The primary objectives of the proposed research are: 1) Develop a new controller design methodology based around a combination of robust control and neural network reinforcement learning techniques. The strategy will combine the best aspects of robust control and reinforcement learning and yield a combined control design that is far more powerful than either acting alone. 2) Provide a theoretical framework for precisely analyzing the stability and performance properties of reinforcement learning controllers by exploiting and extending results from robust control theory. 3) Apply these results to the design of closed loop controllers by exploiting and extending results from robust control theory. 4) Implement and test these ideas on an experimental HVAC system. 5) Educate practitioners in industry about this new technology. 6) Integrate these ideas into systems and control curriculum. To accomplish these objectives, an interdisciplinary team has been formed consisting of a specialist in robust control from the Electrical Engineering Department, a specialist in reinforcement learning for neural networks from the Department of Computer Science, and a specialist in design, modeling and control of HVAC systems from the Mechanical Engineering Department. This interdisciplinary approach will advance the state-of-the-art in the theory of robust reinforcement learning control design, demonstrate these new methods on an experimental HVAC system and provide much needed improved methods for controlling HVAC systems in buildings. Eventual wide spread implementation of these schemes in buildings around the world will reduce energy consumption and extend equipment life.
