Extending the boundaries of design optimization by integrating fast optimization techniques with machine-code-based, linear genetic programming

Optimized models of complex physical systems are difficult to create and time consuming to optimize. The physical and business processes are often not well understood and are therefore difficult to model. The models of often too complex to be well optimized with available computational resources. Too often approximate, less than optimal models result. This work presents an approach to this problem that blends three well-tested components. First: We apply Linear Genetic Programming (LGP) to those portions of the system that are not well understood-for example, modeling data sets, such the control settings for industrial or chemical processes, geotechnical property prediction or UXO detection. LGP builds models inductively from known data about the physical system. The LGP approach we highlight is extremely fast and builds rapid to execute, high-precision models of a wide range of physical systems. Yet it requires few parameter adjustments and is very robust against overfitting. Second: We simulate those portions of the system-for example, the cost model for the processes-these are well understood with human built models. Finally: We optimize the resulting meta-model using Evolution Strategies (ES). ES is a fast, general-purpose optimizer that requires little pre-existing domain knowledge. We have developed this approach over a several years period and present results and examples that highlight where this approach can greatly improve the development and optimization of complex physical systems. (C) 2003 Elsevier Inc. All rights reserved.