Machine learning-based algorithm selection for irregular three-dimensional packing in additive manufacturing

In additive manufacturing (AM), a common problem is the efficient arrangement of arbitrary three-dimensional objects, subject to geometric constraints. This can be mapped to three-dimensional irregular packing (3DIP) problems, which have been systematically addressed by many academics and practitioners. This study demonstrates the utilisation of machine learning for algorithm selection to find efficient layout configurations during the design stage of the AM process. The choice of the most suitable approach to use, typically a packing algorithm, is not trivial, and depends on the non-obvious relationship between the characteristics of the instance in hand and the portfolio of algorithms available. The matching between problem features and algorithm performance forms the basis of the well-known algorithm selection problem. This study introduces the first empirical investigation of algorithm selection for 3DIP problems, conducting extensive experiments with hundreds of combinations of well-known supervised machine learning classifiers and different parameter settings to identify an initial state-of-the-art for this problem. We generate a comprehensive dataset, labelled with the performance of two of the most popular 3DIP algorithms, and analyse the features which can be used to support decision making when selecting a method to solve a 3DIP instance. Our results show that deploying machine learning-based algorithm selection methods are able to outperform the results obtained by the individual constituent packing algorithms applied independently, with the best algorithm selection method obtaining a 1.48% higher average build volume utilisation over the 2000 problem instances tested.