yyds-xtt/cdrl

Collaborative Deep Reinforcement Learning

Collaborative Deep Reinforcement Learning

Collaborative Deep Reinforcement Learning (CDRL) is a collaborative framework to enable knowledge transfer among heterogeneous tasks.

Dependencies

• Python 2.7
• six (for py2/3 compatibility)
• TensorFlow 0.11
• gym
• gym[atari]
• universe
• opencv-python
• numpy
• scipy

Folder structure

|>CDRL

The code for experiments in paper section 5.4 Collaborative Deep Reinforcement Learning

|>heterogeneousTransfer

The code for experiments in paper section 5.3 Certificated Heterogeneous Transfer

|>model

The pre-trained model used for section heterogeneous transfer

How to run:

|>heterogeneousTransfer

python train.py --num-MTL-workers 8 0 --env-id PongDeterministic-v3_Bowling-v0

|>CDRL

python train.py --num-MTL-workers 8 8 --env-id PongDeterministic-v3_Bowling-v0

Preprint:

Collaborative Deep Reinforcement Learning
Kaixiang Lin, Shu Wang and Jiayu Zhou

Acknowledgement

This code refers to OpenAI universe starter agent.