Deep Deterministic Policy Gradient (DDPG) with Prioritized Experience Replay

Authors: Luca Iezzi and Giulia Ciabatti.

This consists of a complete reimplementation of DDPG with PrioritizedExperience Replay, and its adaptation on Pendulum-v1 and MountainCarContinuous-v0, from OpenAI Gym.

Agent playing

Usage

This implementation is based on Python 3.8 and PyTorch Lightning. To install all the requirements:

$ pip install -r requirements.txt

Hyperparameters

PER_BUFFER	ACTOR	CRITIC
hyperparameter value WARM_POPULATE 10000 ALPHA 0.6 BETA 0.1 PRIORITIZED_REPLAY_ALPHA size 0.6 PRIORITIZED_REPLAY_BETA0 size 0.4 PRIORITIZED_REPLAY_BETA_ITERS None PRIORITIZED_REPLAY_EPS rate 1e-6 BATCH_SIZE 64 Latent size 64 EPISODES 150	hyperparameter value OU_NOISE_STD 0.8 OPTIMIZER Adam LEARNING RATE 1e-4 GAMMA 0.99 TAU 5e-3	hyperparameter value OPTIMIZER Adam LEARNING RATE 5e-4 GAMMA 0.99 TAU 5e-3

Results

Running

The complete pipeline to train the 3 model components:

1. Train the agent

In simple_config.py, set ENV=[gym env you want to train on], set TRAIN=True and run:

$ python main.py

1. Test the agent

In main.py, manually copy the path of one of the checkpoints in ckpt/ in the variable model, set TRAIN=False and RENDER=True in simple_config.py and run:

$ python main.py

Credits

Some of the implementations (e.g. SumTree and MinTree) have been taken by existing repos, but it's been impossible to track the original author :( . If you recognize your code there, don't hesitate to drop me an email and I will add your repo to the credits!