Special Issue on General Game Playing

Abstract

Computer game playing has always been of special interest to Artificial Intelligence from the early days on. The cultural history of game-playing machines reaches even farther back than the advent of the modern computer. In the 18th century, an (allegedly) purely mechanical automaton, The Turk, outplayed almost any human chess player. The main reason for The Turk’s popularity was that people considered human-level intelligence a basic prerequisite for being a good chess-player. This motivated the first AI researchers to regard chess-playing as an excellent benchmark for the creation of intelligent machines. But when in 1997 a computer finally defeated the then world chess champion in a six-game match, it was clear that progress in computer chess had little if any implications on how to build intelligent machines in general. This insight has led to the recent interest in general game playing, which is concerned with systems that can learn to play just any game by being given nothing but the bare rules of that game. Figure 1 provides a schematic illustration of this principle. Because neither the specifics nor even the type of games to be played are known at programming time, a general game player cannot be endowed with fixed heuristics, tailor-made for one specific game. Rather, the player must be able to design a suitable strategy all by itself whenever it is asked to play a new game. Successful general game-playing programs can thus be seen as a new generation of AI systems that can be told what we want from them and that learn to carry out possibly radically different tasks without human intervention.