Exploiting C-TÆMS Models for Policy Search

We describe our experience developing software for nding centralized multiagent optimal policies for problems speci- ed in the C-T˘MS modeling language, which differs from more classical state-based modeling languages in specifying task relationships explicitly. This software has been used by the DARPA Coordinators program for evaluating other real- time, decentralized algorithms that do not guarantee optimal solutions. The search algorithm is AO*, but we focus discus- sion on our customization of the search for C-T˘MS, how we addressed uncertain duration, and the resulting challenges in trying to scale to larger problems. We will describe differ- ent techniques for reducing the number of states created and how they succeeded or failed.