Anticipating Computational Demands when Solving Time-Critical Decision-Making Problems

An agent embedded in a dynamic environment may need to respond in a timely manner to sequences of events outside the agent's control. By anticipating computational demands and allocating processing time accordingly the agent can avoid costly delays often arising from trying to respond to a dynamic environment with high-complexity decision procedures. Deliberation scheduling, the process of allocating processing time among competing decision procedures to explicitly account for the costs and beneets of computational delays, may aid an agent that must solve time-critical decision-making problems in which the time spent in decision-making aaects the quality of the responses generated. The more accurate the agent is in anticipating the computational demands of forthcoming problems the more successful it can be in allocating its decision-making time. We present an approach to solving time-critical decision-making problems by taking advantage of domain structure to expand the amount of time available for processing diicult combinatorial tasks. Our approach uses predictable variability in anticipated computational demands to allocate on-line deliberation time and exploits problem regularity and stochastic models of environmental dynamics to restrict attention to small subsets of the state space. This approach demonstrates how slow, high-level systems (e.g. for planning and scheduling) might interact with faster, more reactive systems (e.g. for real-time execution and monitoring) and enables us to generate timely solutions to diicult combinatorial planning and scheduling problems such as the traac control of multiple robot vehicles.