Evaluation of a Master-Slave Parallel Evolutionary Algorithm Applied to Artificial Intelligence for Games in the Xeon-Phi Many-Core Platform

S. RODRIGUEZ LEOPOLD; F. PARODI; S. NESMACHNOW; E. MOCSKOS

High Performance Computing

Springer International Publishing

Lugar: Cham; Año: 2017; p. 161 - 176

Evolutionary algorithms are non-deterministic metaheuristic methods that emulate the evolution of species in nature to solve optimization, search, and learning problems. This article presents a parallel implementation of evolutionary algorithms on Xeon Phi for developingan artificial intelligence to play the NES Pinball game. The proposed par-allel implementation offloads the execution of the fitness function evalua-tion to Xeon Phi. Multiple evolution schemes are studied to get the mostefficient resource utilization. A micro-benchmarking of the Xeon Phi co-processor is performed to verify the existing technical documentation and obtain detail knowledge of its behavior. Finally, a performance analysis of the proposed parallel evolutionary algorithm is presented, focusing on the characteristics of the evaluated platform.