next up previous
Next: Identical Branches Up: Miscellaneous Issues Previous: Dependence on Outcomes

One-sided Contingencies

The preceding discussion notwithstanding, a plan involving no contingencies is not always superior to a plan involving a contingency. This is why a planner might in fact construct a plan like the Western/Ashland one. To take a more clear-cut example, suppose Pat needs $50 to bet on a horse. She might try to borrow the $50 from Chris, but the outcome of this action is uncertain---Chris might refuse. Alternatively, she could rob a convenience store. While the robbery plan would (we shall stipulate) involve no uncertainties, it is a bad plan for other reasons. It would be better to first try to borrow $50 from Chris, and then, if that fails, rob the convenience store. could generate this plan. In order to make it prefer the plan to the contingency-free alternative, however, its search metric would have to take into account the estimated costs of various actions, and to perform something akin to an expected value computation. (See, for example, Feldman and Sproull 1977, Haddawy and Hanks 1992 for discussions of decision-theoretic measures applied to planning.) In order to execute the plan properly, it would also be necessary for it to have some way of knowing that the borrowing plan should be preferred to the robbery plan when if it were possible to execute either of them.

Louise Pryor <>;
Last modified: Wed May 1 11:43:54 1996