Agent-Organized Networks for Multi-Agent Production and Exchange

As multi-agent systems grow in size and complexity, social networks that govern the interactions among the agents will directly impact system behavior at the individual and collective levels. Examples of such large-scale, networked multiagent systems include peer-to-peer networks, distributed information retrieval, and agent-based supply chains. One way of dealing with the uncertain and dynamic nature of such environments is to endow agents with the ability to modify the agent social network by autonomously adapting their local connectivity structure. In this paper, we present a framework for agent-organized networks (AONs) in the context of multiagent production and exchange, and experimentally evaluate the feasibility and efficiency of specific AON strategies. We find that decentralized network adaptation can significantly improve organizational performance. Additionally, we analyze several properties of the resulting network structures and consider their relationship to the observed increase in organizational performance. Introduction and Related Work The success of both real and artificial organizations is dependent upon a structure that facilitates effective and efficient behavior at the individual and organizational levels. In many multi-agent system applications, groups of agents must coordinate to solve problems, efficiently distribute goods or services, form teams to accomplish tasks, and collect and share information. In these domains, the organizational structure, or the agent social network, has a direct impact on the performance of the agent society. Our goal is to enable agents to autonomously adapt their local network connectivity, providing organizations with a level of social intelligence and an organizational learning capability. In this paper, we develop an approach for implementing agent-organized networks (AONs) in the context of a multiagent production and exchange economy (Wilhite 2003). Potential application domains for this work include the management and formation of supply chain networks (Geunes & Pardalos 2003; Fox, Barbuceanu, & Teigen 2000), peer-topeer (P2P) networks (Ramanathan, Kalogeraki, & Pruyne 2002), and distributed information retrieval (Yu & Singh Copyright c © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. 2003). Researchers have recently emphasized the necessity of decentralized, bottom-up supply chain formation: “To achieve the oft-expressed visions of dynamically forming and dissolving business interactions requires automated support for supply chain formation, the process of bottom-up assembly of complex production and exchange relationships.” (Walsh & Wellman 2000) Related to our work in economically motivated AONs, network formation has been studied in economic game theory (Jackson 2003), including applying reinforcement learning to refine partner selection in repeated games (Skyrms & Pemantle 2000). Small-scale trade networks and partner selection have been studied in agent-based computational economics (Tesfatsion 1997), and reinforcement learning has been proposed as a way to learn effective agent interactions based on reputation in multi-agent market environments (Tran & Cohen 2003). Dutta and Sen examined learning cooperative relationships for efficient task completion using a “simple reinforcement scheme” (2003). Our work extends and differs from the previous research in two important ways. First, we take a bottom-up approach to network formation, maintaining the resource, cognitive, and communication limitations implied by an initial agent social network. Second, we explicitly analyze the structure of the agent social network resulting from various decentralized adaptation strategies. Two assumptions guiding our approach are the lack of a central broker or global search capability and the assumption that the agents may not be under the control of a single authority. The goals of this paper are to provide a framework for bottom-up AONs and to demonstrate the feasibility of applying AONs to improve the performance of an organization of economically motivated agents. We are particularly interested in developing agent-initiated network adaptation strategies that are realistic, feasible, and efficient. We first describe a general multi-agent production and exchange model and the AON framework. We then present experimental results demonstrating the feasibility of AONs for multi-agent production and exchange. Multi-Agent Production and Exchange In order to study mechanisms for AONs in multi-agent economies, we selected a simple, generic, yet realistic model of a production and exchange economy. In this section, we describe the model, and discuss the effects of agent social structures on the organization’s ability to distribute goods. A Model of Production and Exchange The basis for the model was first presented by Wilhite (2001; 2003). Each agent is given an initial endowment of two distinct goods, and has a fixed production capacity. At each time step, each agent is allowed to choose whether to trade or to produce. The goal of the individual agents is to maximize their utility. The model assumes that agents are purely selfish (i.e., they select the action that maximizes their utility) and completely truthful (i.e., they always provide perfect information during negotiation and trade). Let there be n agents in the economy and two goods, g1 and g2, where g2 is infinitely divisible and g1 must be traded in whole units.1 The utility of agent i is