Planning agents in JAMES

Testing is an obligatory step in developing multiagent systems. For testing multiagent systems in virtual, dynamic environments, simulation systems are required that support a modular, declarative construction of experimental frames, that facilitate the embedding of a variety of agent architectures and that allow an efficient parallel, distributed execution. We introduce the system JAMES (a Java based agent modeling environment for simulation). In JAMES, agents and their dynamic environment are modeled as reflective, time-triggered state automata. Its possibilities to compose experimental frames based on predefined components, to express temporal interdependencies, to capture the phenomenon of proactiveness and reflectivity of agents are illuminated by experiments with planning agents. The underlying planning system is a general-purpose system, about which no empirical results exist besides traditional static benchmark tests. We analyze the interplay between heuristics for selecting goals, viewing range, commitment strategies, explorativeness, and trust in the persistence of the world and uncover properties of the the agent, the planning engine, and the chosen test scenario: TILEWORLD