Mass Programmed Agents at the Service of Simulation Systems

In this paper we introduce a MAS-based methodology for simulating the behavior of large scale systems populated by human individuals with diverse strategies and goals. The great challenge in simulating this kind of systems derives primarily from the need to predict how the different strategic behaviors of the individuals change as a result of changes in system settings, the types of information available by the system and the strategic behavior exhibited by the other individuals. Our unique approach to the problem relies on a Multi-Agents based simulation in which each participating agent is pre-programmed by a different person, in a distributed manner, capturing her strategy to be exhibited in scenarios similar to those that need to be simulated. The general agent architecture used isolates the strategy layer from the rest of the system modules, suggesting two inherent advantages. First, the strategy programming task is not accompanied by any agentor server-communication related programming overhead, thus can be distributed to a large group of individuals with very basic programming skills. This enables capturing a large set of realistic human strategic behaviors in the specific domain the simulation attempts to examine. Second, it removes many of the programmatic constraints associated with managing the development of a distributed system, in which different parts are coded by different people. The paper presents the general guidelines and the basic system architecture and reports a successful experience of applying the proposed principles in the legacy parking space search domain, where most former work is based on a restricted number of basic parking strategies. In particular we utilize the parking space search simulation to demonstrate how the system can support the process of capturing the value of different information models supplied in the simulated environments.