Organization properties of open networks of cooperative neuro-agents.
Jean-Pierre Mano, Pierre Glize · The European Symposium on Artificial Neural Networks · 2005
Our researches on adaptive systems are inspired by their ability of building by themselves a representation of their surrounding world. Using cooperation as a local criterion of self-organization, we study in a network of neuro-agents the evolution of the system and in the same way, the emergence of a functioning coherent with the environmental feed-back. In this paper we expose the abilities of the neuro-agents that give them the autonomy required to build an ab nihilo topology. And finally we want to emphasize the dynamic of the network organization that is basically its best property of adaptation to a changing environment. The work presented here concerns our approach to self-organization in a dynamic neural network. We present some experiments of emergent learning based on our knowledge of complex systems through multi-agent systems technology and cellular biology. In fact we are most interested in the mechanisms involved in the construction of a network able to stabilize its structure in a given environment, than in the efficiency of the learning supported by this structure. Obviously those two ontogenetic processes are strongly coupled and appear simultaneously as mutual consequences. Under those conditions, we conceive artificial systems presenting an inherent complexity close to biological systems as they are open and as their structure evolves from nothing (at least an empty shell) up to an organized, stable and functional network of neuro-agents. Our working hypothesis tells minimal entities of such a system can be built and given with sufficient behavioral potentialities in order to organize themselves to satisfy the whole system activity. In this study, we want to show that a coherent and useful system can emerge without any global model for its functioning, solely from activity and organizational rules of its sub-parts; only agents are designed, not the topology of the network they will belong to. Our thought process supposes four distinct steps that description will be formerly extended. Firstly the neuro-agent unit has to be designed, their properties, their abilities (calculus, communication) and their knowledge. Secondly, initial system is constituted of a minimal number of unlinked neuro-agents. Those pre-built neuro-agents are only receptors (inputs) and actuators (outputs) of the network. Thirdly the global system has to be plunged in a dynamic and open environment in which it will have to adapt to; regarding neuromimetic systems, this phase suppose