An Ecological Approach to Agent Population Management

The problem of maintaining a desired number of mobile agents on a network is not trivial, especially if what is required is a completely decentralized solution. Decentralized control makes a system more robust and less susceptible to partial failures. The problem of agent population management is exacerbated on wireless ad hoc networks where host mobility can result in significant changes in the network size and topology. System stability is also of critical importance. This paper analyzes the stability of a previously proposed ecology-inspired approach to agent population management, and proposes improvements. The stability of the new ecology based strategy is proved theoretically, and the conclusions are verified with a set of experiments. Introduction The problem of balancing supply and demand on multiagent systems (MAS) is not new, and is becoming more prevelent as MAS and mobile ad hoc networks (MANET) become more popular. One specific type of supply-demand relationship is the demand for jobs of a certain kind and the supply of agents to complete those jobs. Satisfying this balance is critical to system performance and stability, since an insufficient number of agents forces jobs to accumulate over time, preventing the system from benefiting from the parallel nature of the networked MAS. On the other hand an excessive number of agents slow the system down with unnecessary processing and communications overhead. This problem is amplified if the agent system is running over a MANET, since such networks may be composed of lightweight devices that are restricted in terms of processing power, memory, available bandwidth and battery power. Previously, attempts to solve this problem used auctioning, voting, cooperative negotiation (Conry et al. 2004) and distributed constraints satisfaction (Mailler & Lesser 2004). Because dynamics of MANETs make centralized approaches impossible, we present an analysis of an ecology based approach that does not require centralization. There are two major approaches to simulating an ecosystem (Flake 1998). One is a species-based view of the system, where large classes of individuals interact in the simulation (e.g., modeling the dynamics of species interaction rather than the Copyright c © 2005, American Association for Artificial Intelligence (www.aaai.org). All rights reserved. interaction of individuals). The second approach is a bottom up construction of an ecological simulator, focusing on simulating individuals and their interactions. What is often missing in current research is the correlation between these two levels. Connecting these levels allows us to: • Predict the system behavior on the global level; • Deduce the set of local interactions leading to desired global behavior; and • Analyze the stability of global behavior based on the stability of local interactions. In this respect, this paper focuses on species-based analysis of individual-based software agent systems. Related Work Agent System Stability Service replication. This approach to improving system robustness looks at replication of agents and/or services on the MAS network. It was investigated by several researchers (Gartner 1999; Marin et al. 2002). Most focus on the methodology of agent/service replication. This paper provides the extended analysis of the agent replication techniques based on an ecological model originally proposed by Peysakhov: (Peysakhov, M.; Cicirello, V.; Regli, W. 2004). Probabilistic models. This research assumes some uncertainty in agent behavior or the agent’s environment, and suggests techniques for estimating and improving the state of the agent-based system. One of the first researchers to analyze probabilistic survivability in an MAS was Kraus in (Kraus, Subrahmanian, & Tacs 2003). Kraus proposed a probabilistic model of MAS survivability founded on two assumptions: 1. The global state of the network is always known. 2. The probabilities of host or link failure are known. Unfortunately, these assumptions rarely hold in real life (Leland & Porche 2004). An alternative approach enables agents to reason about the state of the network and security of their actions (Peysakhov et al. 2004). Stability theory. The problem of system stability is by no means unique to agent systems. Agent based systems are a subset of a much larger class of discrete event systems. Ramadge and Wonham (Ramadge & Wonham 1989) summa-