Differential evolution requires a prior setting of its parameters. Appropriate values are not always easy to determine, even more since they may change during the optimisation process. This is where parameter control comes in. Accordingly, a scheme inspired by ant colony optimisation for controlling the crossover-rate and mutation factor is proposed. Conceptually, artificial ants guided by a pheromone model select parameter value pairs for each individual. The pheromone model is steadily updated in order to reflect the current optimisation state. Promising results were achieved in the comprehensive experimental analysis. Nonetheless, much room for potential improvements is available.

Differential Evolution

Source

Cite

Citations (1)

Effectiveness of differential evolution in training radial basis function networks for classification

2022 International Conference on Smart Systems and Technologies (SST) (2016)

Dražen Bajer Bruno Zorić Goran Martinović

Building classification models often presents a significant problem that requires the selection of a classifier and a corresponding training approach. Radial basis function networks are a frequent choice among the classifiers for which a large spectre of training approaches exist. In that regard, an important role is played by bio-inspired methods, and differential evolution, as an representative example, has been applied for training such networks. This paper investigates the behaviour of differential evolution in training radial basis function networks primarily from the perspective of fitting the model to available (training) data rather than its performance on unknown (testing) data. This is believed to provide a clearer insight into optimiser efficiency. Another important issue considered is a steady emergence of new bio-inspired methods claiming superior performance that can be witnessed in the literature. It may raise the question whether differential evolution is still competitive to those approaches. In light of this, the canonical differential evolution algorithm has been compared to a couple of recently proposed and a well established swarm intelligence algorithm.

Differential Evolution

Basis (linear algebra)

10.1109/sst.2016.7765655

Cite

Citations (2)

Empirical Analysis of Artificial Bee Colony Algorithm Parameters

2022 International Conference on Smart Systems and Technologies (SST) (2018)

Dražen Bajer Bruno Zorić Goran Martinović

The design of smart systems frequently includes various problems such as parameter estimation and optimisation, feature subset selection or model tuning. As a rule, these problems are quite challenging and as a way of tackling them, bio-inspired algorithms are recently becoming the approach of choice. Although a multitude of these algorithms is available, the artificial bee colony algorithm represents a viable candidate due to its good performance that has been demonstrated in a wide array of applications. However, the aforementioned performance is heavily reliant on the chosen parameter values. Tuning those parameters represents a significant ordeal. This paper is aimed at the empirical analysis of parameter influence. To this end, a somewhat detailed experimental analysis is conducted in order to compare the effectiveness of numerous parameter combinations. The paper also endeavours to provide certain guidelines with the hope of easing the utilisation of said algorithm for researchers and practitioners alike.

Empirical Research

10.1109/sst.2018.8564632

Cite

Citations (2)

Evaluation and analysis of bio-inspired optimization algorithms for feature selection

Dražen Bajer Bruno Zorić Mario Dudjak Goran Martinović

Wrapper-based feature (subset) selection is widely used as an effective means for decreasing the dimensionality of datasets. However, it is not the most efficient approach in terms of computational cost. Hence, the choice of the wrapper is paramount. Ideally, the wrapper should be simple to use and understand, whilst yielding good solutions as fast as possible. Bioinspired optimisation algorithms are a common choice in that regard, but not all are made equally. This paper investigates a number of optimisers on diverse datasets in order to provide an insight into their efficiency and behaviour with respect to the problem of dimensionality reduction for classification needs. Correspondingly, some guidelines concerning the choice of the wrapper are given.

Feature (linguistics)

Source

Cite

Citations (0)

Adaptive k-tournament mutation scheme for differential evolution

Applied Soft Computing (2019)

Dražen Bajer

Tournament selection

Benchmark (surveying)

Differential Evolution

Base (topology)

10.1016/j.asoc.2019.105776

Cite

Citations (10)

Predicting public transport arrival time and congestion based on BLE beacon crowdsourced data

Bruno Zorić Mario Dudjak Dražen Bajer

Public transport networks play an important role in minimising congestion and improving environmental sustainability of developed cities. However, they face a number of challenges in achieving these goals, especially during the ongoing pandemic. In order to overcome these challenges, at least to some extent, public transport must be made accessible and attractive to potential passengers. To this end, a design for augmenting the transit network is proposed in this paper. The utilisation of Bluetooth low energy beacons as one of the key components makes it cost-effective and easily applicable in such an environment. Additionally, it incorporates a simple mobile application used to enable crowdsourced data acquisition on which machine learning-based models can be built to predict information relevant to consumers, like arrival time and congestion estimates. A prototype of the proposed system design, albeit of limited functionality, was deployed and evaluated on a tram route in the city of Osijek, Croatia. Promising results were obtained in terms of congestion and arrival time prediction, but some challenges remain to be addressed, like motivating users to participate in the crowdsourced data collection.

Beacon

Arrival time

Crowdsourcing

10.1109/zinc55034.2022.9840632

Cite

Citations (2)

Feature extraction procedures for chronic wound tissue classification

Bruno Zorić Dražen Bajer Mario Dudjak

Feature (linguistics)

Chronic wound

10.1109/zinc61849.2024.10579393

Cite

Citations (0)

SMOTE Inspired Extension for Differential Evolution

Lecture notes in computer science (2022)

Dražen Bajer Bruno Zorić Mario Dudjak

Benchmark (surveying)

Differential Evolution

10.1007/978-3-031-21094-5_16

Cite

Citations (0)

Unaprjeđenja algoritma diferencijalne evolucije podešavanjem parametara i izborom počrtne populacije

Dražen Bajer

The need or tendency to improve different systems or models can be encountered in numerous forms of engineering and science. For that purpose, frequently optimization problems must be tackled. Such problems usually posses different properties that make them hard to solve. Besides, it is not uncommon that they are like black boxes which only provide responses to given inputs. Dealing with numerical optimization problems, differential evolution (DE) as an representative evolutionary algorithm (EA) may be pointed out, which is also in the center of the thesis. Although relatively simple, its performance is acclaimed. Three enhancements of DE are proposed in the thesis. Due to sensitivity to parameter values and the problem of determining appropriate ones, a self-adaptive scheme for controlling the scale factor and crossover-rate is proposed. It features, for each population member, the maintenance of a number of previously successful parameter values that are used for generating new ones. Also, the initial population may play an important factor in the performance of DE, and an initialization method is proposed which is based on data clustering and Cauchy deviates. The key element and difference between DE and other common EAs is its mutation. A mutation that employs an adaptive k-tournament selection for determining the base vector is proposed. All proposed enhancements were extensively tested on a set of selected standard functions and benchmark functions prepared for the CEC 2014 competition on numerical optimization. The analysis of the obtained results led to the conclusion that the enhancements considerably improve the performance of the algorithm incorporating them, but also that they compare favorably against similar enhancements from the literature. Finally, the behavior and performance of the standard DE algorithm were investigated in designing radial basis function networks for classification. The design of such networks represents a complex global optimization problem primarily from the viewpoint of the number of parameters that need to be adjusted and expansive evaluations. Although the standard algorithm is a viable choice for tackling this problem, the performed testing and analysis showed that the proposed enhancements yield improved performance in this case as wel

Initialization

Benchmark (surveying)

Differential Evolution

Tournament selection

Source

Cite

Citations (0)