The worldwide growth of e-commerce has created new challenges for logistics companies, such as delivering products quickly and cheaply. This paper presents a heuristic to solve the last-mile route creation problem dynamically. The heuristic is based on a multi-agent system integrated with trajectory data mining techniques to extract territorial patterns and use them to solve the Dynamic Capacitated Vehicle Routing Problem with Stochastic Customers. Our solution approach is focused on a linear-time heuristic that depends only on the Warehouse system configurations and not on the total number of packages processed, which is suitable for express delivery logistics companies that must process a large number of packages per day. We compare our proposal with benchmark algorithms from the literature; additionally, we evaluate its performance and robustness under different scenarios. Results show that our solution approach is effective for scenarios in which routes must be set dynamically from a continuous stream of packages.

This paper proposes a new evolving fuzzy model constructed with an unsupervised recursive clustering algorithm with participatory learning and multivariable Gaussian membership functions. The proposed model, called evolving fuzzy with multivariable Gaussian participatory learning and multi-innovations recursive weighted least squares, uses first-order Takagi-Sugeno functional rules. The rules are extracted by the clustering algorithm that can add a new cluster, delete, merge or update existing clusters. The clusters are created using a compatibility measure and an arousal mechanism. The compatibility measure is computed by Euclidian or Mahalanobis distance according to the cluster's number of samples. The clusters exclusion method combines age and population to exclude inactive clusters. Redundant clusters are merged if there is a noticeable overlap between two clusters. The consequent parameters are updated by a multi-innovations weighted least squares recursive algorithm. The performance of the eFMI is evaluated and compared with alternative state-of-the-art evolving models in times series forecasting and non-linear system identification problems. Computational experiments and comparisons suggest the proposed model performs better or comparable than the alternative evolving models.

This paper suggests a new evolving fuzzy approach called eFCE (evolving Fuzzy with Multivariable Gaussian Participatory Learning and Recursive Maximum Correntropy). The approach uses a single pass learning procedure based on a recursive clustering algorithm with participatory learning and multivariable Gaussian membership functions. The eFCE employs first-order Takagi-Sugeno functional rules that may be added, excluded, merged and/or updated depending on the input data. The rules' antecedent is extracted from the clusters, and the consequent parameters are updated by a recursive algorithm of maximum correntropy. The method to create rules uses a compatibility measure and an arousal index. The compatibility measure employs both Euclidean and Mahalanobis distance. The rules elimination procedure combines age and population to exclude inactive rules. Redundant rules are merged if there is a noticeable overlap between two clusters. The performance of the approach is evaluated using instances of times series forecasting. Computational results and comparisons against alternative state-of-the-art evolving models show that the eFCE has better or comparable performance.

Policy gradient methods are amongst the most efficient for on-policy, model-free reinforcement learning. However, they suffer from high variance in gradient updates, making them unstable during training. Subtracting a baseline from the rewards is an effective strategy to reduce variance, such as in actor-critic models. This work presents a variation of the actor-critic model that uses a fuzzy system instead of a neural network to estimate the state value function. The fuzzy value approximation is inspired by previous value-based methods such as fuzzy Q-learning. Experiments with the cart-pole benchmark show that fuzzy value approximation outperforms several reinforcement learning algorithms in terms of sample-efficiency.

Many applications have been producing streaming data nowadays, which motivates techniques to extract knowledge from such sources. In this sense, the development of data stream clustering algorithms has gained an increasing interest. However, the application of these algorithms in real systems remains a challenge, since data streams often come from non-stationary environments, which can affect the choice of a proper set of model parameters for fitting the data or finding a correct number of clusters. This work proposes an evolving clustering algorithm based on a mixture of typicalities. It is based on the TEDA framework and divide the clustering problem into two subproblems: micro-clusters and macro-clusters. Experimental results with benchmarking data sets showed that the proposed methodology can provide good results for clustering data and estimating its density even in the presence of events that can affect data distribution parameters, such as concept drifts. In addition, the model parameters were robust in relation to the state-of-the-art algorithms.

Evolving systems emerge from the synergy between systems with adaptive structures, and the recursive methods of machine learning. Evolving algorithms construct models and derive decision patterns from stream data produced by dynamically changing environments. Different components can be chosen to assemble the system structure, rules, trees, and neural networks being amongst the most prominent. Evolving systems concern mainly with time-varying environments, and processing of nonstationary stream data using computationally efficient recursive algorithms. They are particularly suitable for on-line, real-time applications, and dynamically changing situations, and operating conditions. This chapter gives an overview of evolving systems focusing on the model components, learning algorithms, and illustrative applications. The aim of to introduce the main ideas and a state of the art view of the area.

In this work, we explore the application of modern deep learning techniques to build a neural model centric search engine. We conduct an in-depth discussion under several quantitative and qualitative criteria, comparing the trade-offs of adopting the proposed neural architecture against the successful and mature traditional information retrieval techniques. We show that a full neural architecture, which employs neural models both in the retrieval and ranking phases, offers good scalability, predictability and evolution properties, and discuss under which conditions one can achieve state-of-the-art results. We conclude that deep learning centric systems still require significant more effort to implement and deploy and demand more computational resources, but this work, together with several others in the research community, sheds a light into that path.

Resumo Este artigo apresenta uma metodologia para agrupamento incremental de dados em fluxos cont́ınuos. O método proposto se baseia nos conceitos de tipicidade e excentricidade e no algoritmo CEDAS, recentemente introduzidos. A cada nova amostra recebida, atualiza-se uma estrutura de micro grupos os quais armazenam, dentre outros parâmetros, a densidade local dos dados e a tipicidade local. Em seguida, uma estrutura de macro grupos é atualizada como sendo uma soma das tipicidades dos micro grupos que se sobrepõem ponderadas pela densidade local de cada um destes micro grupos. Ao final tem-se um modelo de mistura de densidades locais que possui a capacidade de agrupar dados de distribuições arbitrárias e gerar como sáıda um valor de pertinência de uma amostra para cada agrupamento. Os resultados preliminares, com bases de dados sintéticas, mostraram que o algoritmo proposto é promissor para aplicações de agrupamento online.

Cryptocurrencies prices forecasting is a complex theme due to the chaotic market behavior and the influence of external events. Therefore, inference models should offer, in addition to a satisfying accuracy, reasonable interpretability, so that investors can decide based on their own knowledge. However, many studies in this subject focus on model accuracy and leave much to be desired in terms of simplicity and interpretability. This work proposes the use of Mamdani interpretable fuzzy inference models for forecasting cryptocurrency price variation. For that, a genetic algorithm to optimize models accuracy is employed, limiting the quantity of rules and antecedents arbitrarily. A set of infeasible rules had to be discarded, in order to generate interesting models, that produce a relevant amount of trades. Data from Kraken exchange were utilized for training, validation and results assessment. Results have shown that, for the cryptocurrencies with the highest validation performances, there are gains in comparison to the simple currency appreciation. Using the interpretable aspect of the models, it should be possible to obtain even higher profits.

Resumo Este trabalho apresenta uma proposta de método map-matching online para pré-processamento de trajetórias de véıculos. O método utiliza um modelo loǵıstico para determinar a probabilidade de localização de uma trajetória pertencer a um segmento que representa uma rua ou estrada. A avaliação do modelo foi feita em duas etapas. A primeira avalia a curva ROC, AUC, acurácia, sensibilidade e especificidade do modelo em diferentes bases de dados. Na segunda etapa foi feita uma comparação com métodos propostos na literatura. A performance dos métodos foi avaliada com relação à média da distância de erro, à mediana da distância do erro e ao Dynamic Time Warping. Os resultados para diferentes bases de dados demonstram que a abordagem é promissora ainda que a medição do GPS seja realizada em situações de diferentes valores de precisão.

The use of time-series from wrist worn accelerometers for Human Activity Recognition is investigated in this work. We employ, as features, coefficients of two-dimensional multivariate/vector autoregressive (AR) models obtained from raw acceleration signals and from estimated wrist attitude roll and pitch angles. It is shown that the simultaneous use of both types of models improves the overall accuracy about 20% when compared to recently published algorithms where only univariate AR models coefficients for each raw acceleration signal are employed.

Multiple Instance Learning (MIL) is a recent paradigm of learning, which is based on the assignment of a single label to a set of instances called bag. A bag is positive if it contains at least one positive instance, and negative otherwise. This work proposes a new algorithm based on likelihood computation by means of Kernel Density Estimation (KDE) called MILKDE. Using the LogitBoost classifier, its performance was compared to that of forty-three MIL algorithms available in the literature using five data sets. Our proposal outperformed all of them for the Elephant (87.40%), Fox (66.80%) and COREL 2000 data sets (77.8%), and achieved competitive results for the MUSK 1 (89.20%) and MUSK 2 (87.50%) data sets, which are comparable to the higher accuracies obtained by other methods for this data sets. Overall results are statistically comparable to those obtained by the most well known methods for MIL described in the literature.

Thermoelectric power plants have critical units, such as the boiler and the turbine-generator, which are complex multivariate systems. These units exhibit non-stationary behavior and multiple operational modes that imply constant changes of set points of key performance variables. A methodology based on MSPC (Multivariate Statistical Process Control) techniques and PCA (Principal Component Analysis) is presented with an adaptive mean estimator that deals with frequent changes of set points, both for design and just in time monitoring. The proposed methodology is implemented in a thermoelectric power plant using a commercial PIMS (Process Information Management System) software suite. Experimental results illustrate and validate the proposition, its just-in-time implementation and usage.

The RBF network is commonly used for classification and function approximation. The center and radius of the activation function of neurons is an important parameter to be found before the network training. This paper presents a method based on computational geometry to find these coefficients without any parameters provided by the user. Experimental results showed that our approach is promising.

After a great advance by the industry on processes automation, an important challenge still remains: the automation under abnormal situations. The first step towards solving this challenge is the Fault Detection and Diagnosis (FDD). This work proposes a batch-incremental adaptive methodology for fault detection and diagnosis based on mixture models trained on a distributed computing environment. The models used are from a family of Parsimonious Gaussian Mixture Models (PGMM), in which the reduced number of parameters of the model brings important advantages when there are few data available, an expected scenario of faulty conditions. On the other hand, a large number of different models rises another challenge, the best model selection for a given behaviour. For that, it is proposed to train a large number of models, using distributed computing techniques, for only then select the best model. This work proposes the usage of the Spark framework, ideal for iterative computations. The proposed methodology was validated in a simulated process, the Tennessee Eastman Process (TEP), showing good results for both the detection and the diagnosis of faults. Furthermore, numeric experiments show the viability of training a large number of models for the best model selection a posteriori.

This paper suggests an evolving approach to develop neural fuzzy networks for system modeling. The approach uses an incremental learning procedure to simultaneously select the model inputs, to choose the neural network structure, and to update the network weights. Candidate models with larger and smaller number of input variables than the current model are constructed and tested concurrently. The procedure employs a statistical test in each learning step to choose the best model amongst the current and candidate models. Membership functions can be added or deleted to adjust input space granulation and the neural network structure. Granulation and structure adaptation depend of the modeling error. The weights of the neural networks are updated using a gradient-descent algorithm with optimal learning rate. Prediction and nonlinear system identification examples illustrate the usefulness of the approach. Comparisons with state of the art evolving fuzzy modeling alternatives are performed to evaluate performance from the point of view of modeling error. Simulation results show that the evolving adaptive input selection modeling neural network approach achieves as high as, or higher performance than the remaining evolving modeling methods.