stefanbo92 / a3c-continuous

Tensorflow implementation of the asynchronous advantage actor-critic (a3c) reinforcement learning algorithm for continuous action space

A3C Continuous Reinforcement Learning

Tensorflow implementation of the asynchronous advantage actor-critic (A3C) reinforcement learning algorithm (paper) for continuous action space. Code is mostly based on Morvan Zhou (github).

Components

• ACNet: This class contains the actor-critic neural network that estimates an action given a certain state and a value for each state. For continuous action states the action is given as an expected value mu and variance sigma.
• Worker: The A3C algorithm employs multiple workers which have their own environment and ACNet and train on these asynchronous. Every few steps they update their weights to the global ACNet.
• Main: The main function creates the global ACNet and multiple workers. They start training until a defined number of training episodes is reached. Reward will be plotted over all steps.

Results

Pendulum environment before training:

After 1500 episodes: