prlearn 0.0.1

Parallel Reinforcement Learning library

PRLearn

PRLearn is a Python library designed for Parallel Reinforcement Learning. It leverages multiprocessing to streamline the training of RL agents, making it easier and more efficient to experiment and develop new RL approaches.

Key Features

• Simple and Flexible: Easy-to-use API built on Gymnasium, enabling seamless integration with existing environments.
• No Dependencies: No mandatory dependencies, but optional use of multiprocess for enhanced parallelism.
• Parallel Data Collection and Training: Efficiently collects and processes data using parallel execution, reducing training time.
• Agent Combination: Combines multiple agents to enhance learning outcomes through methods like averaging, boosting performance and stability.

Installation

Install PRLearn using pip:

pip install prlearn

or

pip install prlearn[multiprocess]

Quick Start

Here's a brief example to get you started with PRLearn:

Define Your Agent

from prlearn import Agent, Experience
from typing import Any, Dict, Tuple


class MyAgent(Agent):
    def action(self, state: Tuple[Any, Dict[str, Any]]) -> Any:
        observation, info = state
        # Define action logic
        pass

    def train(self, experience: Experience):
        obs, actions, rewards, terminated, truncated, info = experience.get()
        # Define training logic
        pass

Using Parallel Training

import gymnasium as gym
from my_best_model import BestAgent
from prlearn import Trainer
from prlearn.collection.agent_combiners import FixedStatAgentCombiner

# Define your environment and agent
env = gym.make("LunarLander-v2")
agent = BestAgent()

# Create and configure the trainer
trainer = Trainer(
    agent=agent,
    env=env,
    n_workers=4,
    schedule=[
        ("train_finish", 1000, "episodes"),
        ("train_agent", 10, "episodes"),
    ],
    mode="parallel_learning",  # optional
    sync_mode="sync",  # optional
    combiner=FixedStatAgentCombiner("mean_reward"),  # optional
)

# Run the trainer
agent, result = trainer.run()

• Environment: We use the LunarLander-v2 environment from Gymnasium.
• Agent: BestAgent is a custom agent class you should define.
• Trainer Configuration: The trainer is configured with 4 parallel workers, a schedule that specifies training completion after 1000 episodes and agent training every 10 episodes.
• Optional parameters include the mode parallel_learning, synchronization mode sync, and a combiner FixedStatAgentCombiner that averages agent rewards.

License

Licensed under the MIT License.