ASAP: Exploiting the Satisficing Generalization Edge in Neural Combinatorial Optimization

Researchers have proposed new frameworks to improve Deep Reinforcement Learning (DRL) in solving Combinatorial Optimization problems, addressing the brittleness of neural solvers when facing distribution shifts.

Deep Reinforcement Learning has emerged as a promising approach for solving Combinatorial Optimization problems, according to recent studies^[1][2]. To address the issue of brittleness when facing distribution shifts, a new framework called Adaptive Selection After Proposal (ASAP) has been proposed^[1]. ASAP 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 effective online adaptation where the selection policy can be rapidly fine-tuned on a new distribution. Experimental results on problems such as 3D Bin Packing Problem (3D-BPP), Traveling Salesman Problem (TSP), and Capacitated Vehicle Routing Problem (CVRP) demonstrate that ASAP improves the generalization capability of state-of-the-art baselines and achieves superior online adaptation on out-of-distribution instances^[1]. Additionally, a unified robustness-oriented framework has been proposed for preference-conditioned DRL solvers for Multi-Objective Combinatorial Optimization Problems (MOCOPs), along with a preference-based adversarial attack to generate hard instances that expose solver weaknesses^[2]. A defense strategy integrating hardness-aware preference selection into adversarial training has also been introduced, with experimental results verifying the effectiveness of the attack and defense methods^[2].