X. Li, C. Han, T. GUO
The Quadratic Assignment Problem (QAP) is one of the NP-hard combinatorial optimization problems and is known for its diverse applications in real life. The metaheuristics are prevalent solution methods for this problem. However, it is difficult to solve in the polynomial time even for small instances and different metaheuristics are applicable to different problems. In this paper, we propose a Graph Pointer Network(GPN) for the QAP using Reinforcement Learning. Our method introduces the Graph Neural Network to capture the relationship between each node based on the traditional Pointer Network. The trained network then outputs approximate solutions in feasible time, without the need to re-train for every new problem instance. We demonstrate the performance of our approach outperforms the previous learning-based methods on the benchmark instances of QAPLIB, a well-known library of QAP instances, and show that our model is able to generalize well: (i) from training on small graphs to testing on large graphs; (ii) from training on one type of random graphs to testing on another type of random graphs; and (iii) from training on random graphs to running on real-world graphs.
Keywords: Quadratic Assignment Problem,Graph Pointer Networks,Reinforcement LearningScheduled
TD2 Quadratic Assignment and Knapsack Optimization
June 10, 2021 2:45 PM
2 - LV Kantorovich