This paper reports briefly on the development of a new approach to evolutionary computation, called the Learnable Evolution Model or LEM. In contrast to conventional Darwinian-type evolutionary algorithms that employ mutation and/or recombination, LEM employs machine learning to generate new populations. At each step of evolution, LEM determines hypotheses explaining why certain individuals in the population are superior to others in performing the designated class of tasks. These hypotheses are then instantiated to create a next generation. In the testing studies described here, we compared a program implementing LEM with selected evolutionary computation algorithms on a range optimization problems and a filter design problem. In these studies, LEM significantly outperformed the evolutionary computation algorithms, sometimes speeding up the evolution by two or more orders of magnitude in the number of evolutionary steps (births). LEM was also applied to a real-world problem of designing optimized heat exchangers. The resulting designs matched or outperformed the best human designs. 1

The recently introduced Learnable Evolution Model (LEM) represents a form of non-Darwinian evolutionary computation that is guided by a learning system. Specifically, LEM "genetically engineers" new populations via hypothesis formation and instantiation. Initial experiments with a preliminary implementation of LEM were highly encouraging, but tentative. This paper presents results from a new study in which LEM was systematically tested on a range of optimization problems and a complex real world design task. The study involved LEM2, a new implementation oriented toward function optimization, and ISHED, an implementation oriented toward engineering design. In all cases of function optimization, LEM2 strongly outperformed tested evolutionary algorithms in terms of the evolution length, measured by the number of fitness function evaluations needed to reach the solution. This evolutionary speedup also translated to an execution speedup whenever the fitness evaluation time was above a small threshold (a fraction of a second). The most important result of the study was that the advantage of LEM2 over the tested Darwinian-style evolutionary methods grew rapidly with the growth of the complexity of the optimized function. Experiments with ISHED on problems of optimizing heat exchangers (evaporators) produced designs that matched or exceeded designs produced by human experts. The obtained very strong results from the LEM application to two diverse domains suggest that it may be useful also in other application domains, especially, those in which the fitness function evaluation is time-consuming or complex.

A new method for optimizing complex engineering designs is presented that is based on the Learnable Evolution Model (LEM), a recently developed form of non-Darwinian evolutionary computation. Unlike conventional Darwinian-type methods that execute an unguided evolutionary process, the proposed method, called LEMd, guides the evolutionary design process using a combination of two methods, one involving computational intelligence and the other involving encoded expert knowledge. Specifically, LEMd integrates two modes of operation, Learning Mode and Probing Mode. Learning Mode applies a machine learning program to create new designs through hypothesis generation and instantiation, while Probing Mode creates them by applying expertsuggested design modification operators tailored to the specific design problem. The LEMd method has been used to implement two initial systems, ISHED1 and ISCOD1, specialized for the optimization of evaporators and condensers in cooling systems, respectively. The designs produced by these systems matched or exceeded in performance the best designs developed by human experts. These promising results and the generality of the presented method suggest that LEMd offers a powerful new tool for optimizing complex engineering systems.