Online Program

Return to main conference page
Thursday, May 17
Machine Learning Applications
Thu, May 17, 6:15 PM - 7:15 PM
Regency Ballroom B
 

Penalized Regression Within the Game Cribbage (304640)

Presentation

Amanda Montoya, The Ohio State Univerisity 
*Christopher Silberstein, The Ohio State Univerisity 

Keywords: Cribbage, data, machine learning, algorithms, game, expected value, efficiency

Cribbage is a card game that is played by many and currently does not contain much complete analysis of all its components. There are three main components of the game. Players cycle between being the dealer, which they must deal both players a six-card hand, in which they must choose four cards to keep and two to throw into the “crib”. A card is then flipped up from the deck and they begin playing cards one at a time within the “pegging” process. After the dealer looks at the four cards within the crib to score. If neither person is at 121 points or higher the game continues. The proposed algorithms for playing Cribbage do not take pegging into consideration. Pegging is an important part of the game where the dealer always scores at least 1 point and it is surprising how the other programs do not take it into account. My project shows that taking pegging into consideration increases the expected value of the hand which increases the probability of winning. The goal of the project is to have the program be able to dynamically shift the way that it plays cribbage based on the initial data of 1000 games from an online cribbage company that was given and through the collection of data as it plays more hands. Using multi-level modeling on the training dataset, we determined that the ace has the highest expected value during pegging. The program uses penalized regression to simultaneously balance the importance of different cards in the different stages of the game (hand, pegging, crib). This means the program will be able to adjusts the way that it plays to maximize the efficiency of every hand. Previous attempts to develop algorithms have ignored different parts of the game, perhaps assuming that because the hand is the highest scoring section of the game, the other parts have little importance. This project demonstrates that integrating the knowledge of how the game is played within the algorithm performs better than ignoring seemingly unimportant parts of the game.