Using grammar induction to model adaptive behavior of networks of collaborative agents

We introduce a formal paradigm to study global adaptive behavior of organizations of collaborative agents with local learning capabilities. Our model is based on an extension of the classical language learning setting in which a teacher provides examples to a student that must guess a correct grammar. In our model the teacher is transformed in to a workload dispatcher and the student is replaced by an organization of worker-agents. The jobs that the dispatcher creates consist of sequences of tasks that can be modeled as sentences of a language. The agents in the organization have language learning capabilities that can be used to learn local work-distribution strategies. In this context one can study the conditions under which the organization can adapt itself to structural pressure from an environment. We show that local learning capabilities contribute to global performance improvements. We have implemented our theoretical framework in a workbench that can be used to run simulations. We discuss some results of these simulations. We believe that this approach provides a viable framework to study processes of self-organization and optimization of collaborative agent networks.