No abstract

Black-Box Search Algorithms (BBSAs) tailored to a specific problem class may be expected to significantly outperform more general purpose problem solvers, including canonical evolutionary algorithms. Recent work has introduced a novel approach to evolving tailored BBSAs through a genetic programming hyper-heuristic. However, that first generation of hyper-heuristics suffered from overspecialization. This poster paper presents a second generation hyperheuristic employing a multi-sample training approach to alleviate the overspecialization problem. A variety of experiments demonstrated the significant increase in the robustness of the generated algorithms due to the multi-sample approach, clearly showing its ability to outperform established BBSAs. The trade-off between a priori computational time and the generated algorithm robustness is investigated, demonstrating the performance gain possible given additional run-time.

Black-Box Search Algorithms (BBSAs) tailored to a specific problem class may be expected to significantly outperform more general purpose problem solvers, including canonical evolutionary algorithms. Recent work has introduced a novel approach to evolving tailored BBSAs through a genetic programming hyper-heuristic. However, that first generation of hyper-heuristics suffered from over-specialization. This paper presents a study on the second generation hyperheuristic which employs a multi-sample training approach to alleviate the over-specialization problem. In particular, the study is focused on the affect that the multi-sample approach has on the problem configuration landscape. A variety of experiments are reported on which demonstrate the significant increase in the robustness of the generated algorithms to changes in problem configuration due to the multi-sample approach. The results clearly show the resulting BBSAs' ability to outperform established BBSAs, including canonical evolutionary algorithms. The trade-off between a priori computational time and the generated algorithm robustness is investigated, demonstrating the performance gain possible given additional run-time.

Many important problem classes lead to large variations in fitness evaluation times, such as is often the case in Genetic Programming where the time complexity of executing one individual may differ greatly from that of another. Asynchronous Parallel Evolutionary Algorithms (APEAs) omit the generational synchronization step of traditional EAs which work in well-defined cycles. This paper provides an empirical analysis of the scalability improvements obtained by applying APEAs to such problem classes, aside from the speed-up caused merely by the removal of the synchronization step. APEAs exhibit bias towards individuals with shorter fitness evaluation times, because they propagate faster. This paper demonstrates how this bias can be leveraged in order to provide a unique type of "elitist" parsimony pressure which rewards more efficient solutions with equal solution quality.

Practitioners often need to solve real world problems for which no custom search algorithms exist. In these cases they tend to use general-purpose solvers that have no guarantee to perform well on their specific problem. The relatively new field of hyper-heuristics provides an alternative to the potential pit-falls of general-purpose solvers, by allowing practitioners to generate a custom algorithm optimized for their problem of interest. Hyper-heuristics are meta-heuristics operating on algorithm space employing targeted primitives to compose algorithms. This paper explores the advantages and disadvantages of expanding a hyperheuristic's primitive-space with additional primitives. This should allow for an increase in quality of evolved algorithms. However, increasing the search space of a meta-heuristic almost always results in longer time to convergence and lower quality results for the same amount of computational time, but also all too often lower quality results at convergence, potentially making a problem impractical to solve for a practitioner. This paper explores the scalability of hyperheuristics as the primitive-space is increased, demonstrating significantly increased quality solutions at convergence with a corresponding increase in convergence time. Additionally, this paper explores the impact that the nature of the added primitives have on the performance of the hyper-heuristic.