Distributed Learning Strategies for Collaborative Agents in Adaptive Decen- tralized Power Systems

For regenerative electric power the traditional top-down and long-term power management is obsolete, due to the wide dispersion and high unpredictability of wind and solar based power facilities. In the R&D DEZENT1 project we developed a multi-level bottom-up solution where autonomous software agents nego-tiate available energy quantities and needs on behalf of consumers and producer groups. We operate within very short time intervals of assumedly constant de-mand and supply, in our case 0.5 sec (switching delay for a light bulb). We prove security against a relevant variety of malicious attacks. In this paper the main contribution is to make the negotiation strategies themselves adaptive across periods. We adapted a Reinforcement Learning approach for defining and discussing learning strategies for collaborative auto-nomous agents that are clearly superior to previous (static) procedures. We report briefly on extensive comparative simulation