The fields of distributed AI and multi-agent systems have traditionally studied task decomposition and collaboration among sets of agents. Most of the work in these domains has focused on situations where there is a high level of communication and coordination among agents: agents modeling each other [16], for example, or explicit agent negotiation [4] [13] [18]. In particular, work by Lander and Lesser [9] takes a task negotiation approach to the problem of distributed search. Somers [17] and Garijo et al. [6] apply distributed AI techniques to manage networks in a decentralized fashion. Our work is similar in that we also focus on decentralized cooperating agents, but differs in that our agents are mobile, living in the network, and tend to be less knowledge-based.
Mobile agents are an active and exciting research topic, especially as contemporary tools such as Java make mobile agent systems relatively simple to implement. A number of arguments exist as to why the mobile agent approach is potentially useful [3] [20]. Much research has focused on performance enhancement from resource distribution, but we believe that the most interesting and novel possibilities of mobile agent architectures lie in their adaptive nature and inherent flexibility [1] [7]. We share some commonalities of direction with developers of experimental active packet infrastructures [19], including a concern with the dynamic characteristics of network systems and a desire to explore the possibilities of movable bits of code in the context of network protocols. Our mobile agents could be recast as active packets -- active routers host the mobile agent programs, which in turn discover the connectivity of the routers and provide topology information to the network.
There are several threads of research on using mobile agents situated in a telecommunications network to manage connectivity and load balancing. Appleby and Steward's work is an early paper suggesting using mobile agents with AI-like strategies to dynamically load-balance a telecommunication network. Followup work using a design inspired by ant behavior [2] [15] extends these ideas in a new direction, using markers in the network environment as a means of indirect inter-agent communication (``stigmergy''). We believe our work is complementary to the ant-based research; the environmentally-mediated approach is efficient and elegant, but trading structured stores of information can be transparent and powerful.
Finally, our particular result regarding the importance of population diversity for efficiency of multi-agent systems has been a theme in biologically-inspired computer system design. Huberman has studied performance characteristics of cooperative processes [8], looking at the concurrent search of a space of possibilities by an ecology of agents and at the effectiveness of hints exchanged among those agents. He also concludes that diversity increases effectiveness. In the field of genetic algorithms [11], the tradeoff between population diversity and efficiency is a well known, if not entirely solved, problem. Results from genetic algorithm work [5] [10] [14] suggest that explicit efforts to maintain population diversity can make genetic algorithms perform more efficiently.
