Discrete Probabilistic Simulated Annealing with Local Search (DPSAwLS)
Madara M. Ogot, Sherif Aly · 10th AIAA/ISSMO Multidisciplinary Analysis and Optimization Conference · 2004
Early conceptual design within a multi-disciplinary framework often involve large, noisy, multimodal design spaces that tend to be global in nature. This paper presents a stochastic robust design optimization method based on simulated annealing (SA). Discrete Probabilistic Simulated Annealing with Local Search (DPSAwLS), works eectively in these types of design spaces and has proven to be eective for global optimization problems encountered during the early stages of the design process. The nature of the design space makes the use of gradient-based methods unsuitable as they tend to get caught in local minima. Gradient-based methods may be applicable later in the design process once the design space has been reduced and the design needs to be refined. DPSAwLS, (1) results in faster and more reliable convergence to the global minimum than regular simulated annealing, (2) accounts for uncertainty in design and environmental variables, (3) can handle problems with continuous or mixed (continuous and discrete) variables, and (4) is simple to code and can therefore be readily wrapped around any existing analysis routines that are treated as ‘black boxes’. With a drastic reduction in the number of iterations required for convergence, DPSAwLS becomes a practical optimization method for inexpensive (seconds per analysis iteration) and moderately expensive (minutes per analysis iteration) global optimization problems. Two aerodynamic shape optimization problems are presented to demonstrate the ecacy of the approach.