This project will address the general task of solving optimization problems where the criterion to be optimized is a very complicated, nonconvex function of the choices available to the user. As a testbed example, it will focus on the classic Traveling Salesman Problem (TSP) for which many practical applications and benchmarks are available. General-purpose learning systems have never been able to perform as well as ad hoc, problem-specific methods here; however, by using new learning approaches, this PI has already achieved results much better than previous efforts of that type. In this project, he plans to go further, by using more sophisticated neural network designs motivated by recent work in intelligent control and reinforcement learning. Better tools for learning-based optimization will be important in the long-term as part of learning-based brain-like systems to make decisions over time and to manage complex network systems.
The project will also include collaboration with K-12 and EPSCOR institutions, and will generate capstone engineering projects for undergraduate students.
