Decision Theoretic Execution Strategies for Modularized Problem Solvers

Large-scale problem solving algorithms will prove to be critical in implementing the long-term vision of cyberphysical systems (CPS). Such algorithms will be highly modularized -that is, they will use subroutines implemented behind abstract interfaces. A characteristic of many problem solving algorithms is that their performance varies wildly depending on which order subroutines are executed, with which parameters, and which computational resources. These choices must be made in the face of unpredictable subroutine behavior. We propose a formalism that treats the problem solving algorithm as a decision-making agent in a belief space, casting the problem of choosing a subroutine execution strategy as a Markov decision process (MDP). This formalism has already been applied successfully to many existing problems in robotics and AI, and holds promise for CPS applications and fundamental CPS research in the future.