Search Results

This chapter describes efforts to add metalevel control capabilities to the Informed Unroller agent (IU-agent), a scheduling agent based on the Markov decision process (MDP) formalism designed to operate in a cooperative multiagent environment. It begins by reviewing related work in metalevel control, especially in the area of performance profiling. It then provides background information on Markov decision processes and the TAEMS modeling language (a derivative of which was used to represent the IU- agent’s tasks) along with a description of the IU-agent. This is followed by a discussion of the implemented metalevel control approach and experimental results indicating its advantages for the IU-agent. The chapter concludes by presenting lessons learned over the course of implementing metalevel control for the IU-agent and suggestions for future work in metalevel control.

This chapter presents a generalized metalevel control framework for multiagent systems and discusses the issues involved in extending single-agent metalevel control to a team of cooperative agents requiring coordination. It presents a methodology for constructing a class of Markov decision processes (MDPs) that can model the interactions necessary for coordinating metalevel control among multiple agents.

This introductory chapter begins with a brief discussion of the concept of metareasoning. It then provides manifesto in an attempt to present in plain language and simple diagrams a description of a model of metareasoning that mirrors the action-selection and perception cycle in first-order reasoning. Many theories and implementations are covered by this model, including those concerning metalevel control, introspective monitoring, distributed metareasoning, and models of self. An overview of the subsequent chapters is also presented.

This chapter describes metareasoning in an image interpretation architecture called GRAVA (Grounded Reflective Adaptive Vision Architecture), where the goal is to produce good image interpretations under a wide range of environmental conditions. GRAVA employs metareasoning because it provides the mechanisms necessary to support two of the core problems of self-adaptive software—a mechanism for reasoning about the state of the computational system and a mechanism for making changes to it. These are referred to as introspective monitoring and metalevel control, respectively.