Multi-objective optimization by Machine Learning

Abstract

Solving multi-objective optimization with machine learning can significantly improve various fields, such as multi-junction traffic management or stock portfolio optimization. These are problems that can have a large amount of relevant and irrelevant data. This thesis targets one such problem area, focusing on multi-objective optimization in trot horse harness racing, specifically the V75.

A large part of the project was data-related, such as data collection, preprocessing, and engineering. The predicting part is divided into two parts single race prediction and system predictions. The single-race prediction utilizes the large amount of data collected to train a neural network to predict the percentage of the horse finishing behind the winner. The system prediction uses the result from the neural network to pick a system. During this process, a greedy algorithm selects more horses in the races that the machine learning deems close and fewer that it deems one-sided.

The performance evaluation showed that the single race predicting performed on par with the more advanced baseline and showed clear signs of finding a pattern between the data and the finishing result. The system prediction found some accuracy but did not surpass the odds baseline.