Resource-dependent behavior through adaptation.

Abstract

An autonomous agent functioning in the real world consumes finite resources as it operates, creating resource constraints upon its execution. The dynamics of the real world cause the environment, including those resource constraints, to change in ways the agent can neither predict nor control and in ways the agent cannot always be preprogrammed to handle. Behavior that is reactive with respect to (responsive to changes in the availability of) those resources is desired in such a situation, but the agent will be able to arrive at this reactivity only through experience in its environment. The agent's success in a dynamic environment will depend upon its ability to adapt to the conditions and constraints imposed upon it by its environment--its ability to create Resource-Dependent Behavior Through Adaptation. The dynamics of the real world necessitate a view of resources, and of the agent's knowledge about its resource consumption, as being dynamic rather than static entities. In my dissertation I present RDB, an agent capable of producing behavior that is sensitive to the availability of critical resources and to its own knowledge about its resource consumption. RDB observes its own execution and monitors its critical resource consumption, allowing it to learn about its requirements from experience--RDB can both acquire new estimates for previously untested actions and can adjust existing consumption estimates that prove to be inaccurate. RDB builds new plans to meet new situations and associates global consumption estimates to the rules forming the new plan, which allows it to determine the plan's overall needs. At each step of execution RDB considers these global estimates and compares them to the available resources it is monitoring and changes its behavior accordingly. RDB is a flexible system in terms of both behavior and knowledge. Its resource knowledge is continually under development in terms of both its quality and its quantity. RDB can use resource knowledge at any stage of its development to help guide planning and execution. Successful applications of RDB demonstrate the high degree of adaptivity to dynamic environments that can be achieved by systems combining learning, planning, and reactivity.