Incorporating simulation-based models into planning systems

General-purpose planners have been proposed but few have shown to work effectively and efficiently enough for enough domains to be really called general purpose. A general-purpose planner that uses a single methodology is often too restrictive and therefore cannot plan effectively for all domains. As planning problems become more complicated, having multiagents of different types in dynamic environments, evaluating candidate plans and choosing the best plan becomes prohibitively complex if not impossible within a single methodology. To overcome this problem, we propose simulation-based planning where simulation is used to evaluate the candidate plans. By allowing appropriate simulation model types to accurately express each type of agent in the domain, the task of measuring the success and effects of each candidate plans is simplified and the resulting evaluation will be more accurate since plans are simulated using dynamic models. We describe an application along with the implementation of simulation-based planning in the domain of mission planning. Possible future experiments related to Soar are also discussed.<>