Optimization and Adaptation of Athlete Training Load With Evolutionary Computation

Managing the relationship between training and performance in athletic populations is a complex problem due to its non-linearity and highly constrained nature. The relationship between a training dose and subsequent improvements in performance can not be adequately described through simple mechanistic models. No formalised theories have emerged to guide the planning and adaptation of athlete training load such that a desired or prescribed effect on performance can be realised. This research aims to develop a system of methods, based on evolutionary computation, that can support the management of the training-performance relationship, through effective planning, optimization, adaptation and modeling. This thesis presents a training load generation system capable of maximizing model based measures of performance or minimizing model based measures of fatigue. This system is based on the grammatical evolution algorithm which allows it to incorporate hierarchical structures, domain knowledge and constraints into the training load optimization and generation process by utilising a formal grammar. We build on this system to develop a complementary intelligent control system which is capable of adapting future training loads in response to deviations away from an optimal training or performance trajectory. When applied to the problem of adapting the future training loads of athletes subject to large deviations away from a training load target trajectory, the system generated future training loads which outperformed alternative random and proportional approaches to the problem. Finally, we explore the use of alternative optimization methods to fit the parameters of a dose-response model. Results demonstrated that performance improvements can be gained through the utilization of heuristic based optimization algorithms, whilst also removing the need to supply informative initial parameter values. We extend our initial focus on parameter optimization to investigate the use of symbolic regression to select both the features and parameters of a model, in addition to combining them into a functional analytical form capable of describing the training-performance relationship.