Continuous Refinement of Agent Resource Estimates

Abstract

The challenge we address is to reason about projected resource usage within a hierarchical task execution framework in order to improve agent effectiveness. Specifically, we seek to define and maintain maximally informative guaranteed bounds on projected resource requirements, in order to enable an agent to take full advantage of available resources while avoiding problems of resource conflict. Our approach is grounded in well-understood techniques for resource projection over possible paths through the plan space of an agent, but introduces three technical innovations. The first is the use of multi-fidelity models of projected resource requirements that provide increasingly more accurate projections as additional information becomes available. The second is execution-time refinement of initial bounds through pruning possible execution paths and variable domains based on the current world and execution state. The third is exploitation of additional semantic information about tasks that enables improved bounds on resource consumption. In contrast to earlier work in this area, we consider an expressive procedure language that includes complex control constructs and parameterized tasks. The approach has been implemented in the SPARK agent system and is being used to improve the performance of an operational intelligent assistant application.

Keywords: Artificial Intelligence, Artificial Intelligence Center, AIC, agents, procedural reasoning, plan execution, resources