Control Structures for Software Agents

Abstract

Software agents have shown considerable promise in many fields. Thus, control structures of such software agents raise important research issues. This dissertation describes control structures designed for software agents that solve real world problems. The dissertation also presents principles of software agents' control mechanisms discovered by doing the research. The first software agent named SUMPY was designed and implemented to test a new control structure. SUMPY lives in and helps to maintain a UNIX file system for better disk space utilization by compressing and backing up files in the system. Built using subsumption architecture, SUMPY displays a plug and play property. A new UNIX maintenance task can be added to SUMPY's repertoire without modification of existing layers performing earlier tasks. One of SUMPY's layers supports a fuzzy control mechanism enabling it to achieve its goals in a real-world manner. Another restricts SUMPY's activity to a light CPU load. Experiments have demonstrated the properties of the architecture and advantages of using SUMPY for such maintenance tasks. A multiple agent version of SUMPY, MultiSUMPY, has been designed and implemented, aimed at demonstrating collaborative behavior among agents and collision avoidance behavior. MultiSUMPY inherits SUMPY's features and extends them along several dimensions, especially in collaboration among agents, system robustness and efficiency. The third agent system, VMattie, has been designed and implemented for exploring an action selection theory toward collaboration and cognitive features. The VMattie architecture has been inauenced by Maes' behavior network architecture and Hofstadter and Mitchell's Copycat architecture and significantly extends them. The instantiation-based action selection mechanism empowers VMattie in that the VMattie behavior network can handle variables very easily, an open problem as Maes pointed out. The Codelet concept enables VMattie to perform behaviors in a natural, parallel, collaborative, and dynamic way. The VMattie architecture enables VMattie to actively gather information from humans, compose announcements of next week's seminars, and mail them each week to a list that she keeps updated, all without the supervision of a human. By doing the research on the three agent systems, a number of principles have been discovered for designing software agent systems.