Empowering distributed constraint optimization with deep learning

Abstract

Chapter 3 is published as PCM-boosted algorithms through extensive empirical evaluations. In the final part, we investigate using DNNs to boost the performance of BP for solving COPs, by learning the optimal dynamic hyperparameters. Our model, DABP, seamlessly integrates BP and DNNs within the message-passing framework to reason about dynamic neighbor weights and damping factors for composing new BP messages. Furthermore, unlike existing neural-based BP variants, we propose a novel selfsupervised learning algorithm for DABP with a smoothed cost, which does not require expensive training labels and also avoids the common out-of-distribution issue through efficient online learning. Extensive experiments show that our model significantly outperforms state-of-the-art baselines.