Genetic Algorithms and Feature Subset Selection for Predicting Athletic Performance: Case of Professional Football.

摘要

Existing predictive modeling in sports analytics often hinges on atheoretical assumptions winnowed from a large and diverse pool of game metrics. Feature subset selection by way of a genetic algorithm to identify and assess the combinatorial advantage for a group of metrics is a viable option to otherwise arbitrary model construction. However, this approach concedes similar arbitrariness as there is no general strategy or common practice design among the tightly coupled nucleus of genetic operators. Instead, extant research typically considers and improves any single operation in the evolutionary workflow, leaving future practitioners a wealth of advanced individual parameters from which to construct a necessarily sub-optimal mix. The resulting dizzying ecosystem of choice is especially difficult to overcome and leaves a residual uncertainty regarding true strength of output, specifically for practical implementations.;Conventional random population initialization in the standard genetic algorithm is probabilistically accountable within a set range of allele frequencies for a binary encoded chromosome. As such, solutions evolve inside a confined pocket of the search space. It remains unobvious if the strongest fit solution would survive beyond the familiar Meta from which it evolved. In nature, extreme environmental change, such as water temperature increases from an El Nino event, threaten unassuming organisms that, despite surviving an ongoing evolutionary process, cannot cope with conditions of an unfamiliar Meta and subsequently die.;In concert with design science research methodology, this study transposes ideas from extreme environmental change into a quasi-deterministic extension of standard genetic algorithm functionality that seeks to punctuate converged populations with individuals from auxiliary metas. This strategy has the effect of challenging what might otherwise be considered shallow fitness, thereby promoting greater trust in output against innumerable alternatives. A subroutine is constructed and evaluated in a larger predictive modeling apparatus for future athletic performance in the National Football League at a specific skill position, running back. Box score data from seasons 2013 and 2014 are considered and the strategy is validated against output from four alternate sources.