Version 2.0 of the program, which had much greater training experience as well as a 2-ply search algorithm, made its public debut at the 1992 World Cup of Backgammon tournament.

13-9, 6-5 -0.014 -0.040 13-9, 24-23 +0.005 +0.005 According to an article by Bill Robertie published in Inside Backgammon magazine [8].

safe play decisions, which is what backgammon really is all about, is nothing short of phenomenal.

As one might expect from such favorable assessments, TD-Gammon has had a rather significant impact on the human expert backgammon community.

This situation confronted Joe Sylvester, the highest-rated player in the world at the time, in the final match of the 1988 World Cup of Backgammon tournament.

If TD-Gammon has been an exciting new development in the world of backgammon, it has been even more exciting for the fields of neural networks and machine learning.

In backgammon this comes about partly due to the dice rolls, and partly due to the fact that one can only move one's pieces in the forward direction.

Based on the surprising quality of results obtained in the backgammon application, it now appears that TD methods may be much more powerful than previously suspected.

In fact, a 3-ply analysis is commonly employed by authors of backgammon texts and annotated matches to explain how to find the best move in certain kinds of endgame situations.

In more complex games such as chess and Go, one would guess that an ability to learn a linear function of the raw board variables would be less useful than in backgammon. In those games, the value of a particular piece at a particular board location is more dependent on its relation to other pieces on the board.

It would also seem that TD approaches to deterministic games would require some sort of external noise source to produce the variability and exploration obtained from the random dice rolls in backgammon. There are many conceivable methods of adding such noise - one could randomize the initial positions at the start of every training game, or one could have the network choose moves according to some probability distribution.

There are a number of methods available to assess the quality of play of a backgammon program; each of these methods has different strengths and weaknesses.