Advances in intelligent systems

The power of different intelligent techniques, such as Artificial Neural Networks (ANNs), have been demonstrated over the years by their successful use in many types of problems with different degrees of complexity and in different fields of application. However, the use of a machine learning solution involves choosing a particular model and properly tuning a set of parameters that, together, can effectively and efficiently solve the specific problem. Some difficulties arise in this process, such as the existence of several machine learning algorithms, the exponential number of combinations of parameter values, and also the need for a priori knowledge on the problem domain. Moreover, certain complex learning problems cannot be solved by a single intelligent technique and each intelligent technique has particular computation properties that can be complementary. This special issue has a collection of papers on intelligent systems design and applications. The papers have a combination of many different techniques including neural networks, fuzzy systems, evolutionary computation, evolving systems and metalearning. Applications examples are from multiobjective and many objective optimization problems, time-changing data and others. The Extreme Learning Machine (ELM) is an efficient training algorithm for single hidden layer feedforward neural networks, known to obtain neural network models with better generalization performance while consuming far less time than other traditional learning algorithms. Nonetheless, the generated models tend to present more neurons in the hidden layer, which affects the time taken for future predictions. Particle Swarm Optimization (PSO) algorithms can be used to find good performing ELM networks with a lower number of neurons. Paper 1, Investigating the use of alternative topologies on performance of the PSO–ELM by Elliacking Figueiredo and Teresa Ludermir, presents an ELM-PSO hybridization, considering different PSO topologies. While the performance of the PSO depends strongly on its topology, there is no outright best topology for all problems. Therefore, the authors investigate the effect of the PSO topologies on the performance of the PSO–ELM for training single hidden layer neural networks. Experimentally, the frequently used global topology showed better results compared to the other topologies tested regarding the root mean squared error on validation data. Learning from data streams is a contemporary and demanding issue because of the constantly increasing rate in size and temporal availability of data. Paper 2, Uninorm Based Evolving Neural Networks and Approximation Capabilities by Fernando Bordgnon and Fernando Gomide, suggests a structure and introduces a learning approach to train uninorm-based hybrid neural networks using extreme learning concepts. Uninorms bring flexibility and generality to fuzzy neuron models as they can benefit from triangular norms, triangular conorms, or operations in between by adjusting identity elements. The Fuzzy c-means is used to granulate the input space, and a scheme based on extreme learning is employed to compute the weights of the neural network approximating any continuous functions in compact domains. Subsequently, an evolving version of the network is developed exploring recursive clustering methods and extreme learning. It is postulated, and computational experiments endorse, that the evolving neuro fuzzy network share equal or better approximation ability in dynamic environments than its static counterpart. The response of natural and artificial systems to external signal has been studied deeply, especially in biological systems. This response to a weak signal can be enhanced in the presence of a specific intensity of stochastic input noise. Such phenomenon highlights that noise may play a constructive role in signal processing. As noise is a common feature in nature, there are many research works concerning on the positive impact of noisy factors, with the objective to explore and understand the phenomenon of stochastic resonance. Some works suggested that the use of excitable neuronal system profited from the noise of the neurons themselves to optimize the response to external stimulus, where the noise is not from the environment fluctuations but from the intrinsic diversity of neurons. The third paper, Effect of nonidentical signal phases on signal amplification of two coupled excitable neurons by Xiaoming Liang and Liang Zhao, investigates the response of two coupled excitable neurons to external subthreshold periodic signals. The neuron's response to the subthreshold signals can be enhanced when the signal phases are nonidentical. Moreover, the response can be strongly improved if the level of nonidentical signal phases is at an intermediate value and thus resulting in a resonance-like phenomenon. Under the influence of nonidentical signal phases the response also exhibits a resonance-like dependency on the coupling coefficient. This paper presents an analysis to understand the mechanism of these two types of resonance behaviors. The robustness of the resonances induced by nonidentical signal phases to neuronal diversity, noise perturbation and different neuron models is verified and can be useful for understanding neural information processing. Most machine learning and data mining techniques consider a scenario where data are centralized. This includes the popular k-means algorithm for clustering data. For dealing with data distributed in separated repositories, distributed versions of k-means have been proposed. Paper 4, Evolutionary k-means for distributed data sets by Murilo Naldi and Ricardo Campello, uses a hybrid evolutionary k-means algorithm for clustering distributed data. The evolutionary algorithm evolves clustering partitions that are refined by the k-means algorithm. A master node combines the information from multiple data nodes, which distribute the data. This algorithm is called Distributed Fast Evolutionary Algorithm for Clustering (DF-EAC). Two different distribution approaches are