Abstract:
Parallel processing is leveraged to tremendous effect in certain areas of optimization; for instance, this is key to enabling the training of large language models. Parallel computing, however, has had relatively little practical impact on important problems such as mixed-integer optimization. Hence for many problems there is a gap between what hardware offers and what optimization software needs. This talk presents two recent projects that address some aspects of this gap. The first project considers the problem of projecting a vector onto a (possibly weighted) simplex; such projection is a central subroutine for various more complex problems. The second involves conflict cut generation and management for (generic) mixed-integer linear programming problems. Both projects share the algorithmic theme of distributed preprocessing, where parallel compute is applied beforehand in order to accelerate a serial main method is run. We will consider both theoretical and practical aspects of this approach in contrast to other possibilities.
Short Bio:
Chen Chen is an Assistant Professor in the Integrated Systems Engineering department at The Ohio State University, as well as a core faculty member of its Sustainability Institute. He received a PhD in Industrial Engineering and Operations Research from UC Berkeley in 2015, and was thereafter a postdoctoral researcher in the Industrial Engineering & Operations Research department at Columbia University until 2017. He is interested in various aspects of global optimization, including: mixed-integer conic optimization; nonconvex continuous problems such as polynomial or signomial optimization; and hardware acceleration. Such work is motivated by a variety of applications, especially pertaining to machine learning and power systems.
