TY - GEN
T1 - Control co-design of a slender structure in high-speed flow via fast multiobjective Bayesian optimization
AU - Carlson, Kade
AU - Renganathan, Ashwin
AU - Huang, Daning
N1 - Publisher Copyright:
© 2024, American Institute of Aeronautics and Astronautics Inc, AIAA. All rights reserved.
PY - 2024
Y1 - 2024
N2 - The control co-design of a slender structure in high-speed (supersonic to hypersonic) flows involves the simultaneous optimization of multiple objectives, namely, structural mass and control effort. To make this optimization sample-efficient, we consider multiobjective Bayesian optimization (MOBO) which places a Gaussian process prior on the objectives followed by sequentially updating the posterior GP in a goal-oriented fashion via Bayesian decision theory. Conventional MOBO, however, involves an "inner" optimization of an intractable "acquisition" function, which can in some cases be a non-trivial computational overhead, in addition to being hard to solve. Inability to solve this inner optimization accurately causes MOBO to be not sample efficient. In this work, we solve the control co-design problem, via a novel, "acquisition-free", MOBO approach called Pareto optimal Thompson sampling (qPOTS). qPOTS is revolutionary in the sense that new candidate(s) are chosen from the Pareto frontier of random GP posterior sample paths obtained by solving a much cheaper multiobjective optimization problem, instead of solving a potentially nonconvex, nonsmooth, and/or stochastic inner optimization. We demonstrate the superior performance of qPOTS compared to the state-of-the-art on several synthetic test functions as well as the control co-design problem.
AB - The control co-design of a slender structure in high-speed (supersonic to hypersonic) flows involves the simultaneous optimization of multiple objectives, namely, structural mass and control effort. To make this optimization sample-efficient, we consider multiobjective Bayesian optimization (MOBO) which places a Gaussian process prior on the objectives followed by sequentially updating the posterior GP in a goal-oriented fashion via Bayesian decision theory. Conventional MOBO, however, involves an "inner" optimization of an intractable "acquisition" function, which can in some cases be a non-trivial computational overhead, in addition to being hard to solve. Inability to solve this inner optimization accurately causes MOBO to be not sample efficient. In this work, we solve the control co-design problem, via a novel, "acquisition-free", MOBO approach called Pareto optimal Thompson sampling (qPOTS). qPOTS is revolutionary in the sense that new candidate(s) are chosen from the Pareto frontier of random GP posterior sample paths obtained by solving a much cheaper multiobjective optimization problem, instead of solving a potentially nonconvex, nonsmooth, and/or stochastic inner optimization. We demonstrate the superior performance of qPOTS compared to the state-of-the-art on several synthetic test functions as well as the control co-design problem.
UR - http://www.scopus.com/inward/record.url?scp=85203454230&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85203454230&partnerID=8YFLogxK
U2 - 10.2514/6.2024-4026
DO - 10.2514/6.2024-4026
M3 - Conference contribution
AN - SCOPUS:85203454230
SN - 9781624107160
T3 - AIAA Aviation Forum and ASCEND, 2024
BT - AIAA Aviation Forum and ASCEND, 2024
PB - American Institute of Aeronautics and Astronautics Inc, AIAA
T2 - AIAA Aviation Forum and ASCEND, 2024
Y2 - 29 July 2024 through 2 August 2024
ER -