Rapid training of thermal agents with single parent genetic programming

The temperature profile across an object can be computed by iterative methods. The time spent waiting for iterative solutions to converge for multiple objects in a complex configuration is an impediment to exploratory analysis of engineering systems. A high-quality rapidly-computed initial guess can speed convergence for an iterative algorithm. A system is described and tested for creating thermal agents that supply such initial guesses. Thermal agents are specific to an object but general across different thermal boundary conditions. During an off-line training phase, genetic programming is used to locate a thermal agent by training on several sets of boundary conditions. In use, thermal agents transform boundary conditions into rapidly-converged initial values on a cellular decomposition of an object. In this study, the impact of using single parent genetic programming on thermal agents is tested. Single parent genetic programming replaces the usual sub-tree crossover in genetic programming with crossover with members of an unchanging ancestor set. The use of this ancestor set permits the incorporation of expert knowledge into the system as well as permitting the re-use of solutions derived on one object to speed training of thermal agents for another object. For three types of experiments, incorporating expert knowledge; re-using evolved solutions; and transferring knowledge between distinct configurations statistically significant improvements are obtained with single parent techniques.