Artificial Intelligence as a Control Problem: Comments on the Relationship between Machine Learning and Intelligent Control

Ultimately, the problem of Artificial Intelligence (and thus of Neural Nets) comes down to that of making a sequence of decisions over time so as to achieve certain goals. AI is thus a control problem, at least in a trivial sense, but also in a deeper sense. This view is to be contrasted with AI's traditional view of itself, in which the central paradigm is not that of control, but of problem solving in the sense of solving a puzzle, playing a board game, or solving a word problem. Areas where the problem solving paradigm does not naturally apply, such as robotics and vision, have been viewed as outside mainstream AI. I think that the control viewpoint is now much more profitable than the problem solving one, and that control should be the centerpiece of AI and machine learning research. If both AI and more traditional areas of engineering are viewed as approaches to the general problem of control, then why do they seem so different? In the 1950's and early 1960's these fields were not clearly distinguished. Pattern recognition, for example, was once a central concern of AI and only gradually shifted to become a separate specialized subfield. This happened also with various approaches to learning and adaptive control. I would characterize the split as having to do with the familiar dilemma of choosing between obtaining clear, rigorous results on the one hand, and exploring the most interesting, powerful systems one can think of on the other. AI clearly took the latter " more adventurous " approach, utilizing fully the experimental methodology made possible by digital computers, while the " more rigorous " approach became a natural extension of existing engineering theory, based on the pencil-and-paper mathematics of theorem and proof. See the figure. This is not in any way to judge these fields. The most striking thing indicated in the figure is not that some work was more rigorous and some more adventurous, but the depth of the gulf between work of these two kinds. Most AI work makes absolutely no contact with traditional engineering algorithms, and vice versa. Perhaps this was necessary for each field to establish its own identity, but now it is counterproductive. The hottest spot in both fields is the one between them. The current enormous popularity of neural networks is due at least in part to its seeming to span these two—the applications …