Graphical Games and Predictive Pricing for Distributed Control

Anders Rantzer, Professor, Department of Automatic Control, Lund University, Sweden

Decomposition of large scale control problems using duality has already been a topic at our previous visits to Caltech. Here we will discuss some possible game formulations in more detail. The purpose is to formulate games that solve control problems with graph structure, using independent decision-makers which all act in their own interest. Several versions are possible, with different properties regarding scalability and information flow.                    
                                                                                    
A common idea is to use two types of players; subunits optimizing their local control action in response to prices, and intermediate players adjusting prices to counteract disagreements between neighboring subunits. For dynamic problems, this leads to control action based on prediction of future prices.