Adaptive Dynamic Programming for Real-Time Cooperative Multi-Player Games and Graphical Games

点击翻译按钮获取中文摘要

The Objective of this research is to design new algorithms for decision and control in multi-player games for complex human-engineered systems interacting on communication graph topologies. Standard differential game solutions are for systems with a single dynamics and multiple action inputs. However, realistic systems are composed of agents having their own individual dynamics and only interacting with their immediate neighbors in a social graph topology. The Approach is to bring together discoveries in neurobiological learning, sociobiological systems having local interactions between agents, and multi-player differential games to develop novel feedback control structures for nonlinear dynamical systems. Intellectual Merit. Standard differential game theory solutions are generally offline design methods that rely on solving nonlinear design equations requiring knowledge of full system dynamics. Approximate Dynamic Programming techniques based on reinforcement learning will be used to develop novel adaptive control structures that learn game theory solutions online in real time using data measured along the system trajectories and without knowing full system dynamics. Solutions for different definitions of game equilibria will be sought including Nash, Pareto, Nash bargaining, and cooperative games.Broader Impacts. The research will help bridge the gap between the Computational Intelligence and the Control Systems communities by bringing together reinforcement learning, game theory, and differential dynamical systems. Applications will be made to cooperative control of distributed electric power microgrids for renewable energy such as wind and solar generation. Existing programs at UTA will be expanded in women in engineering, research for US students, high school engineering technology, and K-12 outreach.