Context-Dependent Pitch Prediction with Neural Networks

-
This semester I took the machine learning class at UMass, and for my final project I developed models for predicting characteristics of pitches based on the context that they were thrown in.  The context consists of relevant and available information known before the pitch is thrown, such as the identities of the batter and pitcher, the batter height and stance, the pitcher handedness, the current count, and many others.  In particular, I looked into modeling the distribution over pitch types and locations.  This problem is challenging primarily because for a particular context (a specific setting of the different features that make up the context), there is usually very little data available for pitches thrown in that context.  In general, for each context feature we include, we have to split up our data into $k$, groups where $k$ is the number of values that the context feature can take on (for batter stance it is $k=2$, for count it is $k=12$, etc).  Thus, developing models by partitioning it into groups based on context then individually constructing models for the pitch type and location would not work well.  To overcome this problem, I developed a neural network that can automatically learn the complex dependencies between the context and the distribution over pitch types and locations.  To model the pitch type, I used a standard neural network for probabilistic classification.  And to model the pitch location, I used something called a mixture density network, which I explain in more detail in the write-up.  I demonstrated that these models outperformed models that don't account for context in a principled fashion, and showed that they generalize effectively to pitches thrown in new contexts.
For more information on this project and these models, please refer to the following resources:
Feel free to get in touch with me over email if you have any questions or want to extend the work.

Comments

Post a Comment

Popular posts from this blog

Optimal Strategy for Farkle Dice

-
In this post, I will discuss my findings in terms of an optimal strategy for Farkle , which is a dice game of chance.  If you are unfamiliar with the game, I encourage you to skim the linked article so that you can better understand this blog post.  All of my findings are based on sound mathematical methods and a computer program was used to determine the optimal strategy.  Before I begin, let me first state the value of different dice rolls I assumed while developing the strategy. Each 1: 100 points Each 5: 50 points 3 1's: 1000 points 3 2's: 200 points ... 3 6's: 600 points 4 1's: 2000 points ... 4 6's 1200 points 5 1's 3000 points ... 5 6's 1800 points Straight 1-6 1500 points 3 Pairs (e.g., 22 33 44): 750 points Everything else is considered to be a Farkle (0 points). There are many variants to Farkle, but I like to play by this set of rules. The "game state" of Farkle roll can roughly be characterized by 2 values: the curre...
Read more

Markov Chains and Expected Value

-
A few weeks ago, I was using a Markov Chain as a model for a Project Euler problem, and I learned about how to use the transition matrix to find the expected number of steps to reach a certain state. In this post, I will derive the linear system that will help answer that question, and will work out specific an example using a 1-dimensional random walk as the model. Terminology Before I begin the derivation, let me define some terms. The transition matrix is an \( n \) by \( n \) matrix that describes how you can move from one state to the next. The rows of the transition matrix must always sum to 1. States can either be transient or absorbing . An absorbing state is one that cannot be left once reached (it transitions to itself with probability 1). A transient state is a state that is not an absorbing state. In many problems, it is of general interest to compute the expected number of steps to reach an absorbing state (from some start state). Derivation  Let \...
Read more

Beat the Streak: Day Three

-
Image
In order to maximize your probability of beating the streak, you should (1) predict the probability that a batter will get a hit in a given game given the game parameters and (2) determine if it's worthwhile to risk your current streak in order to possibly improve it by 1 or 2 games. In this blog post, I outline my solution to (2). In previous blog posts, I've hinted at what I do to solve (1) and will continue that discussion in a later blog post. Motivation When your current streak is short, the optimal strategy is to pick the best player every day, regardless of how likely their probability of getting a hit is (to a certain extent). However, as your streak grows, you have an important decision to consider: is it better to pick the best player today and possibly lose your current streak, or skip picking a player and instead maintain your current streak. Naturally, this decision should be guided by the players probability of getting a hit, as well as the distribution of the...
Read more