How to adapt efficiently using distributed resources and multiple models to time varing dynamic systems

Yale University, New Haven CT

Investigators

Abstract

When the environment in which a person or a device is operating is relatively constant, but uncertain, information can be collected slowly and used to make decisions. However, the situation is quite different when critical decisions have to be made rapidly as the environment changes. Such situations arise in medical emergencies, trading in the stock market, conflict management, and taking counter-terror measures. In such cases, there is (i) incomplete information about the situation (ii) uncertain or unreliable information and (iii) limited resources. The aim of the proposed research is to use the given resources as efficiently as possible to find adequate responses. The central idea of the research is to exploit the information coming from distributed sources, and decide at every instant how to combine the information for decision making. Systems that continuously monitor their own performance, and adjust their control strategies to improve the performance are considered. These are both stable and robust for time-invariant plants when the parametric uncertainty is small. However, in recent years the need for analytic tools for reacting effectively to large uncertainties and rapidly varying environments, in the presence of input and output disturbances, has been arising in a variety of fields including biology and medicine, economics and finance, and various engineering problems such as energy management, aircraft and automotive control, and security. In many of these applications, it has been found that classical adaptive algorithms result in large and oscillatory (and possibly unstable) responses. To address such problems, the use of different methods based on multiple models has been proposed by the PI. These include a) Switching b) Switching and Tuning c) Interactive adaptation and d) Second Level adaptation. The first method involves switching between fixed models, while the second method combines adaptive switching and continuous adaptation. The third method is based on novel ways of locating models, while the fourth method confined search of unknown parameters to convex regions. In complex problems, the variations that can occur cannot be classified easily. The proposal attempts to study methods for adapting efficiently by distributing resources.