StreamRL with SPR

Paper : Squeezing More from the Stream : Learning Representation Online for Streaming Reinforcement Learning

Evaluation of Self-Predictive Representations (SPR) combined with streaming reinforcement learning algorithms on Atari environments.

Overview

This repository implements and compares baseline streaming RL algorithms with their SPR-augmented variants:

• DQN / DQN-SPR: Deep Q-Network with optional SPR
• QRC / QRC-SPR: Quick Replay Candidates with optional SPR
• StreamQ / StreamQ-SPR: Streaming Q-learning with optional SPR

Installation

pip install -r requirements.txt

Usage

Run an algorithm on an Atari environment:

python algorithms/dqn.py --env_type atari --env_id ALE/Pong-v5