Enhancing Graph Partitioning with Reinforcement Learning-based Initialization

Graph partitioning is fundamental in many computer-aided design (CAD) applications, as it decomposes large problems into more manageable subproblems. State-of-the-art frameworks often count on the Fiduccia-Mattheyses (FM) algorithm to iteratively refine partition quality. However, the solution quality of FM heavily depends on the quality of the initial partition, which is typically generated randomly and can lead to suboptimal results. To address this issue, we propose a reinforcement learning (RL)-based algorithm to generate high-quality initial partitions that improve the final solution quality of the FM algorithm. By designing a reward function that promotes cut size reduction while penalizing partition imbalance, our RL agent learns to produce balanced graph partitions with minimized cut sizes. Evaluated on a set of industrial circuits, our experiments show that running FM on our RL-based initial partitions yields significantly better cut sizes than with random initialization.