Resource-Bounded Reasoning in Phoenix

I started working in resource-bounded reasoning as part of the PHOENIX project (Cohen et al. 1989; Hart & Cohen 1990; Greenberg & Westbrook 1990), while a graduate student in the Experimental Knowledge Systems Laboratory (EKSL) at UMass, Amherst. The PHOENIX problem domain is forest fire fighting, for which EKSL built a sophisticated computer simulation of forest fires in Yellowstone National Park (based on satellite data) and of autonomous agents that put out the fires. The PHOENIX system consists of an instrumented discrete event simulation, an autonomous agent architecture that integrates multiple planning methods, and a hierarchical organization of agents capable of improving their fire-fighting performance by adapting to the simulated environment. The PHOENIX agent architecture includes several innovative features that support adaptable planning under real-time constraints, including a least-commitment planning style called lazy skeletal refinement and a combination of reactive and deliberative planning components operating at different time scales. The problem involves resourcebounded reasoning because the simulator ensures that the fire continues to burn while the agents are planning and acting. The deadlines are somewhat soft, because the agent can gain time by sacrificing more forest, but it sets a deadline for itself as part of committing to a plan, and it is costly to replan at that point. In any event, there is a natural time pressure on the agent, which it can monitor by observing the fire as well as its own progress.