Decades after the potential of evolutionary inspirations in global optimization was acknowledged, there is still some confusion concerning the way the essential evolutionary operators act in basic adaptive situations. This is because the dynamics of evolving populations are quite complex and often counterintuitive. Therefore appropriate visualization programs might be really helpful in providing the required expertise. We present such a program together with a number of examples that illustrate some aspects of evolution of phenotypes in unlimited, multidimensional, real spaces under different kinds of selection.
Belief networks encode joint probability distribution functions and can be used as fitness functions in genetic algorithms. Individuals in the genetic algorithm's population then represent instantiations, or explanations, in the belief network. Computing the most probable explanations (belief revision) is thus cast as a genetic algorithm search in the joint probability distribution space. At any time, the best fit individual in the genetic algorithm population is an estimate of the most probable explanation. This paper argues that joint probability distribution functions represented by belief networks typically are multimodal and highly variable. Thus the genetic algorithm techniques known as sharing and scaling should be of help. It is shown empirically that this is indeed the case, in particular that niching combined with scaling significantly improves the quality of a genetic algorithm's estimate of the most probable explanations. A novel scaling approach, root scaling, is also introduced.
For function approximation using fuzzy if-then rules, Nomura et al. (1992) proposed a genetic algorithm-based method for adjusting the fuzzy partition of an input space. In this paper, we apply their method to pattern classification problems. We have already extended the coding method to the case where intervals and trapezoidal membership functions can be used for antecedent fuzzy sets (Ishibuchi & Murata, 1996). There are, however, two drawbacks in these methods. One is that the resolution of each axis on which the fuzzy partition is adjusted should be prespecified by a decision-maker. The other is that the number of fuzzy if-then rules exponentially increases as the number of attributes increases. To cope with these drawbacks, we propose a genetic algorithmbased fuzzy partition method that has the following advantages; We show the effectiveness of the proposed method by computer simulations on iris data with 4 attributes and wine data with 13 attributes.
The task of finding minimal elements of a partially ordered set is a generalization of the task of finding the global minimum of a real-valued function or of finding Pareto-optimal points of a multicriteria optimization problem. It is shown that evolutionary algorithms are able to converge to the set of minimal elements in finite time with probability one, provided that the search space is finite, the time-invariant variation operator is associated with a positive transition probability function and that the selection operator obeys the so-called ‘elite preservation strategy.’
This paper presents a new approach to evolutionary artificial neural networks, based on the integration of feedforward neural networks, messy genetic algorithms (GAs), and singular value decomposition (SVD). The set of competing hidden nodes with variable number of connections from the input layer represents an evolving neural network. Selection of hidden nodes is based on their estimation via SVD. The resulting singular values are used to determine significance of hidden nodes for the network's output. To represent connectivity of hidden nodes and to process the topology of connections between input and hidden layers, we employ the approach of messy GAs. This establishes a framework for processing strings of variable length which codes this topology and allows one to search for useful combinations of input variables. The proposed approach is tested using sonar data classification.
DNA sequencing methods are the subject of continued interest in molecular biology for use in a wide variety of applications. Sequencing DNA by hybridization on a “DNA chip” has been estimated to increase the rate of DNA sequencing by as much as one-million fold. In this process, the sequence of a target molecule is reconstructed by the complementary binding of a pool of random probe molecules. For each target, an appropriate probe length must be used to unambiguously determine the sequence of a given target sequence of length N. Using evolutionary programming, we have simulated the binding of probes of length four nucleotides to a series of target lengths to determine most optimal target length that can be unambiguously reconstructed. Evolutionary programming is demonstrated to be well suited to sequence reconstruction problems and could also be extended for gene expression monitoring with DNA chip technology.
Shape matching techniques are important in machine intelligence, especially in applications such as robotics. Currently, there are three major approaches to shape recognition: statistical, syntactic and neural approaches. This paper presents a fourth approach: evolutionary algorithms. A steady state memetic algorithm is shown to be successful in matching shapes even when they are partially obscured, and even in the presence of noise in the input image.
The particle swarm algorithm has been shown to optimize a wide variety of complex functions. This paper examines a simplified version of the algorithm in an effort to understand the trajectories of particles as they search for solutions. Findings address optimal parameter values, point out issues for future research, and contribute to understanding this new optimization method.
Local selection (LS) is a very simple selection scheme in evolutionary algorithms. Individual fitnesses are compared to a fixed threshold, rather than to each other, to decide who gets to reproduce. LS, coupled with fitness functions stemming from the consumption of shared environmental resources, maintains diversity in a way similar to fitness sharing; however it is generally more efficient than fitness sharing, and lends itself to parallel implementations for distributed tasks. While LS is not prone to premature convergence, it applies minimal selection pressure upon the population. LS is therefore more appropriate than other, stronger selection schemes only on certain problem classes. This papers characterizes one broad class of problems in which LS consistently out-performs tournament selection.
Two meta-evolutionary optimization strategies described in this paper accelerate the convergence of evolutionary programming algorithms while still retaining much of their ability to deal with multi-modal problems. The strategies, called directional mutation and recorded step in this paper, can operate independently but together they greatly enhance the ability of evolutionary programming algorithms to deal with fitness landscapes characterized by long narrow valleys. The directional mutation aspect of this combined method uses correlated meta-mutation but does not introduce a full covariance matrix. These new methods are thus much more economical in terms of storage for problems with high dimensionality. Additionally, directional mutation is rotationally invariant which is a substantial advantage over self-adaptive methods which use a single variance per coordinate for problems where the natural orientation of the problem is not oriented along the axes.
This paper discusses the question how ES/EP-like algorithms perform the evolutionary search in real-valued N-dimensional parameter spaces. It will be shown that the sometimes invoked model of a perturbed gradient search does not seem to give an appropriate picture of the search process. Instead, the search behavior is described as the antagonism of exploitation and exploration, where exploitation works in one dimension, whereas the exploration is a random walk on a (N−1)-dimensional manifold in the search space. As an example the exploration dynamics on the sphere model will be investigated.
In this paper we introduce a novel approach to the use of automatically defined functions (ADFs) with the help of a genetic library builder (GLiB). The new technique uses the two mutation operators of GLiB to automatically create subpopulations of ADFs during evolution, where these are termed evolution-defined functions (EDFs). Our approach consists of dynamically specifying separate subpopulations for each identified ADF, where a further population of programs uses individuals from these subpopulations during evaluations. Using a multiplexer problem and two classification tasks we compare a number of existing methods with this co-evolutionary approach. It is shown that dynamically creating ADF subpopulations (according to worth) proves more beneficial than specifying them a priori. It is also shown that the approach performs better than existing approaches — GP with ADFs and GP with GLiB — at all three tasks. Further, we extend the approach to allow the number of EDFs to emerge during the course of evolution, removing the need to specify how many functions are available a priori.
Control laws for backing up a simulated truck-and-trailer to a loading dock in a parking lot are developed. Evolutionary programming (EP) with a set of mutation operators is used to automatically generate optimal control laws. EP produces an optimal nonlinear control strategy that takes the state variables as inputs and determines the direction and angle by which the truck's front wheels must be steered. Results indicate that the generated control laws successfully back up the truck-and-trailer system in an optimal manner and are capable of generalizing well over previously unseen input states.
We present a family of related test problems for genetic programming. These test problems form a very simple test environment that nevertheless possesses some degree of algorithmic subtlety. We term this genetic programming environment plus-one-recall-store (PORS). This genetic programming environment has only a pair of terminals, 1 and recall, and a pair of operations, plus and store, together with a single memory location. We present an extensive mathematical characterization of the PORS environment and report experiments testing the benefits of incorporating expert knowledge into the initial population and into the operation of crossover. The experiments indicate that, in the test environment, expert knowledge is best incorporated only in the initial population. This is a welcome result as this is the computationally inexpensive choice of the two methods of incorporating expert knowledge tested.
In this paper, genetic programming (GP) is employed to model learning and adaptation in the overlapping generations model, one of the most popular dynamic economic models. Using a model of inflation with multiple equilibria as an illustrative example, we show that our GP-based agents are able to coordinate their actions to achieve the Pareto-superior equilibrium (the low-inflation steady state) rather than the Pareto-inferior equilibrium (the high-inflation steady state). We also test the robustness of this result with different initial conditions, economic parameters, and GP control parameters.
This paper first analyzes the impact that inertia weight and maximum velocity have on the performance of the particle swarm optimizer, and then provides guidelines for selecting these two parameters. Analysis of experiments demonstrates the validity of these guidelines.
This paper compares two evolutionary computation paradigms: genetic algorithms and particle swarm optimization. The operators of each paradigm are reviewed, focusing on how each affects search behavior in the problem space. The goals of the paper are to provide additional insights into how each paradigm works, and to suggest ways in which performance might be improved by incorporating features from one paradigm into the other.
This paper reports work investigating various evolutionary approaches to vertex cover (VC), a well-known NP-Hard optimization problem. Central to each of the algorithms is a novel encoding scheme for VC and related problems that treats each chromosome as a binary decision diagram. As a result, the encoding allows only a (guaranteed optimal) subset of feasible solutions. The encoding also incorporates features of a powerful traditional heuristic for VC that allow initial evolutionary algorithm (EA) populations to be seeded in known promising regions of the search space. The resulting EAs have displayed exceptionally strong empirical performance on various vertex cover, independent set, and maximum clique problem classes.
Evolutionary algorithms (EAs) have been applied to many optimisation problems successfully in recent years. The genetic algorithm (GA) and evolutionary programming (EP) are two of the major branches of EAs. GAs use crossover as the main search operator and mutation as a background operator in search. EP typically uses mutation only. This paper investigates a novel EP algorithm for cutting stock problems. It adopts a mutation operator based on the concept of distance between a parent and its offspring. Without using crossover, the algorithm is less time consuming and more efficient in comparison with a GA-based approach. Experimental studies have been carried out to examine the effectiveness of the EP algorithm. They illustrate that EP can provide a simple yet more efficient alternative to GAs in solving some combinatorial optimisation problems.
As the software industry matures, the design and implementation of individual software components that can be assembled to form an application is gaining popularity as a method of software development. This is because component-level design shortens development time, increases maintainability, and results in a more flexible software product. The domain of evolutionary computation encompasses a wide and diversified set of algorithms. Some of these algorithms serve the same propose, but are better suited to particular domains. As evolutionary computation gains viability as a means for investigating complex problems, it would benefit from the application of component software design to build frameworks that are flexible enough to support customization, and fundamental enough to support reuse. In this paper, fundamental concepts in component software design are considered and applied to a simple evolutionary computation framework.
