핵심 개념
The proposed Instance-Conditioned Adaptation Model (ICAM) effectively integrates instance-conditioned information, such as problem scale and node-to-node distances, into both the encoding and decoding processes of a neural combinatorial optimization model, enabling it to achieve state-of-the-art performance on large-scale Traveling Salesman Problem and Capacitated Vehicle Routing Problem instances.
초록
The paper presents a novel Instance-Conditioned Adaptation Model (ICAM) to improve the large-scale generalization performance of reinforcement learning-based neural combinatorial optimization (NCO) methods.
Key highlights:
- ICAM incorporates instance-conditioned information, including problem scale and node-to-node distances, into both the encoding and decoding processes of the NCO model through a powerful yet lightweight Adaptation Attention Free Module (AAFM) and a new compatibility calculation.
- The authors develop a three-stage reinforcement learning-based training scheme that enables the model to learn cross-scale features efficiently without any labeled optimal solutions.
- Experimental results show that ICAM outperforms state-of-the-art RL-based constructive NCO methods on Traveling Salesman Problem and Capacitated Vehicle Routing Problem instances with up to 1,000 nodes, achieving significantly faster inference times.
- The authors also demonstrate ICAM's superior performance on larger-scale instances (up to 5,000 nodes) compared to other classical and learning-based solvers.
The paper presents a comprehensive approach to address the crucial limitation of existing RL-based NCO methods in achieving large-scale generalization, which is essential for practical applications.
통계
The paper reports the following key metrics:
Solution lengths (Obj.) for TSP and CVRP instances of different scales
Optimality gaps (Gap) compared to optimal solutions
Total inference times (Time) for generating solutions
인용구
"To the best of our knowledge, our model achieves state-of-the-art performance among all RL-based constructive methods for TSP and CVRP with up to 1,000 nodes."
"On CVRP instances with scale ≥1000, our method outperforms the other methods, including GLOP with LKH3 solver and all TAM variants, on all problem instances except for CVRP3000."