Previous: Static Analysis Tables
Up: Constructing a Metagame-player
Previous Page: Static Analysis Tables
Next Page: Examples of Material Analysis

Weights for Advisors

The last major issue concerning the construction of the strategic evaluation function involves assigning weights to each advisor, or more generally, developing a function for mediation among advisors [\protect\citenameEpstein, 1989a]. While this issue is already difficult in the case of existing games, it is correspondingly more difficult when we move to unknown games, where we are not even assured of the presence of a strong opponent to learn from. However, by the construction of some of the advisors, we do have one significant constraint on their possible values. For advisors which anticipate goal-achievement (such as promote-distance and the threat advisors), it would seem that their weight should always be at most 1. The reason is that the value they return is some fraction of the value which would be derived if the goal they anticipate were to be achieved. If such an advisor were weighted double, for example, the value of the threat would be greater than the anticipated value of its execution, and the program would not in general choose to execute its threats.

Beyond the constraint on such advisors, this issue of weight assignment for Metagame is an open problem. One idea for future research would be to apply temporal-difference learning and self-play [\protect\citenameTesauro, 1993] to this problem. It would be interesting to investigate whether the ``knowledge-free'' approach which was so successful in learning backgammon also transfers to these different games, or whether it depends for its success on properties specific to backgammon. In the meantime, we have been using METAGAMER with all weights set to 1.