonMIT LicenseMIT

Introduction

Python implementation of policy gradient reinforcement learning methods, built on TensorFlow.

Flexible reinforcement learning framework that does not require dynamics of the environment.

Given a sequence of state-action pairs and rewards, adjusts policy based on gradient steps with respect to its parameters.

Interface designed to minimize the amount of developer interaction needed.

Installation

Quick installation using the pip package manager.

pip install -i https://testpypi.python.org/pypi policy_gradient

Interface

Quick, clean interface for training a neural network using policy gradient methods.

Designed to minimize the amount of interaction with the interface.

Just initialize the learner, pass in the dynamics/rewards functions and the initial state, and you're good to go!

# Initialize learner
learner = PolicyGradient(net_dims, 'tanh')
# Train policy
learner.train_agent(dynamics_func, reward_func, initial_state)
# Retrieve predicted best actions based on learned policy
learner.get_action(new_state)

Documentation

Full Documentation