Unlikely Best Response


This article is about computer programmer Tom Murphy.  Tom has created a computer program with artificial intelligence, which learns as it completes tasks. The task that Tom Murphy gave his program was to play and win old video games.  The computers parameters for winning the game, were set as the bytes of memory being used and not losing.  The longer the computer played the game the more bytes of memory would be used.  The idea was that as the computer played the game for longer and longer periods of time,the computer would learn best responses to the challenges in the game.  The best response being to increasing bytes used in the game, and not losing.  The computer program is using Game Theory to better complete it’s allotted task, finding the best strategies to the challenges at hand.  The computer while playing exhibited some very amazing responses.  Responses that could not be executed by humans, who are limited by motor skills.  There was however one response that was not expected.  The computer after playing some of the games determined that it’s dominant strategy was to simply not play.  It is best illustrated in the end of the video where the AI is given the task of playing Tetris.  When the game becomes too hard and the computer is about to lose the computer pauses the game.  Pausing the game satisfies the not losing parameter of the game.  It also satisfies the increasing bytes used to play the game. While the game was paused the music and screen still progressed, which increased the bytes used.  The computer would for many game simply pause the game and consider the game won because it did not lose. I believe this a good example of artificial intelligence executing its best response within a game.  The computers payoffs were either wining, which when paused the computer believed it was doing, or the payoff of 0 if it lost. The computer could acquire byte at a faster rate by playing the game, but had the possible pay off of 0.  When paused, the computer could increase bytes slowly and eliminate the possibility of a 0 payoff.   In the end the best response to loosing is to not play.