In 1950, Claude Shannon published his seminal work on how to program a computer to play chess. Since then, developing game-playing programs that can compete with (and even exceed) the abilities of the human world champions has been a long-sought-after goal of the AI research community. In Shannon's time, it would have seemed unlikely that only a scant 50 years would be needed to develop programs that play world-class backgammon, checkers, chess, Othello, and Scrabble. These remarkable achievements are the result of a better understanding of the problems being solved, major algorithmic insights, and tremendous advances in hardware technology. Computer games research is one of the important success stories of AI. This article reviews the past successes, current projects, and future research directions for AI using computer games as a research test bed.

Although TD-Gammon is one of the major successes in machine learning, it has not led to similar impressive breakthroughs in temporal difference We werelearning for other applications or even other games. Instead we apply simple hill-climbing in a relative fitness environment. These results and further analysis suggest of Tesauro's program had more to do with thethat the surprising success of the learning task and the dynamics of theco-evolutionary structure backgammon game itself. 1 INTRODUCTION It took great chutzpah for Gerald Tesauro to start wasting computer cycles on temporal of Backgammon (Tesauro, 1992). After all, the dream ofprogram play itself in the hopes computers mastering a domain by self-play or "introspection" had been around since the early days of AI, forming part of Samuel's checker player (Samuel, 1959) and used in Donald Michie's MENACE tictac-toe learner (Michie, 1961). However such self-conditioning or nonexistent internal representations, had generally beensystems, with weak of scale and abandoned by the field of AI.

We here discuss insights gained about the structure of evaluation functions for a large domain such as backgammon. Evaluation began as a single linear polynomial of backgammon features. Later, we introduced Mate-classes, each with its own evaluation function. This improved the play, but caused problems with odge-effects between state-classes. Our latest effort uses models of position potential to select across the set of best members of each represented state-class. "This has produced a significant jump in performance of BKG. Because of the localization of knowledge, state-classes permit relatively easy modification of knowledge used in evaluation. They also permit the building of opponent models based upon what evidence shows the opponent knows in each state-class.