Identification of underground formation lithology from well-log data is an important task in petroleum exploration and engineering. Due to the cost or imprecision of some methods applied in this activity, there is a need to automate the procedure of reservoir characterization. Machine learning techniques can be efficient alternatives to lithology identification. To acquire proper performance, usually, some parameters of these techniques should be adjusted, and this can become a hard task depending on the complexity of the underlying problem. This letter integrates the gradient boosting (GB) with a differential evolution (DE) for formation lithology identification using data from the Daniudui gas field and the Hangjinqi gas field. This letter's contributions include the use of an evolutionary algorithm to adjust optimally the hyperparameters of the GB, and the results show improvements when compared with those obtained in the literature.

The Island Model (IM) is an alternative to implement evolutionary algorithms to be executed in parallel architectures. An important feature of the IM is the process called migration where islands exchange solutions between themselves periodically along iterations of their algorithms. Parameters to be set by the user define how the migration will occur. Different strategies for the migration process have already been proposed and evaluated in the literature. This paper extends the dynamic Island Model (D-IM) proposed in the literature and proposes a new strategy to evaluate the attractiveness of the islands in the model. Some properties of the two configurations for the D-IM were compared. Besides the quality of the solutions, the adjustment of the topology and the movement of solutions between islands were objects of interest in this work.

The basic principle of the building energy efficiency is to use less energy for operations such as heating, cooling, lighting and other appliances, without impacting the health and comfort of its occupants. In order to measure energy efficiency in a building, it is necessary to estimate its heating and cooling loads, considering some of its physical characteristics such as geometry, material properties as well as local weather conditions, project costs and environmental impact. Machine Learning Methods can be applied to solve this problem by estimating a response from a set of inputs. This paper evaluates the performance of Gaussian Processes, also known as kriging, for predicting cooling and heating loads of residential buildings. The dataset consists of 768 samples with eight input variables and two output variables derived from building designs. The parameters were selected based on exhaustive search with cross validation. Four statistical measures and one synthesis index were used for the performance assessment and comparison. The results show Gaussian Processes consistently outperform other machine learning techniques such as Neural Networks, Support Vector Machines and Random Forests. The proposed framework resulted in accurate prediction models contributing to savings in the initial phase of the project avoidlng the modeling and testing of several designs.

In this paper, a Particle Swarm Optimization algorithm is used to adjust the parameters of an Extreme Learning Machine and select features in order to predict mechanical properties of lightweight aggregate concretes. Unlike the approaches found in the literature, the proposed procedure set the model parameters and select the most beneficial subset of features while simultaneously estimates two important outcomes: the compressive strength and elasticity modulus. These properties can be modeled as a function of up to four features: water/cement fraction, lightweight aggregate volume, cement quantity and lightweight aggregate density. The Particle Swarm Optimization algorithm performs the parameter and feature selection and automatically tunes the number of neurons in the hidden layer and the activation function. The results are compared with a model selection based on exhaustive search on the parameter space. The proposed approach arises as an alternative tool to select the most relevant features and to estimate the mechanical properties of lightweight aggregate concretes.

The short circuit gas metal arc welding is widely used in the industry, having an important role in manufacturing processes. In order to achieve high-quality control levels in weldments, researches are carried out aiming to find a way to monitor the quality of weldments in real time. In this context, this work proposes to use a type-1 and singleton fuzzy logic system classifier to identify the gas flow rate in short circuit gas metal arc welding. The investigated dataset was performed in laboratory, consisting of the current and voltage signals related with weld beads, for different gas flow rates. In addition, the feature extraction uses the statistics from the data signal and a criterion to quantify the metal transfer stability in short circuit processes, denominated Laprosolda Criterion. Based on the excellent performance of the model, the proposal is suitable to be used in welding quality monitoring.

The Island Model is a mechanism that promotes improvement in the quality of the results produced by evolutionary algorithms and speed up their executions. A reason of the impact caused by the Island Model in the quality of results is the migration of solutions between islands that occurs periodically during the search process. The migration process depends on decisions such as the choice of solutions that will be send, the destination islands etc. This set of decisions is known as migration policy. This paper proposes a migration policy to the Island Model in which the destination island for an emigrant solution is defined according to the attractiveness of the islands in the model. In the proposed model the attractiveness between islands also influences the connection between them and affect the topology of the model. This paper evaluated if the proposed model is able to maintain the two main characteristics of the Island Model. The movement of solutions and the states of the connections were evaluated too.

This paper presents a study on the use of fitness inheritance as a surrogate model to assist a genetic algorithm (GA) in solving optimization problems with a limited computational budget.We compared the impact to the evolutionary search introducing three surrogate models: (i) averaged inheritance, (ii) weighted inheritance and (iii) parental inheritance. Numerical experiments are performed in order to assess the applicability and the performance of the proposed approach. The results show that when using a fixed reduced budget of expensive simulations, the surrogate-assisted genetic algorithm allows for improving the final solutions when compared to the standard GA. We find that the averaged and parental inheritance are more effective when compared to weighted inheritance, and they are recommended for expensive of optimization problems using GA-based search.

In order to find a satisfactory solution, genetic algorithms, in spite of their ability to solve difficult optimization problems, usually require a large number of fitness evaluations. When expensive simulations are required, using genetic algorithms as optimization tools can become prohibitive. In this paper we present a strategy for introducing surrogate models into genetic algorithms in order to enhance the quality of the final results, where a fixed budget of simulations is imposed. In this strategy, only a fraction of the population is evaluated by the exact function, thus allowing for more generations to evolve the population. The results obtained indicate that the proposed framework arises as an attractive alternative to improve the performance of the genetic algorithm within a fixed budget of expensive fitness evaluations.

A genetic algorithm (GA) is hybridized with an artificial immune system (AIS) as an alternative to tackle constrained optimization problems in engineering. The AIS is inspired in the clonal selection principle and is embedded into a standard GA search engine in order to help move the population into the feasible region. The procedure is applied to mechanical engineering problems available in the literature and compared to other alternative techniques.

Penalty methods are often used to handle constraints in optimization problems. However, to find the optimal or near optimal set of penalty parameters is a hard task. Also, such values are problem dependent. This paper introduces the stochastic ranking approach to balance objective and penalty functions stochastically in a rank-based ACO metaheuristic. The results presented show that the simple inclusion of the procedure leads to an improved search performance, with respect to the standard penalty technique, when applied to discrete structural optimization problems