StatDiamond

AI-Driven Insights into Baseball Performance, Business, and Analytics

---

Reinforcement Learning in Baseball

The math that serves as the foundation for reinforcement learning is a Markov Decision Process (MDP). I am interested in representing pitching behavior by a MDP. I believe Value Iteration can help batters strategically counter a pitcher's pitch sequence. In order to compute the Value Iteration algorithm, I need a probability transition matrix.

Generate Transition Probability

pitcher_mdp.py is a Python script that creates a MDP class. This class is equipped with various methods for MDP related matter to conduct reinforcement learning. The method get_markov_chain() will generate a transition probability matrix (markov chain) among 12 pitch states, (0,0) (1,0) (2,0) (3,0) (0,1) (0,2) (1,1) (1,2) (2,1) (2,2) (3,1) (3,2)

Value Iteration

Once you've generated a markov chain of the pitching states, you can pass it into the value_iteration() function. Value iteration will find the best sequence of actions (swing or stand) to maximize the expectation at each state.

Usage

Since pitcher_mdp.py is a Python class, we can instantiate it in another python file. For example, find_batting_strategy.py instantiates a pitcher_mpd object and generates a markov chain and calls the value iteration method.

If you want to visualize the markov chain (without actions) call the plot_markov_chain() function.

pitcher_mdp.py should take in most MLB pitchers. I use the pybaseball package to acquire pitching data for an individual player. In order to correctly get data from pybaseball follow the parser arguments in the find_batting_strategy.py file.

For example,

python find_batting_strategy.py \
        --first_name Zac \
        --last_name Gallen \
        --start_dt 2022-04-16 \
        --end_dt 2022-10-04

---