Model Free Algorithms Save

Status. Archive (code is provided as-is, no updates expected)

Note. Please refer to my new repo for reinforcement learning algorithms in TF2.x

Algorithms Implemented

Algorithms are implemented in algo.

• TD3 [code]
• SAC [code]
• IQN [code]
• Rainbow [code]
• Apex [code]
• Noisy Nets[code]
• PER [code]
• MultiStep [code]]
• PPO [code]
• A2C [code]
• GAE [code]
• NAE [code]

Overall Architecture

This repository is designed to represent a nice Tensorboard graph, which is useful for debugging.

A typical graph looks like this:

Notes

Distributed Algorithms are implemented using Ray, a flexible, high-performance distributed execution framework.

Due to the lack of a Mujoco license, all algorithms for continuous control are first tested on the LunarLanderContinuous-v2 and later on BipedalWalker-v2 environments from OpenAI's Gym and solve them.

Rainbow, IQN is tested on CartPole-v0 and and steadily solves it. For Rainbow and IQN on Atari games, please refer to my another project

Performance figures and some further experimental results are recorded in on-policy algorithms and off-policy algorithms.

Best arguments are kept in "args.yaml" in each algorithm folder. If you want to modify some arguments, do not modify it in "args.yaml". It is better to first pass the experimental arguments to gs defined in run/train.py to verify that they do improve the performance.

Requirements

It is recommended to install Tensorflow from source following this instruction to gain some CPU boost and other potential benefits.

# Minimal requirements to run the algorithms. Tested on Ubuntu 18.04.2, using Tensorflow 1.13.1.
conda create -n gym python
conda activate gym
pip install -r requirements.txt
# Install tensorflow-gpu or install it from scratch as the above instruction suggests
pip install tensorflow-gpu

Running

# Silence tensorflow debug message
export TF_CPP_MIN_LOG_LEVEL=3
# When running distributed algorithms, restrict numpy to one core
# Use numpy.__config__.show() to ensure your numpy is using OpenBlas
# For MKL and detailed reasoning, refer to [this instruction](https://ray.readthedocs.io/en/latest/example-rl-pong.html?highlight=openblas#the-distributed-version)
export OPENBLAS_NUM_THREADS=1

# For full argument specification, please refer to run/train.py
python run/train.py -a sac

To add monitor so as to save videos automatically, set argument log_video to True in args.yaml.

Paper References

Timothy P. Lillicrap et al. Continuous Control with Deep Reinforcement Learning

Matteo Hessel et al. Rainbow: Combining Improvements in Deep Reinforcement Learning

Marc G. Bellemare et al. A Distributional Perspective on Reinforcement Learning

Scott Fujimoto et al. Addressing Function Approximation Error in Actor-Critic Methods (TD3)

Tuomas Haarnoja et al. Soft Actor-Critic: Off-Policy Maximum Entropy Deep Reinforcement Learning with a Stochastic Actor.

Tuomas Haarnoja et al. Soft Actor-Critic Algorithms and Applications

Dan Horgan et al. Distributed Prioritized Experience Replay

Hado van Hasselt et al. Deep Reinforcement Learning with Double Q-Learning

Tom Schaul et al. Prioritized Experience Replay

Meire Fortunato et al. Noisy Networks For Exploration

Ziyu Wang et la. Dueling Network Architectures for Deep Reinforcement Learning

Will Dabney et al. Implicit Quantile Networks for Distributional Reinforcement Learning

Berkeley cs294-112

Code References

OpenAI Baselines

Homework of Berkeley CS291-112

Google Dopamine