Soraya B. Rana

Evaluating evolutionary algorithms

Test functions are commonly used to evaluate the effectiveness of different search algorithms. However, the results of evaluation are as dependent on the test problems as they are on the algorithms that are the subject of comparison. Unfortunately, developing a test suite for evaluating competing search algorithms is difficult without clearly defined evaluation goals. In this paper we discuss some basic principles that can be used to develop test suites and we examine the role of test suites as they have been used to evaluate evolutionary search algorithms. Current test suites include functions that are easily solved by simple search methods such as greedy hill-climbers. Some test functions also have undesirable characteristics that are exaggerated as the dimensionality of the search space is increased. New methods are examined for constructing functions with different degrees of nonlinearity, where the interactions and the cost of evaluation scale with respect to the dimensionality of the search space.

Island Model genetic Algorithms and Linearly Separable Problems

Parallel Genetic Algorithms have often been reported to yield better performance than Genetic Algorithms which use a single large panmictic population. In the case of the Island Model Genetic Algorithm, it has been informally argued that having multiple subpopulations helps to preserve genetic diversity, since each island can potentially follow a different search trajectory through the search space. It is also possible that since linearly separable problems are often used to test Genetic Algorithms, that Island models may simply be particularly well suited to exploiting the separable nature of the test problems. We explore this possibility by using the infinite population models of simple genetic algorithms to study how Island Models can track multiple search trajectories. We also introduce a simple model for better understanding when Island Model Genetic Algorithms may have an advantage when processing linearly separable problems.

Neural network classification of remote-sensing data

Abstract Neural nets offer the potential to classify data based upon a rapid match to overall patterns using previously calculated weighting factors, rather than point-by-point comparisons involving algorithmic logic applied to individual data values. Analytical tasks thus are greatly reduced. This paper describes an example of the use of artificial neural networks to classify remotely sensed data, determining that the networks can provide a useful level of categorization. The addition of texture data improves general discrimination ability of the network but diminishes its ability to distinguish between specific types of vegetation. Procedures for optimizing net design were successfully identified. This study validates use of neural networks by application to a larger data set than has been employed previously, and extends previous findings in several ways: 1. (1) It documents a method for designing and training networks which may be used to achieve within-class discrimination for a given data set at a level comparable to human classification. 2. (2) It incorporates texture analysis of the input data. At the expense of extra computation, this permits analysis of the spatial relationships among pixels, instead of being limited to considering the pixels individually. 3. (3) It provides a working prototype system which may be used for generalized standard classification of other land image data sets created by multispectral scanner.

Searching in the Presence of Noise

In this paper, we examine the effects of noise on both local search and genetic search. Understanding the potential effects of noise on a search space may explain why some search techniques fail and why others succeed in the presence of noise. We discuss two effects that are the result of adding noise to a search space: the annealing of peaks in the search space and the introduction of false local optima.

Genetic Algorithm Behavior in the MAXSAT Domain

Random Boolean Satisfiability function generators have recently been proposed as tools for studying genetic algorithm behavior. Yet MAXSAT problems exhibit extremely limited epistasis. Furthermore, all nonzero Walsh coefficients can be computed exactly for MAXSAT problems in polynomial time using only the clause information. This means the low order schema averages can be computed quickly and exactly for very large MAXSAT problems. But unless P=NP, this low order information cannot reliably lead to the global optimum, thus nontrivial MAXSAT problems must be deceptive.

Search, Binary Representations and Counting Optima

Choosing a good representation is a vital component of solving any search problem. However, choosing a good representation for a problem is as difficult as choosing a good search algorithm for a problem. Wolpert and Мacready’s No Free Lunch theorem proves that no search algorithm is better than any other over all possible discrete functions. We elaborate on the No Free Lunch theorem by proving that there tend to be a small set of points that occur as local optima under almost all representations. Along with the analytical results, we provide some empirical evaluation of two representations commonly used in genetic algorithms: Binary Reflected Gray coding and standard Binary encoding.