Optimizing Deep Reinforcement Learning Through Vectorized and Parallel NeuroEvolution

Hesham Magdy

doi:10.1016/j.future.2025.108334

Return

Optimizing Deep Reinforcement Learning Through Vectorized and Parallel NeuroEvolution

2025-12-22

PRE

Hesham Magdy

Amany Sarhan *

Mohammad Ali Eita

DOI：10.1016/j.future.2025.108334

Original

Original request for help

Save

Abstract

En 中文

Deep reinforcement learning (DRL) has achieved remarkable success in solving complex decision-making problems; however, training efficiency, scalability, and ease of implementation remain significant challenges. It is evident that the advances in hardware accelerators such as GPUs and TPUs can reduce the time obstacle that faces the development in the deep learning field. As a novel approach, we propose a framework for NeuroEvolution-based deep reinforcement learning, leveraging JAX, a high-performance numerical computing library optimized for machine learning and automatic differentiation, for efficient parallel execution. Our framework enables seamless vectorized policy optimization, significantly reducing computational overhead while maintaining sample efficiency. We introduce a user-friendly interface designed for accessibility and flexibility, allowing researchers to easily experiment with diverse evolutionary strategies, leading to broader exploration and improved performance in certain environments. Additionally, we extend the application of NeuroEvolution-based DRL to environments that have not been previously explored using such methods, further demonstrating the versatility of our approach. Finally, we incorporate recent evolutionary algorithms into the training process, achieving better results. Through extensive benchmarking, we show that our framework outperforms traditional evolutionary strategies and gradient-based DRL methods in both convergence speed and scalability, achieving a speedup of up to 34 × compared to state-of-the-art approaches.