Abstract
This paper develops a new message passing protocol that can be used to speed up the inference process of Generalized Distributive Law (GDL) based constraint optimization algorithms. In particular, we parallelize the inference process of the GDL-based constraint optimization algorithms which have been widely used to solve Constraint Optimization Problems (COPs) in different real world applications, such as wire routing, image analysis, computer-aided gas pipeline operation, and numerical optimization. It is worth noting that the proposed new parallel approach can be applied to accelerate any GDL-based message passing algorithms, such as Max-Sum, Max-Product or Sum-Product. In this work, we make use of the available Central Processing Unit (CPU) cores of a system to minimize the time it requires to execute a given algorithm. Consequently, this reduced computation time will accelerate the GDL-based COP algorithms which can be used in larger systems. However, the main challenge is to maintain the quality of the solution while minimizing the completion time. Our proposed protocol specifically takes this trade-off into account, and our empirical results depict that it is able to produce the same solution quality while accelerating the inference process 2-10 times faster compared to the state-of-the-art.