Helmut A. Mayer

Evolutionary and coevolutionary approaches to time series prediction using generalized multi-layer perceptrons

The prediction of future values of a time series generated by a chaotic dynamical system is an extremely challenging task. Amongst several nonlinear models employed for the prediction of chaotic time series, artificial neural networks (ANNs) have gained major attention in the past decade. One widely recognized aspect of ANN design in order to achieve sufficient prediction performance is the structure of the network. We automatize this procedure by evolving ANN topologies of low complexity, guiding the evolutionary process towards ANNs of increased generalization ability. Specifically, a genetic algorithm (GA) is utilized to construct the architecture of generalized multi-layer perceptrons (GMPs) trained by error backpropagation. Another less investigated but important factor of ANN prediction quality is the size and composition of the training data set (TDS). Henceforth, we subject the selection of training data to artificial evolution in the environment of an ANN with fixed structure. A natural way to exploit the mutual dependencies of ANN structures and TDSs is symbiotic (cooperative) coevolution, where the fitness of an ANN is equally credited to the TDS it has been trained with. We compare these methods (ANN evolution, TDS evolution, and coevolution) with a standard ANN architecture given in the literature by predicting the Mackey-Glass time series.

Symbiotic Coevolution of Artificial Neural Networks and Training Data Sets

Among the most important design issues to be addressed to optimize the generalization abilities of trained artificial neural networks (ANNs) are the specific architecture and the composition of the training data set (TDS). Recent work has focused on investigating each of these prerequisites separately. However, some researchers have pointed out the interacting dependencies of ANN topology and the information contained in the TDS. In order to generate coadapted ANNs and TDSs without human intervention we investigate the use of symbiotic (cooperative) coevolution. Independent populations of ANNs and TDSs are evolved by a genetic algorithm (GA), where the fitness of an ANN is equally credited to the TDS it has been trained with. The parallel netGEN system generating generalized multi-layer perceptrons being trained by error-back-propagation has been extended to coevolve TDSs. Empirical results on a simple pattern recognition problem are presented.

Hybrid Evolutionary Approaches to CNC Drill Route Optimization

A problem in computer numeric control (CNC) manufacturing is the minimization of the distance a drilling tool has to move in order to drill holes at given locations. After introducing the problem being an instance of the traveling salesman problem (TSP) we describe the route optimizer RO3 based on an evolutionary algorithm (EA). In experiments with real-world problem instances we were able to improve the results of RO3 by hybridization with the 2-opt heuristic to route lengths being 6% above the optimum. In its hybrid version RO3 achieves machine time savings of about 10% compared to visual optimization by a human expert

Coevolution of neural Go players in a cultural environment

We present experiments (co)evolving Go players based on artificial neural networks (ANNs) for a 5/spl times/5 board. ANN structure and weights are encoded in multi-chromosomal genotypes. In evolutionary scenarios, a population of generalized multi-layer perceptrons (GMLPs) has to compete with a single Go program from a set of three players of different quality. Two coevolutionary approaches, namely, a dynamically growing culture, and a fixed-size elite represent the changing environment of the coevolving population. The playing quality of the (co)evolved players is measured by a strength value derived from games against the set of three programs. We also report on first experiments employing recurrent networks, which allow a direct structural representation of the Go board. Finally, the quality of all the best (co)evolved players is evaluated in a round robin tournament.

Board Representations for Neural Go Players Learning by Temporal Difference

The majority of work on artificial neural networks (ANNs) playing the game of Go focus on network architectures and training regimes to improve the quality of the neural player. A less investigated problem is the board representation conveying the information on the current state of the game to the network. Common approaches suggest a straight-forward encoding by assigning each point on the board to a single (or more) input neurons. However, these basic representations do not capture elementary structural relationships between stones (and points) being essential to the game. We compare three different board representations for self-learning ANNs on a 5 times 5 board employing temporal difference learning (TDL) with two types of move selection (during training). The strength of the trained networks is evaluated in games against three computer players of different quality. A tournament of the best neural players, addition of alpha-beta search, and a commented game of a neural player against the best computer player further explore the potential of the neural players and its respective board representations

Multi-chromosomal representations and chromosome shuffling in evolutionary algorithms

We present experiments investigating the use of multichromosomal representations in evolutionary algorithms. Specifically, the conventional representation of parameters on a single chromosome is compared to a genotype encoding with multiple chromosomes on a set of test functions. In this context we present chromosome shuffling, a genetic operator recombining complete chromosomes based on biological evidence. The hypothesis that the multichromosomal representation ameliorates the transmission of good subsolutions to the population is tested on functions of varying degree of complexity.

Ptgas---genetic algorithms evolving noncoding segments by means of promoter/terminator sequences

In this article we present work on chromosome structures for genetic algorithms (GAs) based on biological principles. Mainly, the influence of noncoding segments on GA behavior and performance is investigated. We compare representations with noncoding sequences at predefined, fixed locations with junk code induced by the use of promoter/terminator sequences (ptGAs) that define start and end of a coding sequence, respectively. AS one of the advantages of noncoding segments a few researchers have identified the reduction of the disruptive effects of crossover, and we solidify this argument by a formal analysis of crossover disruption probabilities for noncoding segments at fixed locations. The additional use of promoter/terminator sequences not only enables evolution of parameter values, but also allows for adaptation of number, size, and location of genes (problem parameters) on an artificial chromosome. Randomly generated chromosomes of fixed length carry different numbers of promoter/terminator sequences resulting in genes of varying size and location. Evolution of these ptGA chromosomes drives the number of parameters and their values to (sub)optimal solutions. Moreover, the formation of tightly linked building blocks is enhanced by self-organization of gene locations. We also introduce a new, nondisruptive crossover operator emerging from the ptGA gene structure with adaptive crossover rate, location, and number of crossover sites. For experimental comparisons of this genetic operator to conventional crossover in GAs, as well as properties of different ptGA chromosome structures, an artificial problem from the literature is utilized. Finally, the potential of ptGA is demonstrated on an NP-complete combinatorial optimization problem.

Evolution of iterated prisoner's dilemma strategies with different history lengths in static and cultural environments

We investigate evolutionary approaches to generate well-performing strategies for the iterated prisoners dilemma (IPD) with different history lengths in static and cultural environments. The length of the history determines the number of the most recent moves of both players taken into account for the current move decision. The static environment constituting the opponents of the evolved players is made up of ten standard strategies known from the literature. The cultural environment starts with the standard strategies and gradually increases by addition of the best evolved players representing a culture. The performance of the various evolved strategies is compared in specific tournaments. Also, the behavior of an evolved player is analyzed in more detail by looking at the specific game sequences (and corresponding decisions), which out of all possible sequences are actually utilized in a tournament.

Improving Statistical Measures of Feature Subsets by Conventional and Evolutionary Approaches

In this paper we compare recently developed and highly effective sequential feature selection algorithms with approaches based on evolutionary algorithms enabling parallel feature subset selection. We introduce the oscillating search method, employ permutation encoding offering some advantages over the more traditional bitmap encoding for the evolutionary search, and compare these algorithms to the often studied and well-performing sequential forward floating search. For the empirical analysis of these algorithms we utilize three well-known benchmark problems, and assess the quality of feature subsets by means of the statistical Bhattacharyya distance measure.

Differentiation of neuron types by evolving activation function templates for artificial neural networks

In this paper we investigate the use of neuron-specific activation functions (AFs) within generalized multilayer perceptrons (GMLP). We utilize the netGEN system not only to evolve the structure of an artificial neural network (ANN), but also to search for a set of AF templates which are assigned to specific neurons by evolution. This may be seen as a loose analogy to neuron differentiation in biological neural networks (BNNs). While BNNs employ different neuron types in functionally different brain areas, neuron differentiation in ANNs might be useful to increase the adaptability to specific problems. The evolution of AF templates is based on evolving the control points of a cubic spline function, hence nonmonotonous AFs of (nearly) arbitrary shape may be generated. We present a number of experiments evolving ANN structure and AF templates using the parallel netGEN system to train the evolved architectures. We compare the evolved cubic spline ANNs with evolved sigmoid ANNs on synthetic classification problems and a time series prediction task so as to assess the benefits of problem-adapted AF templates.