Constraint-based design exploration for product portfolio optimization of variant-rich configurations using symbolic and subsymbolic AI

Designing mass customizable products to meet customer-specific requirements on a large scale continues to present a challenge due to increasingly complex product portfolios. As a result, variant-rich products are designed by interdisciplinary teams and realized through modular design and reusable product architectures that are organized across multiple domains summarized in e.g. Product Line Engineering (PLE). Existing dependencies in between and within domains are modelled as rule-based constraints e.g. in families of feature models or in matrix-based notations in Design Structure Matrixes (DSMs). This results in growing complexity and incomprehensible configurations for both designers and customers, requiring sophisticated decision support systems throughout the whole process. Streamlining internal variance while maintaining external variants presents a decisive competitive advantage. Therefore, an adequate optimized variant management is indispensable to employ data-driven methods and tools for modelling the underlying dependencies in symbolic logic to foster the application of subsymbolic Artificial Intelligence (AI). The combination of deterministic reasoning and probabilistic prediction systematically guides design exploration without violating the limited solution space. This paper introduces an unsupervised machine learning approach to explainable product portfolio optimization by examining historical configuration data. To this end, patterns of frequently combined features are identified, while maintaining consistency and satisfiability with respect to the underlying constraints formulated in Boolean logic. The correlations found are proposed to support the early design process and assist the designer in architectural and modular decision through additional constraints, engineering design changes, and better guidance through the configuration process for the customer. To evaluate the approach, this method has been tested successfully on synthetic data using a generic product description. Additionally, this paper addresses the known limitations for chronological changes over time and highlights three potential strategies for incorporating the gained knowledge into the development of multi-generation products.