Pretrained Cost Model for Distributed Constraint Optimization Problems

Distributed Constraint Optimization Problems (DCOPs) are an important subclass of combinatorial optimization problems, where information and controls distributed among multiple autonomous agents. Previously, Machine Learning (ML) has been largely applied to solve problems by learning effective heuristics. However, existing ML-based heuristic methods often not generalizable different search algorithms. Most importantly, these usually require full knowledge about the be solved, which suitable for settings centralization is realistic due geographical limitations or privacy concerns. To address generality issue, we propose a novel directed acyclic graph representation schema DCOPs leverage Graph Attention Networks (GATs) embed representations. Our model, GAT-PCM, then pretrained with optimally labelled data in offline manner, so as construct heuristics boost broad range DCOP algorithms evaluating quality partial assignment critical, such local backtracking search. Furthermore, enable decentralized model inference, embedding GAT-PCM each agent exchanges only embedded vectors, show its soundness complexity. Finally, demonstrate effectiveness our combining it algorithm. Extensive empirical evaluations indicate that GAT-PCM-boosted significantly outperform state-of-the-art various benchmarks.