A critical analysis of genetic algorithms for global optimization
Pradeepkumar V. Annaiyappa · 1992
The general problem of global optimization is difficult. Very few special cases can be solved analytically, and exhaustive search of high dimensional spaces is intractable. There are many techniques based on probabilistic sampling that provide some hope of obtaining a feasible solution, but the general problem remains unsolved. The process of evolution in the biological system has been described as an optimization process that attempts to generate a population of organisms that best fit a given environment. Researchers are trying to apply this biological process to artificial problems. One particular model called Genetic Algorithms that was proposed by John Holland has been applied to many different problems with varying degrees of success. Research in Genetic Algorithms has been isolated and has never mixed with the main stream optimization literature. This has caused the exposition of problems and proposed solutions using this paradigm to remain focused on the specific model proposed by John Holland. However, when applied to artificial problems, this strict conformance to the original model has disclosed the brittleness of Genetic Algorithms. This dissertation discusses many of the problems of Genetic Algorithms in its current state of the art. It provides a different perspective of the string representation used by the Genetic Algorithm that reveals its fragility. Contrary to claims by some researchers in the field of Genetic Algorithms, there are several methods in the optimization literature that are very similar in their operational scheme to the Genetic Algorithm. However, most of the studies that attempt to compare the Genetic Algorithms with other optimization techniques have incorporated only gradient descent methods. This dissertation presents several methods that are similar to the Genetic Algorithm and compares them for their performance. These optimization methods that work from a population of search points can be generalized into a parameterized operational scheme. This results in a space of optimization methods which can then be analyzed for their effectiveness on different types of problems. A generic parameterized model is proposed to capture the essence of random sampling methods for Global Optimization.