Can Combinatorial Optimization Benefit from Latent Space Search Policy Adaptation?

Original title: Combinatorial Optimization with Policy Adaptation using Latent Space Search

Authors: Felix Chalumeau, Shikha Surana, Clement Bonnet, Nathan Grinsztajn, Arnu Pretorius, Alexandre Laterre, Thomas D. Barrett

The challenge of solving complex problems like Combinatorial Optimization tasks often relies on Reinforcement Learning (RL), which offers flexible heuristics. However, existing RL methods struggle to surpass established solvers. A new approach, COMPASS, changes this by introducing a unique RL method. COMPASS pre-trains diverse specialized policies using a continuous latent space, aiming to predict strong search strategies before actual problem-solving. Testing this method on problems like Travelling Salesman, Capacitated Vehicle Routing, and Job-Shop Scheduling, COMPASS outperforms existing methods on 11 benchmark tasks and excels in generalization across 18 different instances. This breakthrough suggests that anticipating effective search strategies during pre-training significantly enhances the solution quality for combinatorial optimization problems, marking a promising direction for future research.

Original article: https://arxiv.org/abs/2311.13569