akashe/deepreinforcementlearning

Deep RL implementations. DQN, SAC, DDPG, TD3, PPO and VPG implemented in pytorch. Tested Env: LunarLander-v2 and Pendulum-v0.

Deep RL algorithms implemented using Pytorch

Algo list:

1. DQN
2. Vanilla policy Gradient
3. Deep Deterministic Policy Gradient
4. Twin Delayed Deep Deterministic Policy Gradient
5. Soft Actor Critic
6. Proximal Policy Optimization - CLIP

Article on deeper Look into policy gradients

Experimental Results:

Algorithm	Discrete Env: LunarLander-v2	Continuous Env: Pendulum-v0
DQN		-
VPG		-
DDPG	-
TD3	-
SAC	-
PPO	-

Usage:

Just run the file/algorithm directly. There is no common structures between algorithms as I implemented them as I learnt them. Different algorithms are inspired from different sources.

Resources:

1. RL course by David Silver
2. Lecture slides for above course
3. Spinning up by OpenAI
4. More exhaustive RL guide by Deeny Britz

Future projects:

1. If time available I will add a simple program for elevator using RL.
2. Better graphs

deepreinforcementlearning's Issues

SAC DDPG

It seems that DDPG performs the best in the simulation. However, according to some papers, it doesn't present well, how does this happen?