IEEE TRANSACTIONS ON KNOWLEDGE AND DATA ENGINEERING, VOL. 4, NO. 2, APRIL 1992 123 Organization Self-Design of Distributed Production Systems Toru Ishida, Les Gasser. and Makoto Yokoo Abstmct4rganiz&ion has emerged as a key concept for struc- turing the activities of collections of problem-solvers. Orguniza- tion self-design (OSD) has been studied as an adaptive approach to long term, strategic work-allocation and load-balancing. In this paper, we introduce two new reorganization primitives, conpo- sition and decomposition. They change the population of agents and the distribution of knowledge in an organization. To create these primitives, we formalize organizational knowledge, which represents knowledge of potential and necessary interactions among agents in an organization. We develop computational OSD techniques for agents with architectures based on production systems to take advantage of the well-understood body of theory and practice. We first extendparallelproduction systems, where global control exists, into distributed production systems, where problems are solved by a society of agents using distributed control. We then introduce OSD into distributed production systems to provide adaptive work allocation. Simulation results demonstrate the effectiveness of our approach in adapting to changing envimn- mental demands. In addition to introducing advanced techniques for flexible OSD, our approach impacts production system design, and improves our ability to build production systems that can adapt to changing real-time constraints. Index Terms-Adaptive problem solving, organization self- design, parallel and distributed processing, production system, real-time problem solving. I. INTRODUCTION I T has been clear for some time that organization is a powerful concept for thinking about how to structure the interactions of collections of problem solvers. Understanding the concept of organization and developing techniques for adaptive reorganization are pressing concerns in distributed artificial intelligence (DAI) [3]. Several conceptual approaches to organization have been introduced, including treating organ- ization as 1) a long term, strategic load-balancing technique [5], 2) a structural set of control and communication relation- ships among agents [28], 3) sets of interaction patterns among agents [19], [12], 4) sets of commitments and expectations among agents, [3], [14], [20], or 5) collections of settled and unsettled questions about knowledge and action [13]. The comparative information processing performance of rigid organization structures was studied by Malone [28]. However, since no single organization is appropriate in all Manuscript received July 1, 1991. This paper is the extended version of the authors’ previous AAAI conference papers [25], (151. T. Ishida and M. Yokoo are with NTI Communication Science Laborato- ries, Sanpeidani, Inuidani, Seika-cho, Soraku-gun, Kyoto, 619.02, Japan. L. Gasser is with the Department of Computer Science, University of Southern California, Los Angeles, CA 90089. IEEE Log Number 9106253. situations, organization self-design (OSD) has been proposed to allow an organization of problem solvers to adapt itself to dynamically changing situations [5]. In this paper, we further explore the process of OSD, and, in doing so, we examine some new ideas about the nature and representation of organizations which are the foundations of OSD. We address OSD by introducing the following new concepts: l Organizational knowledge: To perform either domain problem solving or reorganization, agents need organi- zational knowledge, which represents both the necessary interactions among agents and their organization. How- ever, the kind of organizational knowledge required for reorganization has not been thoroughly investigated in prior research. In this paper, we formalize organizational knowledge as a collection of agent-agent relationships and agent-organization relationships, which represent how agents’ local decisions affect both other agents’ decisions and the behavior of the entire organization. l Reorganization primitives: In previous research, reorga- nization mechanisms typically changed agent roles or inter-agent task ordering [5], [7] [9]. In this paper, how- ever, we take the approach of formalizing reorganization primitives, which can perform OSD through repeated ap- plication. The new reorganization primitives, composition and decomposition of agents, dynamically change inter- agent relationships, the knowledge agents have about one another, the size of the agent population, and the resources allocated to each agent. Up to now, OSD has been investigated using compar- atively complex agents, such as blackboard-based agents. However, here we discuss OSD using a problem-solving model based on production systems, to take advantage of a well-understood body of theory and practice, while retaining general applicability. Production systems have the advantage of providing a formal characterization of both the knowledge needed to solve a problem and the ways in which parts of that knowledge interact. In addition, production rules can be used as general abstractions of organizational and problem-solving processes of many kinds. (For example, Zisman has provided a well-known application of production systems to modeling asynchronous organizational work and problem-solving [34].) Though we use production systems here as a theoretical and modeling foundation, our concepts of OSD and organizational knowledge can be generalized to apply to other problem- solving models and other types of problem solvers. In addition to advancing OSD techniques, our approach im- pacts production system design. Previous research, attempted 1041-4347/92$03.00 0 1992 IEEE