Skip to content

Latest commit

 

History

History
47 lines (29 loc) · 3.58 KB

README.md

File metadata and controls

47 lines (29 loc) · 3.58 KB
Raw

Basic Ray RLlib Project Skeleton

This repository provides a structured, modular setup for training Reinforcement Learning (RL) models using the Ray RLlib library. It accommodates environments compliant with the OpenAI Gym API, enabling you to work with a vast array of environments out-of-the-box. The project centers on a registry-based design, allowing extensibility, and an elegant plug-and-play style configuration for RL training setups.

Project Structure:

.
├── basic_project
│   ├── common                # Module containing commonly used constants and methods
│   ├── config                # Module containing configuration files (main config and trainable configurations)
│   ├── environment           # Module containing custom environment implementation along with creator function
│   ├── delegator             # Main module containing classes responsible for delegating tasks
│   │   ├── abstract          # Submodule containing abstract base classes
│   │   ├── construct         # Submodule containing different RL algorithm constructs
│   ├── tune.py               # Main script to start the training process

Key Features:

  • Environment Activation: Activate the working environment using a shell script. This script sets up and triggers the environment required to run the project.

  • OpenAI Gym Environment Compatibility: The framework accommodates environments adhering to the OpenAI Gym API.

  • Modular RL Algorithm Configuration: Specify the RL algorithm for training as a separate construct. This approach encourages modularity and code reuse. Constructs for RL algorithms can be found in delegator/construct and can be easily extended with additional algorithms.

  • Nested Hyperparameter Configuration: Define and adjust the hyperparameters for the RL algorithm via the YAML configuration files located in the config module. Different configuration files can be created for various training scenarios.

  • Training Control: Define stopping conditions and checkpointing policies for the training process through the main configuration file.

  • Hyperparameter Optimization: Integration with Ray Tune for hyperparameter optimization of RL models.

  • Registry-Based Design: The registry-based design handles multiple training algorithms, promoting extensibility and a modular design. New algorithm constructs are registered in the ConstructRegistry.

Getting Started:

  1. Activate Environment: Use the following command to set up and activate the necessary environment for running the project:

source activate_env.sh

  1. Create/Modify Configuration Files: Provide a YAML configuration file detailing the training parameters under config. This includes specifying the RL algorithm construct to use and its associated hyperparameters.

  2. Run Training: Execute the tune.py script. The script reads the configuration files, initializes the appropriate delegator classes with the loaded configuration, applies the parameters, and commences the RL training process.

  3. Extend with New Algorithms: When adding new RL algorithms to the project, ensure to register the new constructs in the ConstructRegistry and follow the established conventions for naming and file structure. The new construct should implement the 'from_construct_registry_directive' class method. The constructs are Python classes representing RL algorithms, located in the delegator/construct directory.

This project provides a flexible and extendable base for RL experimentation with the Ray RLlib library. Happy experimenting!