Abstract: In this paper we review models, methods and ideas all intended to improve performance of mixed integer programming (MIP) solvers, such as Gurobi. Aside from task instances as an input, MIP solvers also receive a set of tunable parameters that affect the performance of a solver. Successful parameter configuration may lead to the solver’s performance increase, but such configurations usually work fine only for specific task family, so for different tasks new configuration must be found. The number of parameters in such configurations varies with each different solver, but it may vary from 20 to 100, which makes the solver’s parameter configuration difficult and time-consuming, as you need to test a specific task set on each configuration. That’s why, in this paper we discuss different ways to automatize configuration process, and also suggest a way to improve its efficiency. Base tasks we use for benchmark are instances of multicore processor scheduling problem. For automated parameter configuration we use variable neighborhood search (VNS) [1]. To optimize configuration process we suggest to use large language models (LLM) to obtain same-size embeddings (vector representations of tasks), even for instances with different dimensionalities, so we could use them to train linear regression model to predict parameters of solver for each particular instance without running VNS on them.

Cite: Ustyugov V.
On Different Methods For Automated MILP Solver Configuration
20-я Международная Азиатская школа-семинар «Проблемы оптимизации сложных систем» 19-30 Jul 2024