SGER: Neural-Network-Based Adaptive Dynamic Programming for Maximizing Survival Probabilities

点击翻译按钮获取中文摘要

Abstract Proposal Number: 0841055Proposal Title: Neural-Network-Based Adaptive Dynamic Programming for Maximizing Survival ProbabilitiesPI Name: Abhijit GosaviPI Institution: Missouri University of Science and TechnologyMost previous work on optimal control and adaptive dynamic programming (ADP) has focused on a kind of best-case optimization, where the system to be controlled is stable and the goal is simply to maximize benefits or minimize costs. This PI has been a leader in the small but important area of risk-sensitive ADP, which still aims to maximize benefits but tries to find less risky strategies for doing so. In this new work, the PI proposes to develop new methods aimed at the worst case situation, where stability or survival can not be guaranteed, and the optimization problem is to maximize the probability of survival. This is a very important class of problems. In addition, he plans to bridge the gap between the decision, management and risk community, where he has worked so far, and the engineering ADP community, by developing new software tools that can apply ADP efficiently to problems with a mix of discrete and continuous variables.Intellectual merit of proposal: This would be the first work to apply ADP (or at least continuous-variable ADP) to an extremely important class of problems. The challenging of maximizing a probability of survival is very important, for example, in trying to understand what brains do. Risk-sensitive ADP (RSADP) will be applied for the first time to the emerging class of ADP methods capable of efficiently coping with discrete and continuous variables at the same time.Broad impact of proposal: New tools in MatLab for this kind of optimization method could be of enormous strategic importance to the development of the entire ADP field. This could also be of more than usual benefit to the PI and to the area he has tried to pioneer at this stage of his career. The results of this project will build the foundation for studying RSADP on numerous problems in engineering where ADP can be used: routing of vehicles, maintenance of structures (especially in the face of risky natural disasters), revenue management in airlines, and supply chains of manufactured products.