FORETELLING THE FUTURE BY ADAPTIVE MODELING

Abstract

Foretelling the future is a pastime that all of us enjoy every day, but one which is not normally viewed as a scientific activity. Yet it is often quite possible to predict the future-albeit in a limited way-and to put the prediction to good use. By supplying an expected, "normal", course of events, a good prediction focuses attention on how unfolding events differ from the norm. In many cases this is more interesting than the events themselves. For example, suppose one computer is sending a stream of numbers to another. If it is possible to predict exactly what the next number is going to be, then there is no point in sending it at all! But suppose the guesses are not so reliable. Then it may be advantageous to send the difference between the guess and the actual value of the next number. The receiver-assuming it can make identical guesses-will be able to reconstruct the correct number from the difference. If the guesses are good, this scheme will reduce the amount of information that needs to be sent. In fact, the reduction achieved is a quantitative measure of success at guessing.