Events Calendar

Learning to Decompose: A Framework For Solving Efficiently Large-Scale Optimization Problems via Decomposition

Tuesday, February 21, 2023
9:30 am - 10:30 am

Location: NHB 1.720

Process systems engineering can play a key role in addressing challenging problems such as the decarbonization of the energy sector and the chemical industry, and the design of sustainable production processes. These problems involve decisions at multiple temporal and spatial scales, leading to large-scale decision-making (optimization) problems that cannot be solved monolithically. Decomposition-based solution algorithms can solve such problems efficiently by exploiting the underlying interaction pattern (structure) between the variables and constraints of the problem. However, methods to find optimal decompositions systematically have been lacking; furthermore, the efficiency of decomposition-based solution algorithms over monolithic ones is not known a priori and their numerical implementation involves many steps that are difficult to configure.

In this seminar, I will present an automated framework, called Learning to Decompose, which combines network science, optimization theory, and machine learning to determine: (i) how to decompose an optimization problem, (ii) how to initialize the solution algorithm for the decomposed problem, and (iii) when to apply a decomposition-based solution algorithm. On the first task, Stochastic Blockmodeling is used to “learn” the underlying structure of an optimization problem based on appropriate network representations of the problem. The learned structure is subsequently linked to the appropriate decomposition-based solution method. On the second task, active learning is used to efficiently learn the optimal starting conditions (initialization) of decomposition-based solution algorithms to reduce computational time. Finally, on the third task, a graph classification approach is developed to determine a priori if a decomposition-based solution should be implemented over a monolithic one.


Ilias Mitrai is a Ph.D. Candidate in Chemical Engineering at the Department of Chemical Engineering and Materials Science at the University of Minnesota, advised by Professor Prodromos Daoutidis. He obtained a Diploma in Chemical Engineering from the Aristotle University of Thessaloniki in Greece. His research interests are in the intersection of process systems engineering, mathematical optimization, control theory, artificial intelligence, network science, and chemical engineering. During his Ph.D., Ilias worked on the efficient solution of large-scale optimization problems using decomposition-based solution algorithms and on the analysis of the reconfiguration of functional brain networks within a novel optimal control framework which rewards control performance and controller sparsity. His research has been recognized through a Doctoral Dissertation Fellowship from the Graduate School of the University of Minnesota.

Speaker: Ilias Mitrai, Univ. of Minnesota