Álinson Santos Xavier - MIPLearn: An Extensible Framework for Learning-Enhanced Optimization

Part of Discrete Optimization Talks: talks.discrete...
Álinson Santos Xavier - Argonne National Laboratory
Speaker webpage: axavier.org/
MIPLearn: An Extensible Framework for Learning-Enhanced Optimization
Abstract: In many practical scenarios, discrete optimization problems are solved repeatedly, often on a daily basis or even more frequently, with only slight variations in input data. Examples include the Unit Commitment Problem, solved multiple times daily for energy production scheduling, and the Vehicle Routing Problem, solved daily to construct optimal routes. In this talk, we introduce MIPLearn, an extensible open-source framework which uses machine learning (ML) to enhance the performance of state-of-the-art MIP solvers in these situations. Based on collected statistical data, MIPLearn predicts good initial feasible solution, redundant constraints in the formulation, and other information that may help the solver to process new instances faster. The framework is compatible with multiple MIP solvers (e.g. Gurobi, CPLEX, SCIP, HiGHS), multiple modeling languages (JuMP, Pyomo, gurobipy) and supports user-provided ML models. Computational experiments are presented for discrete optimization problems from different domains.
Bio: Alinson Santos Xavier is a Computational Scientist at Argonne National Laboratory’s Energy Systems and Infrastructure Analysis division. His research focuses on solving challenging computational problems that arise daily in the production and transmission of electric power, through a combination of Mathematical Optimization, High-Performance Computing and Machine Learning. Dr. Xavier holds a PhD. in Mathematics (Combinatorics & Optimization) from University of Waterloo, Canada, and a MSc. in Computer Science from Universidade Federal do Ceara, Brazil.