Learning to Win Games in a Few Examples: Using Game-Theory and Demonstrations to Learn the Win Conditions of a Connect Four Game

摘要

Teaching robots new skills using minimal time and effort has long been a goal of artificial intelligence. This paper investigates the use of game theoretic representations to represent interactive games and learn their win conditions by interacting with a person. Game theory provides the formal underpinnings needed to represent the structure of a game including the goal conditions. Learning by demonstration, has long sought to leverage a robot's interactions with a person to foster learning. This paper combines these two approaches allowing a robot to learn a game-theoretic representation by demonstration. This paper demonstrates how a robot can be taught the win conditions for the game Connect Four using a single demonstration and a few trial examples with a question and answer session led by the robot. Our results demonstrate that the robot can learn any win condition for the standard rules of the Connect Four game, after demonstration by a human, irrespective of the color or size of the board and the chips. Moreover, if the human demonstrates a variation of the win conditions, we show that the robot can learn the respective changed win condition.