Distributed Problem Solving using Evolutionary Learning in Multi-Agent Systems

This chapter presents a new framework for solving distributed control problems in a cooperative manner via the concept of dynamic team building. The distributed control problem is modeled as a set of sub-problems using a directed graph. Each node represents a sub-problem and each link represents the relationship between two nodes. A cooperative ensemble (CE) of agents is used to solve this problem. Agents are assigned to the nodes in the graph and each agent maintains a table of link relationship with all the other nodes of the problem. In the cooperative ensemble, each agent generates three sets of outputs iteratively based on the input variables it receives. They are, namely, the need for cooperation, the level of cooperation and the control directives. These outputs are used for dynamic team building within the cooperative ensemble. Agents within each team can issue a collaborative control directive and they take into account the mistakes of all the members in the team. In addition, each agent has a neuro-biologically inspired memory structure containing the addictive decaying value of all its previous errors and it is used to facilitate the dynamic update of the agent’s control parameters. The cooperative ensemble has been implemented in the form of distributed neural traffic signal controllers for the distributed real-time traffic signal control. It is evaluated in a large simulated traffic network together with several existing algorithms. Promising results have been obtained from the experiments. The cooperative ensemble is seen as a potential framework for similar distributed control problems.