next up previous
Next: Discussion Up: Related Work Previous: Interleaving Planning and

Reactive Planning

A different approach to the problem of planning under uncertainty is taken in the reactive planning paradigm. In this approach, no specific sequence of actions is planned in advance. Just as for contingency planning, the planner is given a set of initial conditions and a goal. However, instead of producing a plan with branches, it produces a set of condition-action rules: for example, universal plans [Schoppers 1987] or Situated Control Rules (SCRs) [Drummond 1989].

In theory, a reactive planning system can handle exogenous events as well as uncertain effects and unknown initial conditions: it is possible to provide a reaction rule for every possible situation that may be encountered, whether or not the circumstances that would lead to it can be envisaged. In contrast, a contingency planner such as Cassandra cannot handle exogenous events as it cannot predict them. Cassandra and other contingency planners focus their planning effort on circumstances that are predicted to be possible (or likely, in the case of a probabilistic contingency planner such as C-BURIDAN).

It would be possible to represent Cassandra's contingency plans as sets of condition-action rules, by using the causal links and preconditions to specify the conditions in which each action should be performed. However, more reasoning is required at execution time to use reaction rules than is required to execute a contingency plan. Instead of simply executing the next step in the plan, reasoning only at branch points, the use of reaction rules requires the evaluation of conditions on every cycle in order to select the relevant rule.

Louise Pryor <>;
Last modified: Mon Mar 18 18:19:13 1996