Model-Based Reinforcement Learning

This repository provides a tensorflow implementation of Model-based reinforcement learning for CartPole environment. This repo also aims at helping beginners in Reinforcement Learning to better understand how to train the agent using the model rather than requiring to use the real environment every time in order to learn the dynamics of the real environment space. A concise explaination of Model-based reinforcement learning algorithm can be found here.

Requirements

• tensorflow
• matplotlib
• numpy

Training Environment Used

• OpenAI Gym's CartPole-v1 Environment

Usage

• Model_based_RL.py file contains code statements related to Model Network and Policy Network that are being used here.
• The training procedure involves switching between training the Model Network using the real environment, and training the agent’s policy using the model environment. By using this approach, the Model Network will be able to learn a policy that allows the trained agent to solve the CartPole task without actually ever training the policy on the real environment.
• All hyperparameters to train the two networks can be found in Model_based_RL.py.
• The episode number, action, reward and mean reward are printed after very epoch during training time.

Results

CartPole-v1 Environment	CartPole-v1 Results