Figure 7: A backgammon endgame position (from [Pollack et al.
1996])
The domain of backgammon fits naturally in this research for several reasons: it was the subject of my Master's thesis [Boyan1992], it is the most celebrated and most successful application of value function approximation [Tesauro1992], and it also appears to be suitable for direct meta-optimization approaches [Pollack et al. 1996]. Backgammon provides a perfect setting for developing both approaches and putting them in head-to-head competition.