Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

ASAP: Exploiting the Satisficing Generalization Edge in Neural Combinatorial Optimization

About

Deep Reinforcement Learning (DRL) has emerged as a promising approach for solving Combinatorial Optimization (CO) problems, such as the 3D Bin Packing Problem (3D-BPP), Traveling Salesman Problem (TSP), or Vehicle Routing Problem (VRP), but these neural solvers often exhibit brittleness when facing distribution shifts. To address this issue, we uncover the Satisficing Generalization Edge, which we validate both theoretically and experimentally: identifying a set of promising actions is inherently more generalizable than selecting the single optimal action. To exploit this property, we propose Adaptive Selection After Proposal (ASAP), a generic framework that decomposes the decision-making process into two distinct phases: a proposal policy that acts as a robust filter, and a selection policy as an adaptable decision maker. This architecture enables a highly effective online adaptation strategy where the selection policy can be rapidly fine-tuned on a new distribution. Concretely, we introduce a two-phase training framework enhanced by Model-Agnostic Meta-Learning (MAML) to prime the model for fast adaptation. Extensive experiments on 3D-BPP, TSP, and CVRP demonstrate that ASAP improves the generalization capability of state-of-the-art baselines and achieves superior online adaptation on out-of-distribution instances.

Han Fang, Paul Weng, Yutong Ban• 2025

Related benchmarks

TaskDatasetResultRank
3D Bin Packing Problem3D-BPP ID-Small Discrete
Utilization (%)87.4
30
3D Bin Packing Problem3D-BPP ID-Large Discrete
Utilization (%)74.2
30
3D Bin Packing Problem3D-BPP Continuous (ID-Large)
Utilization63.9
26
3D Bin Packing Problem3D-BPP ID-Small Continuous
Utilization (%)72.6
26
Capacitated Vehicle Routing ProblemImplosion CVRP50
Computation Time0.5
20
Traveling Salesman ProblemImplosion TSP100
Computation Time (mixed units)0.3
20
Traveling Salesman ProblemExplosion TSP-1000
Baseline Optimality Gap10.35
17
Discrete 3D Bin Packing3D-BPP Default
Utilization (%)84.8
13
Discrete 3D Bin Packing3D-BPP ID-Medium
Utilization (%)79.9
13
Discrete 3D Bin Packing3D-BPP OOD
Utilization (%)65.6
13
Showing 10 of 28 rows

Other info

Follow for update