Optimal Pitch Selection Policies Via Markov Decision Processes

Abstract

In this thesis, we employ Markov Decision Processes to consider the challenge of optimizing pitch selection in Major League Baseball. Using empirical transition frequencies from pitch-by- pitch data, transition probabilities are estimated between states that consist of the count, number of outs, and configuration of base runners. A Markov Decision Process determines that MLB pitchers could have, on average, reduced their earned run average by around 0.54 by adopting a policy that throws many more breaking balls than MLB pitchers currently tend to throw. Sensitivity analyses are conducted to confirm the robustness of these results to uncertainty in the estimated transition probabilities. We then consider two extensions to attempt to model more realistic pitcher behavior: a regularized setting where pitchers are forced to mix the pitch types they throw, and a simplified game-theoretic setting where data from the 2017 Houston Astros’ season models the behavior of the batter. In both cases, the results also suggest that pitchers would benefit from relying less on fastballs.