Now a chess playing machine could not be constructed in this way, because here the number of possible positions is too immense. Nevertheless chess playing machines exist. They operate not by builtin prescribed moves but by built-in general rules, "heuristics", as they are called. A heuristic is a rule which prescribes how to evaluate each chess position according to certain criteria. The positions examined are the positions which can result from a given position after a certain small number of moves, say two, by each player. Although the number of such positions can be very large, it is not unmanageable for a high speed computer. The move selection can then be the move which will lead to the most valuable position after two moves, assuming that the opponent will make the best moves available to him.

Now let us see how the evaluation rules can be assigned. Obviously if a contemplated move provides the opponent an opportunity to checkmate, the contemplated move must be eliminated. Thus, immediate checkmate threats are parried. Next, if the machine can "see" a checkmate to the opponent on the immediate or on the next move, regardless of what the opponent does, this move (or moves) will be assigned the greatest value. Thus, the opportunity to impose checkmate will always be utilized. If none of these situations are imminent, the quality of future possible positions will be evaluated by various general considerations such as value of each piece, control of center, mobility, etc.

Each of the positions can be given an exact numerical value to the extent that the above desiderata are more fully realized for the machine and less fully realized for the opponents This value is compounded of a weighted sum of the values of the positions by the listed criteria. The set of weighting factors for the several criteria used obtaining this sum constitutes a particular heuristic. Once a numerical value is assigned to each of the foreseen positions, the machine will choose the move which will certainly lead to the most valuable position for itself, assuming that the opponent makes his best moves.

Now the automaton built on these principles has not been "told" what move to play in every conceivable situation. It has only been told what rules to apply in choosing moves, that is it has been told only how to play. . . . It must be admitted, this step is in the direction of "thought", at least according to a certain understanding of thought.

But now consider a further modification. Suppose after each game the machine adjusts the weighting factors which determine the relative importance of the criteria for evaluating positions, Suppose, moreover, that it makes such adjustments at first in a random fashion, as if conducting a search. If it looses too many games, it will make adjustments roughly opposite to the ones preceding the loosing streak; if it has a winning streak, it will make further adjustments in the same direction. This machine will behave like a learning organism, making variations in its behavior pattern which are at first random, then more and more systematic leading finally to the elimination of "punished" patterns and the fixation of "rewarded" ones. . . . This machine was not "trained"; it was "educated," that is taught how to find out things for itself. (A, pp. 97-98)