next up previous contents
Next: Beliefs about plan rules Up: Improvements to design features Previous: Improvements to design features   Contents

Evaluation of game trees using a logical model of belief

While the planner has been developed using the maximum expected utility rule for evaluation of game trees, an alternative evaluation rule is to check whether particular dialogue outcomes are possible, and to choose the current act based on those. The probability of the outcomes is of no concern. This was discussed in Section 3.4.5, but was not implemented. Such evaluation would be an implementation of Pollack's [53] claim that agents plan dialogues using different sets of beliefs about plan rules and the domain state, but would retain the traditionally preferred logical model of nested beliefs. Such planning would be interesting where it is very important that there is no possibility of plan failure. Implementation of this idea is straightforward - the continuous utility function should be replaced with a two valued one of success and failure, and the maximum expected utility rule used at chance nodes should be replaced by one that selects the worst outcome of the two alternatives, given the evaluating agent's beliefs. A less elegant but effective way to achieve the same effect with the current planner is to use extremely large negative utility values to represent failures. These outcomes would be rejected even if the probability of the outcome is small, if the negative utility is large enough.


next up previous contents
Next: Beliefs about plan rules Up: Improvements to design features Previous: Improvements to design features   Contents
bmceleney 2006-12-19