Deep Recurrent Q Networks

This repository demonstrates the implementation of Deep Recurrent Q-Networks (DRQN) for Partially Observable environments. Utilizing recurrent blocks with Deep Q-Networks can actually make the agent receiving single frames of the environment and the network will be able to change its output depending on the temporal pattern of observations it receives. DRQN does this by maintaining a hidden state that it computes at every time-step. Furthermore, A brief explaination of DRQNs for partial observability can be found here.

Requirements

• tensorflow
• numpy

Training Environment Used

• The Spatially Limited Gridworld Environment

In this new version of the GridWorld, the agent can only see a single block around it in any direction, whereas the environment itself contains 9x9 blocks. Each episode is fixed at 50 steps, there are four green and two red squares, and when the agent moves to a red or green square, a new one is randomly placed in the environment to replace it.

Usage

• To train a new network : run training.py
• To test a preTrained network : run test.py
• To see the DRQN implementation, please check Model.py
• Other imperative helper utilities to properly train the Deep Recurrent Q-Networks can be found in helper.py file.
• All hyperparameters to control training and testing of DRQNs are provided in their respective .py files.

Results

Gridworld Environment	Gridworld Results