An Advisor Concept for Distributed Self-organizing Systems Acting in Highly Connected Environments

We present an extension to the concept of an advisor for distributed self-organizing systems that allows to improve the efficiency of such systems that act in environments where the actions of a system component have an influence on the whole environment or large parts of it. An advisor periodically reviews the history of the system and identifies recurring tasks that the system did not perform well. For these tasks it then computes exception rules for the agents that indicate how to perform these tasks better. These rules are communicated to the agents when communication is possible and are from then on used if they are triggered. To deal with highly connected environments so-called group exceptions are needed that consist of individual rules for several agents and that require the agents to signal to the other agents that in their local view the group rules should be triggered. We instantiated this group advisor concept to a self-organizing system for controlling water distribution networks that is based on digital infochemical coordination. In our experiments, using the group advisor resulted in improvements in energy use between 1 and 21 percent, with an average savings of $ 5,000 per day for randomly generated water demand behaviors for the real network of a small North-American city.