README.md

Git Branches

master branch: Original DQN implemented in this repo
knowledge_distill branch: Implementation of policy distillation method.
multi_dpn branch: Implementation of multi-task DQN and joint training with policy distillation method.

Running Environment

Python 3.6 && Pytorch 0.4.0

Training and Testing

Downloading pretrained weights from here. Unzip it and put it in the root directory of the project.
Train: Run sh train.sh in root directory of the project.
Test: Run sh test.sh in root directory of the project.
The trained model weights are stored in saved_agents/ directory.
The training and testing logs are store in log/ directory.