Thomas Philip Runarsson

Stochastic ranking for constrained evolutionary optimization

This paper analyzes a (1,

Reliability of quantitative EEG features

\lambda

Temporal Difference Learning Versus Co-Evolution for Acquiring Othello Position Evaluation

)-Evolution Strategy, a randomized comparison-based adaptive search algorithm, optimizing a linear function with a linear constraint. The algorithm uses resampling to handle the constraint. Two cases are investigated: first the case where the step-size is constant, and second the case where the step-size is adapted using cumulative step-size adaptation. We exhibit for each case a Markov chain describing the behaviour of the algorithm. Stability of the chain implies, by applying a law of large numbers, either convergence or divergence of the algorithm. Divergence is the desired behaviour. In the constant step-size case, we show stability of the Markov chain and prove the divergence of the algorithm. In the cumulative step-size adaptation case, we prove stability of the Markov chain in the simplified case where the cumulation parameter equals 1, and discuss steps to obtain similar results for the full (default) algorithm where the cumulation parameter is smaller than 1. The stability of the Markov chain allows us to deduce geometric divergence or convergence , depending on the dimension, constraint angle, population size and damping parameter, at a rate that we estimate. Our results complement previous studies where stability was assumed.Penalty functions are often used in constrained optimization. However, it is very difficult to strike the right balance between objective and penalty functions. This paper introduces a novel approach to balance objective and penalty functions stochastically, i.e., stochastic ranking, and presents a new view on penalty function methods in terms of the dominance of penalty and objective functions. Some of the pitfalls of naive penalty methods are discussed in these terms. The new ranking method is tested using a (/spl mu/, /spl lambda/) evolution strategy on 13 benchmark problems. Our results show that suitable ranking alone (i.e., selection), without the introduction of complicated and specialized variation operators, is capable of improving the search performance significantly.

Constrained Evolutionary Optimization by Approximate Ranking and Surrogate Models

OBJECTIVE To investigate the reliability of several well-known quantitative EEG (qEEG) features in the elderly in the resting, eyes closed condition and study the effects of epoch length and channel derivations on reliability. METHODS Fifteen healthy adults, over 50 years of age, underwent 10 EEG recordings over a 2-month period. Various qEEG features derived from power spectral, coherence, entropy and complexity analysis of the EEG were computed. Reliability was quantified using an intraclass correlation coefficient. RESULTS The highest reliability was obtained with the average montage, reliability increased with epoch length up to 40s, longer epochs gave only marginal improvement. The reliability of the qEEG features was highest for power spectral parameters, followed by regularity measures based on entropy and complexity, coherence being least reliable. CONCLUSIONS Montage and epoch length had considerable effects on reliability. Several apparently unrelated regularity measures had similar stability. Reliability of coherence measures was strongly dependent on channel location and frequency bands. SIGNIFICANCE The reliability of regularity measures has until now received limited attention. Low reliability of coherence measures in general may limit their usefulness in the clinical setting.

Support vector machines and dynamic time warping for time series

This paper compares the use of temporal difference learning (TDL) versus co-evolutionary learning (CEL) for acquiring position evaluation functions for the game of Othello. The paper provides important insights into the strengths and weaknesses of each approach. The main findings are that for Othello, TDL learns much faster than CEL, but that properly tuned CEL can learn better playing strategies. For CEL, it is essential to use parent-child weighted averaging in order to achieve good performance. Using this method a high quality weighted piece counter was evolved, and was shown to significantly outperform a set of standard heuristic weights

Automatic Sleep Staging using Support Vector Machines with Posterior Probability Estimates

The paper describes an evolutionary algorithm for the general nonlinear programming problem using a surrogate model. Surrogate models are used in optimization when model evaluation is expensive. Two surrogate models are implemented, one for the objective function and another for a penalty function based on the constraint violations. The proposed method uses a sequential technique for updating these models. The quality of the surrogate models is determined by their consistency in ranking the population rather than their statistical accuracy. The technique is evaluated on a number of standard test problems.

Coevolving Game-Playing Agents: Measuring Performance and Intransitivities

Effective use of support vector machines (SVMs) in classification necessitates the appropriate choice of a kernel. Designing problem specific kernels involves the definition of a similarity measure, with the condition that kernels are positive semi-definite (PSD). An alternative approach which places no such restrictions on the similarity measure is to construct a set of inputs and let each example be represented by its similarity to all the examples in this set and then apply a conventional SVM to this transformed data. Dynamic time warping (DTW) is a well established distance measure for time series but has been of limited use in SVMs since it is not obvious how it can be used to derive a PSD kernel. The feasibility of the similarity based approach for DTW is investigated by applying the method to a large set of time-series classification problems.

Supervised learning linear priority dispatch rules for job-shop scheduling

This paper describes attempts at constructing an automatic sleep stage classifier using EEG recordings. Three different feature extraction schemes were compared together with two different pattern classifiers, the recently introduced support vector machine and the well known k-nearest neighbor classifier. Using estimates of posterior probabilities for each of the sleep stages it was possible to devise a simple post-processing rule which leads to improved accuracy. Compared to a human expert the accuracy of the best classifier is 81%

Ordinal regression in evolutionary computation

Coevolution is a natural choice for learning in problem domains where one agents behavior is directly related to the behavior of other agents. However, there is a known tendency for coevolution to produce mediocre solutions. One of the main reasons for this is cycling, caused by intransitivities among a set of players. In this paper, we explore the link between coevolution and games, and revisit some of the coevolutionary literature in a games and measurement context. We propose a set of measurements to identify cycling in a population and a new algorithm that tries to minimize cycling in strictly competitive (zero sum) games. We experimentally verify our approach by evolving weighted piece counter value functions to play othello, a classic two-player perfect information board game. Our method is able to find extremely strong value functions of this type.

Approximate Evolution Strategy using Stochastic Ranking

This paper introduces a framework in which dispatching rules for job-shop scheduling problems are discovered by analysing the characteristics of optimal solutions. Training data is created via randomly generated job-shop problem instances and their corresponding optimal solution. Linear classification is applied in order to identify good choices from worse ones, at each dispatching time step, in a supervised learning fashion. The method is purely data-driven, thus less problem specific insights are needed from the human heuristic algorithm designer. Experimental studies show that the learned linear priority dispatching rules outperforms common single priority dispatching rules, with respect to minimum makespan.