The need for verification and validation techniques for adaptive control system

Adaptive systems are systems whose function evolves over time, as they improve their performance through learning. The advantage of adaptive systems is that they can, through judicious learning, react to situations that were never individually identified and analyzed by the designer. If learning and adaptation are allowed to occur after the control system is deployed, the system is called an online adaptive system. Online adaptive systems are attracting increasing attention in application domains where autonomy is an important feature, or where it is virtually impossible to analyze ahead of time all the possible combinations of environmental conditions that may arise. An archetype of the former are long term space missions where communication delays to ground stations are prohibitively long, and we have to depend on the systems' local capabilities to deal with unforeseen circumstances. An archetype of the latter are flight control systems, which deal with a wide range of parameters, and a wide range of environmental factors. In recent years NASA conducted experiments evaluating adaptive computational paradigms (neural networks, AI planners) for providing fault tolerance capabilities in control systems following sensor and/or actuator faults. Experimental success suggests significant potential for future use. The critical factor limiting wider use of neural networks and other soft-computing paradigms in process control applications, is our (in)ability to provide a theoretically sound and practical approach to their verification and validation.