Issues for a Contingency Planner

A contingency planner must be able to construct plans that can be expected to succeed despite unknown initial conditions and uncertain outcomes of nondeterministic actions. An effective contingency planner must possess the following capabilities:

• It must be able to anticipate outcomes of nondeterministic actions;
• It must be able to recognize when an uncertain outcome threatens the achievement of a goal;
• It must be able to make contingency plans for all possible outcomes of the various sources of uncertainty that affect a given plan;
• It must be able to schedule sensing actions that detect the occurrence of a particular contingency;
• It must produce plans that can be executed correctly regardless of which contingency arises.

The design of Cassandra addresses these issues. However, there are several issues that have not been addressed:

• We have not considered the problem of determining whether it is worth planning for a particular outcome;
• Cassandra is not a probabilistic planner: it cannot make use of any information about the likelihood or otherwise of any events;
• We have ignored the possibility of interleaving planning and execution (but see Section 7.4);
• Cassandra does not handle exogenous events;
• The version of Cassandra described here cannot solve Moore's bomb in the toilet problem [McDermott 1987]: it can only find plans that involve deciding between courses of action that will succeed in different contingencies (but see Section 6.5.5).