Welcome

This is my personal practice of implementing various algorithms of RL from scratch.

Most of them will be in jupyter notebook and some of them involving multiprocess would be in normal python files.

The framework will always be PyTorch, as a personal practice too.

Normally I use cartpole for easy algorithms in this project and I skip the visual input part. (which is quite trivial if you add few conv layers).

And for harder and visual-related algorithms I will pick various atari game as my environment.

Due to time limit, I will not provide systematic analysis to any particular algorithm. And be aware these are personal usage so bugs do appear frequently.

If the project is mature, I will accept open issues. For now, however, let me dive in. (I guess no one even read this repo though)

Project file structure will be changed continuously to match my needs.

PLAN:

Model-Free RL

Policy Gradient

• [x] REINFORCE
• [x] Off-Policy REINFORCE
• [x] Basic Actor Critic
• [x] Advantage Actor Critic using Huber loss and Entropy
• [x] A3C
• [x] A2C
• [x] DDPG
• [ ] D4PG
• [ ] MADDPG
• [ ] TRPO
• [ ] PPO
• [ ] ACER
• [ ] ACTKR
• [ ] SAC
• [ ] SAC with AAT(Automatically Adjusted Temperature
• [ ] TD3
• [ ] SVPG
• [ ] IMPALA

Deep Q Learning

• [X] Dueling DDQN
• [x] Dueling DDQN + PER
• [ ] Rainbow DQN
• [ ] Ape-X

Distributed RL

• [ ] C51
• [ ] QR-DQN
• [ ] IQN
• [ ] Dopamine (DQN + C51 + IQN + Rainbow)

Policy Gradient with Action-Dependent Baselines:

• [ ] Q-prop
• [ ] Stein Control Variates

Path-Consistency Learning

• [ ] PCL
• [ ] Trust-PCL

Q-learning + Policy Gradient:

• [ ] PGQL
• [ ] Reactor
• [ ] IPG

Evolutionary Algorithm

Monte Carlo Tree (Alpha Zero)

Exploration RL

Intrinsic Motivation

• [ ] VIME
• [ ] CTS-based Pseudocounts
• [ ] PixelCNN-based Pseudocounts
• [ ] Hash-based Counts
• [ ] EX2
• [ ] ICM
• [ ] RND

Unsupervised RL

• [ ] VIC
• [ ] DIAYN
• [ ] VALOR

Hierachy RL

Memory RL

Model-Based RL

Meta-RL

Scaling-RL