Environments for Multi-Agent Systems: First International Workshop, E4MAS, 2004, New York, NY, July 19, 2004, Revised Selected Papers

The modern field of multiagent systems has developed from two main lines of earlier research. Its practitioners generally regard it as a form of artificial intelligence (AI). Some of its earliest work was reported in a series of workshops in the US dating from 1980, revealingly entitled, “Distributed Artificial Intelligence,” and pioneers often quoted a statement attributed to Nils Nilsson that “all AI is distributed.” The locus of classical AI was what happens in the head of a single agent, and much MAS research reflects this heritage with its emphasis on detailed modeling of the mental state and processes of individual agents. From this perspective, intelligence is ultimately the purview of a single mind, though it can be amplified by appropriate interactions with other minds. These interactions are typically mediated by structured protocols of various sorts, modeled on human conversational behavior. But the modern field of MAS was not born of a single parent. A few researchers have persistently advocated ideas from the field of artificial life (ALife). These scientists were impressed by the complex adaptive behaviors of communities of animals (often extremely simple animals, such as insects or even microorganisms). The computational models on which they drew were often created by biologists who used them not to solve practical engineering problems but to test their hypotheses about the mechanisms used by natural systems. In the artificial life model, intelligence need not reside in a single agent, but emerges at the level of the community from the nonlinear interactions among agents. Because the individual agents are often subcognitive, their interactions cannot be modeled by protocols that presume linguistic competence. The French biologist Grass ́e observed that these interactions are typically achieved indirectly, through modifications of a shared environment [1].All interaction among agents of any sort requires an environment. For an AI agent whose interactions with other agents are based on speech act theory, the environment consists of a computer network that can convey messages from oneagent’s outbox to another agent’s inbox. For an ALife agent, the environment is whatever the agent’s sensors sense and whatever its effectors try to manipulate. In most cases, AI agents (and their designers) can take the environment for granted. Error-correcting protocols ensure that messages once sent will arrive in due course. Message latency may lead to synchron