by571 / normalized-advantage-function-naf- Goto Github PK

View Code? Open in Web Editor NEW

PyTorch implementation of the Q-Learning Algorithm Normalized Advantage Function for continuous control problems + PER and N-step Method

License: MIT License

Jupyter Notebook 89.74% Python 10.26%

normalized-advantage-function-naf-'s Introduction

Normalized Advantage Function (NAF)

PyTorch implementation of the NAF algorithm based on the paper: Continuous Deep Q-Learning with Model-based Acceleration.

Two versions are implemented:

Jupyter notebook version
Script version (results tracking with wandb)

Recently added PER and n-step method

To run the script version: python naf.py

with the arguments:

'-env' : Name of the environment (default: Pendulum-v0)
'-info' : Name of the Experiment (default: Experiment-1)
'-f', --frames : Number of training frames (default: 40000)   
'-mem' : Replay buffer size (default: 100000)
'-b', --batch_size : Batch size (default: 128)
'-l', --layer_size : Neural Network layer size (default: 256)
'-g'--gamma : Discount factor gamma (default: 0.99)
'-t', --tau : Soft update factor tau (default: 1e-3)
'-lr', --learning_rate : Learning rate (default: 1e-3)
'-u', --update_every : update the network every x step (default: 1)
'-n_up', --n_updates : update the network for x steps (default: 1)
'-s', --seed : random seed (default: 0)
'-per', choices=[0,1] : Use prioritized experience replay (default: 0)
'-nstep' : nstep_bootstrapping (default: 1)
'-d2rl': Using Deep Dense Network if set to 1 (default: 0)
'--eval_every': Doing an evaluation of the current policy every X frames (default: 1000)
'--eval_runs': Number of evaluation runs - performance is averaged over all runs (default: 3)

In the paper they compared NAF with DDPG and showed faster and more stable learning: We show that, in comparison to recently proposed deep actor-critic algorithms, our method tends to learn faster and acquires more accurate policies.

To verify and support their statement I tested NAF on Pendulum-v0 and LunarLanderConinuous-v2 and compared it with the results of my implementation of DDPG.

The results shown do not include the model-based acceleration! Only the base NAF algorithm was tested.

Indeed the results show a faster and more stable learning!

TODO:

Test with Double Q-nets like SAC
Test with Entropy Regularization (like sac)
Test with REDQ Q-Net ensemble

Feel free to use this code for your own projects or research:

@misc{Normalized Advantage Function,
   author = {Dittert, Sebastian},
   title = {PyTorch Implementation of Normalized Advantage Function},
   year = {2020},
   publisher = {GitHub},
   journal = {GitHub repository},
   howpublished = {\url{https://github.com/BY571/NAF}},
 }

normalized-advantage-function-naf-'s People