Multi- and Many-Threaded Heterogeneous Parallel Grammatical Evolution

• Published in: Handbook of Grammatical Evolution pp. 219 - 244
• Main Authors: Dufek, Amanda Sabatini, Augusto, Douglas Adriano, Barbosa, Helio José Corrêa, da Silva Dias, Pedro Leite
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 2018; Springer International Publishing
• ISBN: 3319787160; 9783319787169
• DOI: 10.1007/978-3-319-78717-6_9

There are some algorithms suited for inference of human-interpretable models for classification and regression tasks in machine learning, but it is hard to compete with Grammatical Evolution (GE) when it comes to powerfulness, model expressiveness and ease of implementation. On the other hand, algorithms that iteratively optimize a set of programs of arbitrary complexity—which is the case of GE—may take an inconceivable amount of running time when tackling complex problems. Fortunately, GE may scale to such problems by carefully harnessing the parallel processing of modern heterogeneous systems, taking advantage of traditional multi-core processors and many-core accelerators to speed up the execution by orders of magnitude. This chapter covers the subject of parallel GE, focusing on heterogeneous multi- and many-threaded decomposition in order to achieve a fully parallel implementation, where both the breeding and evaluation are parallelized. In the studied benchmarks, the overall parallel implementation runtime was 68 times faster than the sequential version, with the program evaluation kernel alone hitting an acceleration of 350 times. Details on how to efficiently accomplish that are given in the context of two well-established open standards for parallel computing: OpenMP and OpenCL. Decomposition strategies, optimization techniques and parallel benchmarks followed by analyses are presented in the chapter.