Plan creation, plan execution and knowledge acquisition in a dynamic microworld

This paper discusses a planning system that works in a dynamic geographic microworld (an abstraction of a small city), The planning system simulates a taxi driver who must take a customer to a destination in this small city, despite the possible presence of unpredictable dynamic obstacles such as other cars, red lights, pedestrians, etc. At the heart of the system is a knowledge base where all knowledge of the city is uniformly represented in routes, one for each trip which the simulated taxi driver has previously completed. A plan for a new trip is formulated by appropriately splicing together routes. The plan is then augmented to handle the dynamic features of the world and is executed. After successful execution, the plan is integrated into the knowledge base as a new route. One key aspect of this approach to planning is that plans are hierarchical, thus allowing the appropriate scoping of “demons” to handle the dynamic features of the microworld. Another key aspect is the similarity in structure of both routes and plans which facilitates the route-splicing approach to planning. Both aspects are useful when acquiring new routes, the similarity in structure allowing successfully executed plans to be straightforwardly added to the collection of routes and the hierarchy providing a framework for abstracting and generalizing information gleaned during plan execution. The most important overall lesson to be learned from this research is the usefulness of taking an integrated view of planning, execution and acquisition.