Optimizing training programs for athletic performance: a Monte-Carlo Tree Search variant method

Abstract

AbstractPurposeUsing a variant of the Monte-Carlo Tree Search (MCTS) algorithm, we compute optimal personalized and generic training programs for athletic performance.MethodsWe use a non-linear performance model with population variability for athletes and non-athletes previously used in the literature. Then, we simulate an in-silico test population. For each individual of this population, we compute the performance obtained after implementing several widely used training programs as well as the one obtained by our variant of the MCTS algorithm. Two cases are considered depending on individual parameters being observed and personalized programs being possible or only parameter distributions being available and only generic training programs being implementable.ResultsCompared to widely used training programs, our optimization leads to an increase in performance between 1.1 (95% CI: 0.9 – 1.4) percentage point of the performance obtained with stationary optimal training dose (pp POTD) for athletes and unknown individual characteristics to 10.0 (95% CI: 9.6 – 10.3) pp POTD for nonathletes and known individual characteristics. The value of information when using MCTS optimized training strategies, i.e. the difference between the performance that can be reached with knowledge of individual characteristics and the performance that can be reached without it is 14.7 (95% CI: 12.8 – 16.7) pp POTD for athletes and 3.0 (95% CI: 2.6 – 3.4) pp POTD for non-athletes.