Parallel Gradient-Based Local Search Accelerating Particle Swarm Optimization for Training Microwave Neural Network Models

This paper presents a novel global optimization technique for training microwave neural network models. Unlike existing sequential hybrid algorithms, the proposed technique implements parallel gradient-based local search in particle swarm optimization (PSO). The whole swarm is divided into subswarms for multiple processors. The particle with the lowest error in the subswarm in each processor is chosen to do further local search using quasi-Newton method. This process is performed in all the subswarms in parallel using the message passing interface (MPI). The proposed technique increases the probability and speed of finding a global optimum. This technique is illustrated by two microwave modeling examples.