Automated Machine Learning for Business Decision Simulation

Abstract

In many industries it is common to support business decisions with the help of simulation models. Such models can be used to predict the outcome of a series of decisions. Often, the number of possible options is high, making it difficult to identify good candidates for simulation. Optimization models can help to find such candidates. Building optimization models can be an arduous task for a variety of reasons, such as computational complexity, algorithmic complexity, or required expertise. Many times, simulation models are readily available. We propose to use black-box optimization techniques on simulation models to identify good business decisions. This allows generic optimization strategies to be applied to simulation models without additional effort or knowledge. There are numerous existing research papers on black box optimization. A traditional application of such routines is parameter optimization for algorithm tuning, e.g. solvers for mixed-integer linear programming. In recent years, a major focus has shifted to automated machine learning. The challenge here is to find good hyperparameters for machine learning approaches or even to choose the best strategy automatically.
We sketch different approaches to automated machine learning and how to apply them on simulation models. To demonstrate the benefit of our approach, we create a use-case for energy optimization in a commercial facility, such as an office building. Finally, the performance of the dierent approaches is evaluated in an experimental study.