Performances analysis in collective systems

The multi agent approach has been used for several years to study complex systems and to give new techniques of resolution both in artificial life to simulate and to analyse insect societies (E. Bonabean and G. Theraulaz, 1994; J.-L. Deneubourg and S. Goss, 1989; 1991), and in robotics to solve problems such as the collecting or the sorting out of elements in a dynamical environment (R. Brooks, 1986; J.-L. Deneubourg and S. Goss, 1991). The reactive agent architecture is based on a simple process of action-reaction often extended with capabilities of adaptation and learning. However, studies that have been carried out on these systems suffer from a lack of formalism, in particular when performances are evaluated. The experimental approach, based on a direct observation (of real or simulated systems) does not allow for quantitative analysis. Mathematical models have been proposed to analyse the behaviour of action-selection, agent specialization and collective work among insects. But these studies give better results on individual agent behaviour than on global collective performances. The study proposes a method to compute the global performances of collective systems given the behaviour of agents, the environment and the kind of events that can happen. Difficulties lie in the fact that these processes contain a lot of random events. Therefore, the problem consists of modelling the system with the right level of description. Thus we do not study issues that are based on emerging phenomena because, as M. Mataric (1994) emphasizes, it is impossible to determine them without testing the system.