Evolutionary Algorithms (EAs) are inherently parallel due to their ability to simultaneously evaluate the fitness of individuals. Synchronous Parallel EAs (SPEAs) leverage this with the intent to gain significant speed-ups when executed on multiple processors. However, many important problem classes lead to large variations in fitness evaluation times, such as is often the case in hyper-heuristics where the time complexity of executing one individual may differ greatly from that of another. Asynchronous Parallel EAs (APEAs) omit the generational synchronization step of traditional EAs which work in well-defined cycles. They can provide scalability improvements proportional to the variation in fitness evaluation times of the evolved individuals, and therefore should be considered for use in hyper-heuristics. This paper provides an empirical analysis of the improvements obtained by applying APEAs, compared to SPEAs, on a case study involving the evolution of conflict-driven clause learning Boolean satisfiability solvers, demonstrating that APEAs are the future of hyper-heuristics.
Available at: http://works.bepress.com/daniel-tauritz/67/