Learning Task Patterns to Improve Efficiency and Coordination in Decentralized Autonomic Computing Systems

We present the concept of an efficiency and coordination advisor for decentralized autonomic computing approaches realized as multi-agent systems for dynamic optimization problems. The problem scenarios targeted contain recurring tasks that our advisor identifies over several runs of the autonomous system. It is thus giving the system some limited way to "look into the future". If the solutions created by the autonomous agents of the system are much worse than the optimally possible solution, the advisor creates exception rules for those agents making the wrong decisions for the recurring tasks. This allows them to do better decisions in the future in very specific situations while still retaining all advantages of the autonomic computing approach. Our experiments with dynamic instances of the pickup and delivery problem including recurring tasks show that instances that result in suboptimal behavior of the autonomous agents without advisor can be improved substantially when the advisor is present. Our advisor approach is also successful if the recurring tasks change over time