Theory and Applications of Difference Evaluation Functions (Doctoral Consortium)

The credit assignment problem (which agents get credit or blame for system performance) is a key research area. For a team of agents collaborating to achieve a goal, the effectiveness of each individual agent must be calculated in order to give adequate feedback to each agent. We use the Difference Evaluation Function to find agent-specific feedback. The Difference Evaluation Function has given excellent empirical results in many domains, including air traffic control and mobile robot control. Though there has been some theoretical work that shows why Difference Evaluation Functions improve system performance, there has been no work to show when and under what conditions such improvements are realized. We apply an evolutionary game-theoretic analysis to show the theoretical advantages of the Difference Evaluation Function. We then focus on how to apply these multiagent learning methods to optimize distributed sensor networks in advanced power generation systems.