This project studies an important class of complex structured planning domains called ``service domains'' using simulators and probabilistic action models. Examples of service domains include optimizing emergency response in a typical city, scheduling doctors and nurses in a hospital, administering tasks in a typical office, optimally delivering products to shops from distribution centers. These domains share many characteristics such as relational structure, parallel actions, multi-time-scale decision making, exogenous events, and the need for human interpretable solutions that make them highly challenging. The project develops scalable and principled planning algorithms for service domains through a variety of techniques including a novel hierarchical framework of multi-time-scale optimization, new model-free simulation-based planning algorithms, and model-based planning via composition of first-order decision diagrams. These techniques are applied to the real-world problem of optimizing the fire and emergency response in cities through a collaboration with the fire department of Corvallis, Oregon.
The results of the project include new algorithms and frameworks to solve service domains, prototype implementations of the algorithms in the emeregency response domain, and new testbeds of service domains for research. The broader impact of the work includes more cost-effective emergency response systems, and development of new research-oriented courses, tutorials and special workshops on the next generation decision support systems for service domains.