Wilian Soares Lacerda

Aplicação de redes neurais artificiais na previsão da produção de álcool

Este trabalho descreve a aplicacao de Redes Neurais Artificiais na tarefa de previsao da producao de alcool no Brasil no ano de 2006, a partir de dados de producao anteriores. E tambem apresentada uma comparacao entre os resultados obtidos por meio da Rede Neural com os obtidos utilizando tecnicas de series temporais, sendo que a Rede Neural obteve melhores resultados.

A fuzzy system for cloacal temperature prediction of broiler chickens

Cloacal temperature (CT) of broiler chickens is an important parameter to classify its comfort status; therefore its prediction can be used as decision support to turn on acclimatization systems. The aim of this research was to develop and validate a system using the fuzzy set theory for CT prediction of broiler chickens. The fuzzy system was developed based on three input variables: air temperature (T), relative humidity (RH) and air velocity (V). The output variable was the CT. The fuzzy inference system was performed via Mamdanis method which consisted in 48 rules. The defuzzification was done using center of gravity method. The fuzzy system was developed using MAPLE® 8. Experimental results, used for validation, showed that the average standard deviation between simulated and measured values of CT was 0.13°C. The proposed fuzzy system was found to satisfactorily predict CT based on climatic variables. Thus, it could be used as a decision support system on broiler chicken growth.

Desenvolvimento de uma rede neuro-fuzzy para predição da temperatura retal de frangos de corte

The goal of this work was to develop and validate a neuro-fuzzy intelligent system (LOLIMOT) for rectal temperature prediction of broiler chickens. The neuro-fuzzy network was developed using SCILAB 4.1, on the ground of three Departamento de Engenharia, Universidade Federal de Lavras (UFLA), Caixa Postal 3037, Lavras/MG, Brasil le.ferreira@gmail.com yanagi@deg.ufla.br alisonzille@gmail.com Desenvolvimento de uma rede neuro-fuzzy para predição da temperatura retal de frangos de corte 222 RITA • Volume 17 • Número 2 • 2010 input variables: air temperature, relative humidity and air velocity. The output variable was rectal temperature. Experimental results, used for validation, showed that the average standard deviation between simulated and measured values of RT was 0.11 °C. The neuro-fuzzy system presents as a satisfactory hybrid intelligent system for rectal temperature prediction of broiler chickens, which adds fuzzy logic features based on the fuzzy sets theory to artificial neural networks.

Modeling productive performance of broiler chickens with artificial neural network.

The objective of the present research was to predict the productive performance of broiler chickens as a function of environmental conditions (air temperature - Tair, air relative humidity - RH, and air velocity - V) through artificial neural networks (ANN). Productive performance was accounted through daily weight gain (DWG), daily feed intake (DFI), and feed conversion (FC). Each ANN developed has three input variables (Tair, V and RH), a hidden layer with six neurons, and one output (DWG, DFI or FC). Learning and momentum rates of 1 and 0.7, respectively, were adopted. The neural network was calibrated and validated with literature data, obtained from climatic chamber studies, via SCILAB ANN toolbox. Preliminary results shown that ANN provided promising results for animal performance prediction, allowing simulation of several scenarios. Standard deviations for DWG, DFI and FC were 0.84 g/day, 2.16 g/day, and 0.03 g/g, respectively.

RRS + LS-SVM: a new strategy for “a priori” sample selection

We present in this work a new Sparse Hybrid Classifier, by using reduced remaining subset (RRS) with least squares support vector machine (LS-SVM). RRS is a sample selection technique based on a modified nearest neighbor rule. It is used in order to choose the best samples to represent each class of a given database. After that, LS-SVM uses the samples selected by RRS as support vectors to find the decision surface between the classes, by solving a system of linear equations. This hybrid classifier is considered as a sparse one because it is able to detect support vectors, what is not possible when using LS-SVM separately. Some experiments are presented to compare the proposed approach with two existent methods that also aim to impose sparseness in LS-SVMs, called LS 2-SVM and Ada-Pinv.We present in this work a new Sparse Hybrid Classifier, by using reduced remaining subset (RRS) with least squares support vector machine (LS-SVM). RRS is a sample selection technique based on a modified nearest neighbor rule. It is used in order to choose the best samples to represent each class of a given database. After that, LS-SVM uses the samples selected by RRS as support vectors to find the decision surface between the classes, by solving a system of linear equations. This hybrid classifier is considered as a sparse one because it is able to detect support vectors, what is not possible when using LS-SVM separately. Some experiments are presented to compare the proposed approach with two existent methods that also aim to impose sparseness in LS-SVMs, called LS2-SVM and Ada-Pinv.

Reconfigurable co-processor for Kanerva's sparse distributed memory

Abstract The implementation on hardware of the first layer of Kanervas sparse distributed memory (SDM) is presented in this work. The hardware consist on a co-processor board for connection on ISA standard bus of an IBM–PCcompatible computer. The board, named reconfigurable co-processor for SDM(RC-SDM), comprises on Xilinx FPGAs, local random access memory and bus interface circuits. Based on in-system reconfiguration capacity of FPGAs, RC-SDM easily allows change of the characteristics of SDM topology implemented. First results show a speed-up of four times of RC-SDM in relation to a software implementation of the algorithm.

PIPELINED ON-LINE BACK-PROPAGATION TRAINING OF AN ARTIFICIAL NEURAL NETWORK ON A PARALLEL MULTIPROCESSOR SYSTEM

This work presents an on-chip learning of artificial neural networks in a FPGA multiprocessor system, where each neuron is implemented in a soft-core processor. In order to take maximum advantage of the distributed architecture, a pipelined version of the on-line back-propagation algorithm is used, providing a high degree of parallelism between neuron layers and, hence, a higher speed-up in relation to a sequential implementation.

USE OF ARTIFICIAL NEURAL NETWORKS FOR PROGNOSIS OF CHARCOAL PRICES IN MINAS GERAIS

Energy is an important factor of economic growth and is critical to the stability of a nation. Charcoal is a renewable energy resource and is a fundamental input to the development of the Brazilian forest-based industry. The objective of this study is to provide a prognosis of the charcoal price series for the year 2007 by using Artificial Neural Networks. A feedforward multilayer perceptron ANN was used, the results of which are close to reality. The main findings are that: real prices of charcoal dropped between 1975 and 2000 and rose from the early 21st century; the ANN with two hidden layers was the architecture making the best prediction; the most effective learning rate was 0.99 and 600 cycles, representing the most satisfactory and accurate ANN training. Prediction using ANN was found to be more accurate when compared by the mean squared error to other studies modeling charcoal price series in Minas Gerais state.

Design of digital classifier circuits with nearest neighbour prior sample selection

A new method for design of digital classification circuits is presented in this paper in order to implement them in hardware (FPGA, PAL, VLSI, ASIC, etc). The method works by first selecting a subset of the training data that is just off the separation margin between the classes. The subset is provided to a Boolean minimization algorithm that, by hypercube expansion, designs a classifier with a smoother separation surface between classes. The obtained circuits have performance comparable to support vector machines and multilayer perceptron trained with a generalization control algorithm.