CADO: From Imitation to Cost Minimization for Heatmap-based Solvers in Combinatorial Optimization
Researchers have introduced CADO, a novel framework designed to improve heatmap-based solvers for combinatorial optimization problems. Unlike traditional supervised learning methods that focus on imitating data structures, CADO directly optimizes the cost of the final decoded solution. This is achieved by formulating the diffusion denoising process as a Markov decision process and employing a Label-Centered Reward system that uses ground-truth labels as baselines. The framework also incorporates Hybrid Fine-Tuning for efficient parameter adaptation, demonstrating state-of-the-art performance across various benchmarks. AI
IMPACT Optimizes combinatorial optimization solvers by directly minimizing solution cost, potentially improving efficiency in complex problem-solving.