A data-driven heuristic decision strategy for data-scarce optimization

Abstract

Algorithmic optimization is a viable tool for solving complex materials engineering issues. In this study, a data-scarce Bayesian optimization model was developed to research the composition of bio-based composites. The proof-of-concept program adjusts the natural materials' weight ratios to optimize towards user-defined mechanical properties. Preliminary results show that the bio-composites proposed by the program had improved properties compared to existing bulk-moulding compounds. However, the algorithm choice is often arbitrary or based on anecdotal evidence. In parallel, this thesis proposed a data-driven framework for general data-scarce optimization problems to adapt the meta-heuristic during optimization. Guided by the 'No Free Lunch' theorem, we verified that the effectiveness over a selection of algorithms is dependent on problem-specific features and convergence. This effectiveness was captured in a unique identifier metric by optimizing a generated training set of optimization problems. The average solution quality was improved by combining several meta-heuristics in series, based on these problem-specifics. During the optimization of problems in the testing set, the same unique identifier was constructed at predefined stages in the optimization process. Subsequently, the problem was classified, and the meta-heuristic was adapted to the best-performing algorithm based on similar training samples. Experiments with various classifiers and a different number of predefined assessment stages were performed. Results show that the data-driven heuristic decision strategy outperformed the individual optimizers on the testing set. Despite the use of binarization techniques, the classification accuracy was heavily influenced by the imbalanced training set. In terms of computational resources, the various adaptions of the data-driven heuristic strategy are 2.5 times faster in runtime compared to the best-performing meta-heuristic Bayesian Optimization. Lastly, the framework was benchmarked against the 'learning to optimize' study and shows excellent performance on the logistic regression problem compared to the autonomous optimizer. In conclusion, it has been shown that even with the limited information of black-box optimization problems, data-driven optimization effectively improves the current standard of materials engineering processes.