next up previous
Next: Related Work Up: Miscellaneous Issues Previous: Fail-safe Planning

Contingent Failure

Cassandra can produce a plan only if it is possible to achieve the goal of the plan in all possible contingencies. Often, however, the goal cannot in fact be achieved in some outcome of the underlying uncertainty. Consider, for instance, Peot and Smith's example of trying to get to a ski resort by car, when the only road leading to the resort is either clear or blocked by snowdrifts [Peot and Smith 1992]. If the road is clear, then the goal can be achieved, but if it is blocked, all plans are doomed to failure.

No planner can be expected to recognize the impossibility of achieving a goal in the general case [Chapman 1987]. However, a possible approach is suggested by Peot and Smith. We could introduce an alternative method of resolving open goal conditions: simply assume that the goal in question fails.

This is an undesirable method of resolving open goal conditions if the subgoal is in fact achievable, so in theory plans involving contingent failure should be considered only after the planner has failed to find a plan in which all goals are achieved. This is sometimes possible, but in general the problem of determining whether there is a successful plan is undecidable. There may always be partial plans that do not involve goal failure but that cannot be completed. For example, as a partial plan is modified it may become more and more complex, the resolution of each open condition involving the introduction of more unachieved subgoals. In this case, plans involving contingent failure will never be considered unless they are ranked above some plans that do not involve contingent failure. In order to be generally useful, the approach must be weakened: instead of considering goal failure only after all other avenues of attack have failed, apply a high fixed penalty to plans involving failed goals. The aim would be to fix the penalty high enough that contingent failure would only apply in genuine cases of goals being unachievable. However, this would of necessity be a heuristic approach and completeness would be lost.



next up previous
Next: Related Work Up: Miscellaneous Issues Previous: Fail-safe Planning



Louise Pryor <louisep@aisb.ed.ac.uk>;
Last modified: Mon Mar 18 18:08:59 1996