Martin Holeňa

Catalyst Development for CO2 Hydrogenation to Fuels

New active and selective catalyst compositions for the hydrogenation of CO2 to mainly fuel‐type higher hydrocarbons were developed by application of an evolutionary strategy. It was shown that Fe and K supported on TiO2 and modified by Cu plus other modifiers resulted in highest selectivity for C5–C15 hydrocarbons at high degrees of CO2 conversion. Co containing catalysts were less suited since they produced methane and light hydrocarbons with high selectivities. A detailed study of reaction conditions showed that selected catalyst compositions were able to reach high CO2 conversion with still low selectivities to methane at higher reaction temperatures and a higher H2/CO2 ratio.

The GUHA method and its meaning for data mining

The paper presents the history and present state of the GUHA method, its theoretical foundations and its relation and meaning for data mining.

Fuzzy hypotheses testing in the framework of fuzzy logic

Abstract Testing hypotheses about the probability distribution underlying the available empirical data is one of the fundamental data-analytic tasks in any application domain. Basically, it consists in checking the null hypothesis that the probability distribution, a~priori assumed to belong to a certain set of distributions, actually belongs to some its narrow subset, which must be precisely delimited in advance. However, sometimes there are not enough clues for such a precise delimitation, especially if the purpose of the data analysis is explorative, a situation encountered increasingly often, due to the growing amount of routinely collected data and the increasing importance of data mining. That is why generalizations of statistical hypotheses testing to vague hypotheses have been investigated for more than a decade, so far based on the most straightforward approach—to replace the set defining the null hypothesis by a fuzzy set. In this paper, a principally different approach is presented, motivated by the observational logic and its success in automated knowledge discovery. Its key idea is to view statistical testing of a fuzzy hypothesis as the application of an appropriate generalized quantifier of a fuzzy predicate calculus to predicates describing the data. The theoretical principles of the approach are explained and its first implementations are briefly sketched.

Surrogate model for continuous and discrete genetic optimization based on RBF networks

Surrogate modelling has become a successful method improving the optimization of costly objective functions. It brings less accurate, but much faster means of evaluating candidate solutions. This paper describes a model based on radial basis function networks which takes into account both continuous and discrete variables. It shows the applicability of our surrogate model to the optimization of empirical objective functions for which mixing of discrete and continuous dimensions is typical. Results of testing with a genetic algorithm confirm considerably faster convergence in terms of the number of the original empirical fitness evaluations.

Benchmarking Gaussian Processes and Random Forests Surrogate Models on the BBOB Noiseless Testbed

Speeding-up black-box optimization algorithms via learning and using a surrogate model is a heavily studied topic. This paper evaluates two different surrogate models: Gaussian processes and random forests which are interconnected with the state-of-the art optimization algorithm CMA-ES. Results on the BBOB testing set show that considerable amount of fitness evaluations can be saved especially during the initial phase of the algorithms progress.

Measures of ruleset quality for general rules extraction methods

The paper deals with quality measures of whole sets of rules extracted from data, as a counterpart to more commonly used measures of individual rules. It sketches the typology of rules extraction methods and of their rulesets, and recalls that quality measures for whole sets of rules have been so far used only in the case of classification rulesets. Then three particular approaches to extending ruleset quality measures from classification to general rulesets are discussed. The paper also recalls the possibility to measure the dependence of classification rulesets on parameters of the classification method by means of ROC curves, and proposes a generalization of ROC curves to general rulesets. Finally, the approach is illustrated on rulesets extracted with four important rules extraction methods from the well-known iris data.

Dynamic classifier aggregation using interaction-sensitive fuzzy measures

In classifier aggregation using fuzzy integral, the performance of the classifier system depends heavily on the choice of the underlying fuzzy measure. However, little attention has been given to the choice of the fuzzy measure in the literature; usually, the Sugeno λ-measure is used. A weakness of the Sugeno λ-measure is that it cannot model the interactions between individual classifiers. That motivated us to develop two novel fuzzy measures and a modification of an existing fuzzy measure which are interaction-sensitive, i.e., they model not only the confidences of classifiers, but also their mutual similarities. The properties of the measures are first studied theoretically, and in the experimental section, the performance of the proposed measures is compared to the traditionally used additive measure and Sugeno λ-measure. Experiments on 23 benchmark datasets and 3 different classifier systems show that the interaction-sensitive fuzzy measures clearly outperform their non-interaction sensitive counterparts.

Neural networks as surrogate models for measurements in optimization algorithms

The paper deals with surrogate modelling, a modern approach to the optimization of objective functions evaluated via measurements. The approach leads to a substantial decrease of time and costs of evaluation of the objective function, a property that is particularly attractive in evolutionary optimization. The paper recalls common strategies for using surrogate models in evolutionary optimization, and proposes two extensions to those strategies - extension to boosted surrogate models and extension to using a set of models. These are currently being implemented, in connection with surrogate modelling based on feed-forward neural networks, in a software tool for problem-tailored evolutionary optimization of catalytic materials. The paper presents results of experimentally testing already implemented parts and comparing boosted surrogate models with models without boosting, which clearly confirms the usefulness of both proposed extensions.

Boosted Neural Networks in Evolutionary Computation

The paper deals with a neural-network-based version of surrogate modelling, a modern approach to the optimization of empirical objective functions. The approach leads to a substantial decrease of time and costs of evaluation of the objective function, a property that is particularly attractive in evolutionary optimization. In the paper, an extension of surrogate modelling with regression boosting is proposed, which increases the accuracy of surrogate models, thus also the agreement between results obtained with the model and those obtained with the original objective function. The extension is illustrated on a case study in materials science. Presented case study results clearly confirm the usefulness of boosting for neural-network-based surrogate models.

Dynamic classifier systems and their applications to random forest ensembles

Classifier combining is a popular method for improving quality of classification - instead of using one classifier, several classifiers are organized into a classifier system and their results are aggregated into a final prediction. However, most of the commonly used aggregation methods are static, i.e., they do not adapt to the currently classified pattern. In this paper, we provide a general framework for dynamic classifier systems, which use dynamic confidence measures to adapt to a particular pattern. Our experiments with random forests on 5 artificial and 11 realworld benchmark datasets show that dynamic classifier systems can significantly outperform both confidence-free and static classifier systems.