In this chapter we studied a first approach to generate suitable rule sets for solving classification problems on systems of autonomous, memory constrained components. It was shown that a multi agent system that uses interacting Pittsburgh-style classifier systems can evolve appropiate rule sets. The system evolves specialists for parts of the classification problem and cooperation between them. In this way the components overcome their restricted memory size and are able to solve the entire problem. It was shown that the communication topology between the components strongly influences the average number of components that a request has to pass until it is classified. It was also shown that the introduction of communication costs into the fitness function leads to a more even distribution of knowledge between the components and reduces the communication overhead without influencing the classification performance very much. If the system is used to generate rule sets to solve classification tasks on real hardware systems, communication cost in the training phase can thus lead to a better knowledge distribution and small communication cost. That is, in this way the system will be more robust against the loss of single components and longer reliable in case of limited energy resources.

In this chapter we studied a first approach to generate suitable rule sets for solving classification problems on systems of autonomous, memory constrained components. It was shown that a multi agent system that uses interacting Pittsburgh-style classifier systems can evolve appropiate rule sets. The system evolves specialists for parts of the classification problem and cooperation between them. In this way the components overcome their restricted memory size and are able to solve the entire problem. It was shown that the communication topology between the components strongly influences the average number of components that a request has to pass until it is classified. It was also shown that the introduction of communication costs into the fitness function leads to a more even distribution of knowledge between the components and reduces the communication overhead without influencing the classification performance very much. If the system is used to generate rule sets to solve classification tasks on real hardware systems, communication cost in the training phase can thus lead to a better knowledge distribution and small communication cost. That is, in this way the system will be more robust against the loss of single components and longer reliable in case of limited energy resources.