LGNEApr 17, 2025

Evolutionary Policy Optimization

arXiv:2504.12568v11 citationsh-index: 56GECCO Companion
Originality Incremental advance
AI Analysis

This addresses the exploration-exploitation problem in reinforcement learning for tasks requiring both global exploration and local optimization, though it is incremental as it integrates existing methods.

The paper tackled the exploration-exploitation trade-off in reinforcement learning by proposing Evolutionary Policy Optimization (EPO), a hybrid algorithm combining neuroevolution and policy gradient methods, which improved policy quality and sample efficiency on Atari benchmarks like Pong and Breakout.

A key challenge in reinforcement learning (RL) is managing the exploration-exploitation trade-off without sacrificing sample efficiency. Policy gradient (PG) methods excel in exploitation through fine-grained, gradient-based optimization but often struggle with exploration due to their focus on local search. In contrast, evolutionary computation (EC) methods excel in global exploration, but lack mechanisms for exploitation. To address these limitations, this paper proposes Evolutionary Policy Optimization (EPO), a hybrid algorithm that integrates neuroevolution with policy gradient methods for policy optimization. EPO leverages the exploration capabilities of EC and the exploitation strengths of PG, offering an efficient solution to the exploration-exploitation dilemma in RL. EPO is evaluated on the Atari Pong and Breakout benchmarks. Experimental results show that EPO improves both policy quality and sample efficiency compared to standard PG and EC methods, making it effective for tasks that require both exploration and local optimization.

Foundations

The foundational work for this paper's niche, ranked by how specifically the neighbourhood builds on it — not by global fame.

Your Notes