COGAP-42: game playing and learning in a non-board, imperfect informtion game
M. Sandra Carberry · Rice Research Repository (Rice University) · 1970
This thesis investigates a new area of mechanical game playing and learning.The first part surveys previous game playing and learning research.The second and third parts develop a game playing/learning machine.The fourth part presents an inductive definition of heuristic parameter and analyzes the heuristic parameters used in this machine.This project expands mechanical game playing and learning to a new class of problem."42" bidding was chosen for investigation because it differs in many respects from previous areas of experimentation.The major differences are: 1) non-board, 2) non-zero-sum, 3) imperfect information, and 4) partnership interaction.These factors required that the machine be designed as a two phase system hand evaluation and bid determi¬ nation with rote and generalized learning acting interdependently.Strategy is introduced via a goaloriented evaluation and learning procedure.This should more closely parallel human decision making functions and strategical considerations than standard non-goal evaluation and learning.A heuristic is developed for estimating the probabilities necessitated by imperfect This research project expands mechanical game playing and learning to a new class of problems.The game chosen for investigation is "42", a dominoes game similar in many respects to the card game bridge.Four players, forming two partnerships, are required; each player is dealt a hand of seven dominoes.After analyzing his hand, each player is allowed one and only one bid; the player making the highest bid obtains the contract for his side.This team then attempts during the course of play to take enough tricks and special point dominoes to make the contract bid.The detailed rules of "42" and sample play are presented in Appendix I. "42" bidding was chosen as the subject of learning research both for its simplicity and its complexity.Each player receives seven' dominoes compared with thirteen cards in bridge; only one round of bidding is allowed instead of several progressively higher bids.However, certain dominoes have special point value and must be given special consideration in any bidding.These factors lessen memory requirements considerably yet maintain a complexity that prevents algorithmic bid determination.This research seeks to investigate several new phases of game playing and learning."42" is a non-board, non zero-sum game requiring cooperative team play.Previous research games have maintained perfect information there is no chance element and each player possesses all infor¬ mation necessary to determine the best line of play."42"