ayl%hutds.hut.fi@MITVMA.MIT.EDU (Antti Ylikoski) (09/15/88)
Date: Sun, 11 Sep 88 15:32 EDT From: Antti Ylikoski <ayl%hutds.hut.fi@MITVMA.MIT.EDU> To: AIList@AI.AI.MIT.EDU Subject: Robotics and Free Will cc: ayl@hutds.hut.fi In a recent AIList issue, John McCarthy presented the problem how a robot could utilize information dealing with its previous actions to improve its behaviour in the future. Here is an idea. Years ago, an acquaintance of mine came across a very simple computer game which was annoyingly overwhelming to its human opponent. The human chose either 0 or 1. The computer tried to guess the alternative he had chosen in advance. He told the alternative he had chosen to the computer, which told him if it had guessed right or wrong. The human got a point if the guess of the machine was incorrect; otherwise the machine got a point. After a number of rounds, the computer started to play very well, guessing the alternative that the human had chosen correctly in some 60-70 per cent of the rounds. Neither of us ever got to know how the game worked. I would guess it had a model of the behaviour of the human opponent. Perhaps the model was a Markov process with states "human chooses 0" and "human chooses 1"; maybe the program performed a Fourier analysis of the time series. This suggests an answer to McCarthy's problem. Make the robot have a model of the behaviour of the environment. Calculate the parameters of the model with a best fit approach from the history data. The robot also might have several possible models and choose the one which produces the best fit to the history data. If the environment is active (other robots, humans) one also could apply game theory. ------------------------------------------------------------------------------ Antti Ylikoski !YLIKOSKI@FINFUN (BITNET) Helsinki University of Technology !OPMVAX::YLIKOSKI (DECnet) Digital Systems Laboratory !mcvax!hutds!ayl (UUCP) Otakaari 5 A ! SF-02150 Espoo, Finland ! ------------------------------------------------------------------------------