Adaptive Knowledge-Driven Optimization for Architecting a Distributed Satellite System

Abstract

As distributed satellite systems gain interest, there is a growing need for design tools that can identify system architectures with good trades in multiple metrics. Evolutionary algorithms have shown promise as effective design-support tools but are computationally inefficient because they require evaluating many architectures before identifying ones with good tradeoffs. Knowledge-driven optimization aims to improve evolutionary algorithms by including a data-mining algorithm that learns design features common in the best architectures found during the optimization. These design features can guide the remainder of the optimization but should be used carefully to prevent misleading the search. This work presents a new knowledge-driven optimization framework, knowledge-driven optimization (KDO)/adaptive operator selection (AOS), that monitors how well the design features guide the search toward good architectures and focuses on using the best features. If none are beneficial to the optimization, KDO/AOS adapts its search strategy to explore regions of the trade space not captured by the design features. KDO/AOS is demonstrated on an architecture design problem for a distributed satellite system, and its adaptive search strategy enables it to discover high-quality architectures with over 800 fewer evaluations when compared with a conventional evolutionary algorithm. KDO/AOS is also more robust than other knowledge-driven optimization algorithms to inaccuracies from data mining.