发表机构
University of Thessaly; University of Glasgow; Trinity College Dublin（色萨利大学; 格拉斯哥大学; 都柏林圣三一学院）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
GPlaceRL是一个开源图强化学习框架，通过GAT编码器和PPO策略在详细布局细化中实现HPWL改进3.27%至32.87%，强调紧凑架构与灵活动作空间的重要性。
AI 中文摘要
强化学习（RL）已成为布局优化的一种有前景的方法，尤其是与捕捉电路连接性的图神经网络（GNN）相结合时。然而，大多数基于学习的布局方法侧重于平面规划、宏单元布局或全局布局，而详细布局细化仍相对未被探索。在本文中，我们提出了GPlaceRL，一个用于详细布局细化的开源图强化学习框架。GPlaceRL将合法化布局表示为图，并提供了一个模块化环境，用于研究图编码器、策略架构、奖励公式和局部布局动作。为了展示GPlaceRL的能力，我们在逐设计优化设置中对使用图注意力网络（GAT）编码器的近端策略优化（PPO）策略进行了系统评估。在五个布局基准上，最佳贪婪评估结果实现了HPWL改进，范围从3.27%到32.87%。结果突出了紧凑的GAT架构和灵活的局部动作空间对布局优化的重要性。总体而言，GPlaceRL为基于RL的详细布局细化的系统研究提供了一个可复现且可扩展的框架。
英文摘要
Reinforcement learning (RL) has emerged as a promising approach for placement optimization, particularly when combined with graph neural networks (GNNs) that capture circuit connectivity. However, most learning-based placement approaches focus on floorplanning, macro placement, or global placement, while detailed placement refinement remains relatively unexplored. In this paper, we present GPlaceRL, an open-source graph reinforcement learning framework for detailed placement refinement. GPlaceRL represents legalized placements as graphs and provides a modular environment for studying graph encoders, policy architectures, reward formulations, and local placement actions. To demonstrate the capabilities of GPlaceRL, we conduct a systematic evaluation of proximal policy optimization (PPO) policies with graph attention network (GAT) encoders in a per-design optimization setting. Across five placement benchmarks, the best greedy evaluation results achieve HPWL improvements ranging from $3.27\%$ to $32.87\%$. The results highlight the importance of compact GAT architectures and flexible local action spaces for placement optimization. Overall, GPlaceRL provides a reproducible and extensible framework for systematic research on RL-based detailed placement refinement.
