RIDGECUT: Learning Graph Partitioning with Rings and Wedges
Researchers have developed RIDGECUT, a novel reinforcement learning framework designed for graph partitioning problems, specifically targeting the Normalized Cut problem. This method incorporates domain knowledge by constraining actions to enforce structural partitioning, inspired by transportation networks where partitions often form rings and wedges. By transforming graphs and utilizing transformer-based policies with Proximal Policy Optimization, RIDGECUT achieves lower normalized cuts and demonstrates strong generalization across various graph sizes and types, outperforming existing methods on synthetic and real-world traffic data. AI
IMPACT Introduces a novel RL approach for graph partitioning, potentially improving efficiency and generalization in combinatorial optimization tasks.