Transformational Planning of Reactive Behavior

Abstract

Abstract : Reactive plans are plans that include steps for sensing the world and coping with the data so obtained. We investigate the application of Al planning techniques to plans of this sort in a simple simulated world. To achieve fast reaction times, we assume that the agent starts with a default reactive plan, while the planner attempts to improve it by applying plan transformations, thus searching through the space of transformed plans. When the planner has what it believes to be a better plan, it swaps the new plan into the agent's controller. The plans are written in a reactive language that allows for this kind of swapping. The language allows for concurrency, and hence, truly 'nonlinear' plans. The planner evaluates plans by projecting them, that is, generating scenarios for how execution might go. The resulting projections give estimates of plan values, but also provide clues to how the plan might be improved. These clues are unearthed by critics that go through the scenario sets, checking how the world state and the agent state evolved. The critics suggest plan transformations with associated estimates of how much they will improve the plan. Plan transformations must be able to edit code trees in such a way that the changes are orthogonal and reversible whenever possible. The system has been tested by comparing the performance of the agent with and without planning. Preliminary results allow us to conclude that the planner can be fast and directed enough to generate improved plans in a timely fashion, and that the controller can often cope with a sudden shift of plan.