The research focuses on the development of new methodologies for robust controller analysis and design, which will be combined with reinforcement learning techniques to develop a new control paradigm: robust learning control. These new analysis and design tools will then be used to address two specific application areas for intelligent building systems: structural control and environmental control. These problems are highly multidisciplinary in nature, and present interesting and important research challenges. At the same time simplified versions of these problems will be used as effective educational tools in a multidisciplinary undergraduate teaching laboratory.
The theoretical and computational part of the work will aim towards developing computationally efficient analysis and synthesis methods for a general class of robust performance problems for complex multivariable uncertain systems. These will allow one to address problems with parametric uncertainty and (possibly nonlinear) dynamic uncertainty, with both unknown disturbances and known fixed inputs. These theoretical results will be used as the basis for studying reinforcement learning controllers, by developing an uncertainty model for the learning process within the above robustness framework. This will in turn be used to develop a new controller design methodology for robust leaning controllers, which combine the best aspects of robust and reinforcement learning control. The controller will have guaranteed insensitivity to plant/parameter variations and disturbance signals, while at the same time it will be capable of precisely tuning itself to the nonlinearities and time-variations of a particular plant.
The first application area for these new techniques will be vibration supression in tall buildings. The buildings will be subjected to loading which might arise from earthquakes and/or high winds. The goal is to equip the building with sensors and actuators under computer control, creating an intelligent building which has the ability to sense and react to its enviroment. Computer simulations, based on mathematical models, will be comtined with wind-tunnel experiments on a dynamically-scaled physical model. A DSP-based real-time digital feedback control scheme will be used to implement advanced feedback controllers, operating at a sufficiently high bandwidth to effect control of wind-induced vibration on the structure.
These techniques will also be applied to design controllers for building environmental systems. These Heating, Ventilation, and Air-Conditioning (HVAC) systems present very challenging control problems because they are complex nonlinear time-varying systems, and yet the controller is required to function on first powering up, preferably without human intervention. Furthermore, high performance is required for energy efficiency, while at the same time robust stability is essential for safety reasons. The new robust learning controllers will be tested both in simulation and on an experimental HVAC system. ***
