rl algos

Collection of (off-policy) rl algorithms. Fully compatible with OpenAI gym.

Real-time monitoring of training done with visdom.

DDPG (not distributed)

Deep Learning extension of deterministic policy gradients (DPG), an off-policy RL algorithm. My implementation uses action and parameter noise to improve exploration at the start of training and then throughout the remainder of the steps.

TD3 (not distributed)

In progress - DDPG with tweaks to counter the tendency of DDPG to overestimate Q-function later during learning. Also uses action and parameter noise to improve exploration

D4PG (In progress, distributed)

In progress - DDPG with Ape-X framework (using ray for this) and PER

D4PG + TD3 (In progress, distributed)

This is an implementation of DDPG in Pytorch with action and parameter noise for exploration.

TODO

• Implement prioritized experience replay
• Implement Ape-X distributed training framework with Ray
• Fix visdom logging
• Clean up unneeded code.
• Integrate into https://github.com/yeshg/deep-rl

Acknowledgements

Code structure, visdom logging from on https://github.com/p-morais/deep-rl

Basic implementations of DDPG and TD3 from official TD3 release repo: https://github.com/sfujim/TD3