Many disciplines involve optimization at the core. In physics, systems are driven to their lowest energy state subject to physical laws. In business, corporations aim to maximize shareholder value. In biology, fitter organisms are more likely to survive. This research focus on optimization from an engineering perspective, where the objective is to design a system that optimizes a set of metrics subject to constraints. The system could be a complex physical system like an aircraft, or it could be a simple structure such as a bicycle frame.

The system might not even be physical; for example, we might be interested in designing a control system for an automated vehicle or a computer vision system that detects whether an image of a tumor biopsy is cancerous. We want these systems to perform as well as possible. Depending on the application, relevant metrics might include efficiency, safety, and accuracy. Constraints on the design might include cost, weight, and structural soundness.

This research is about the algorithms, or computational processes, for optimization. Given some representation of the system design, such as a set of numbers encoding the geometry of an airfoil, these algorithms will tell us how to search the space of possible designs with the aim of finding the best one. Depending on the application, this search may involve running physical experiments, such as wind tunnel tests, or it might involve evaluating an analytical expression or running computer simulations. 

Decades of advances in large scale computation have resulted in innovative physical engineering designs as well as the design of artificially intelligent systems. The intelligence of these systems has been demonstrated in games such as chess, Jeopardy, and Go. IBM’s Deep Blue defeated the world chess champion Garry Kasparov in 1996 by optimizing moves by evaluating millions of positions. In 2011, IBM’s Watson played Jeopardy! against former winners Brad Futter and Ken Jennings. Watson won the first place prize of $1 million by optimizing its response with respect to probabilistic inferences about 200 million pages of structured and unstructured data.

Since the competition, the system has evolved to assist in healthcare decisions and weather forecasting. In 2017, Google’s AlphaGo defeated Ke Jie, the number one ranked Go player in the world. The system used neural networks with millions of parameters that were optimized from self-play and data from human games. The optimization of deep neural networks is fueling a major revolution in artificial intelligence that will likely continue.

Leave a Reply

Your email address will not be published. Required fields are marked *