Extracting and Applying Knowledge with Adaptive Knowledge-Driven Optimization to Architect an Earth Observing Satellite System Conference Paper uri icon

abstract

  • © 2017, American Institute of Aeronautics and Astronautics Inc, AIAA. All rights reserved. Knowledge-driven optimization (KDO) applies machine learning methods during the optimization to extract design principles that define relationships between design variables that are common in high-quality designs. In an attempt to accelerate the convergence of the optimization, the extracted knowledge is applied during the remainder of the search by encoding it as constraints or knowledge-dependent operators to modify solutions. Applying the heuristics as constraints, however, prevents sufficient tradespace exploration, which can potentially lead to convergence on local optimal. In addition, current approaches can only apply one knowledge-dependent operator at a time during the optimization process, which is limiting because some of the learned design principles will be more helpful than others in finding improving solutions. This work presents a new framework KDOAOS that can generate and apply multiple knowledge-dependent operators during the optimization through the use of adaptive operator selection (AOS). KDOAOS also uses knowledge-independent operators such as crossover and mutation alongside knowledge-dependent operators and continually reallocates computational resources to the effective operators to maintain an efficient search strategy and escape local optima. The efficacy of KDOAOS is demonstrated on a design problem where the goal is to architect an Earth observing satellite system that maximizes scientific benefit and minimizes lifecycle cost. The results show that KDOAOS statistically outperforms an analogous knowledge-independent algorithm and an algorithm that applies the extracted knowledge through constraints.

author list (cited authors)

  • Hitomi, N., Bang, H., & Selva, D.

publication date

  • January 1, 2017 11:11 AM