Pedro Henrique Gouvêa Coelho

A complex EKF-RTRL neural network

The purpose of the paper is to use extended Kalman filter (EKF) techniques in the complex real time recurrent learning (RTRL) neural network in order to have faster convergence as an alternative to standard gradient methods, usually used in RTRL neural networks training which are known to be slow. This characteristic is desired in many engineering applications, particularly those which are carried out online.

Flow estimation using an Elman networks

This paper investigates the application of partially recurrent artificial neural networks (ANN) in the flow estimation for Sao Francisco river that feeds the hydroelectric power plant of Sobradinho. An Elman neural network was used, suitably arranged to receive samples of the flow time series data available for Sao Francisco river shifted by one month. The data used in the application concern to the measured Sao Francisco river flow time series from 1931 to 1996, in a total of 65 years from what 60 were used for training and 5 for testing. The obtained results indicate that the Elman neural network is suitable to estimate the river flow for 5 year periods monthly. The average estimation error was less than 0.2%.

A new state space model for a complex RTRL neural network

The purpose of the work is to represent the complex real time recurrent learning (RTRL) fully recurrent neural network in a state space model for engineering applications such as mobile channel equalization. This representation extends Haykins (1999) for complex valued inputs, yielding a compact formulation useful in possible changes in the training of a fully recurrent neural network. Numerical results are presented to illustrate the method.

Further results on the EKF-CRTRL equalizer for fast fading and frequency selective channels

This paper shows further results on the EKF-RTRL (extended Kalman filter-real time recurrent learning) equalizer comparing its performance with the PSP-LMS (per survivor processing-least mean squares) equalizer for fast fading selective frequency channels using the WSS_US (wide sense stationary-uncorrelated scattering) model. The EKF-RTRL is a symbol by symbol neural equalizer and the PSP-LMS equalizer uses the maximum likelihood criterion for symbol sequence estimation and the per survivor processing principle. The performance here presented depicts several scenarios regarding the channel variation speed. The performance considered in this paper is the symbol error rate (SER). A comparison involving the computational complexity of both equalizers is also carried out.

An extended RTRL training algorithm using Hessian matrix

An extended real time recurrent learning (RTRL) algorithm using Hessian matrix is proposed. The algorithm is suitable for small fully recurrent neural networks present in several applications. Simulation results indicate that the training algorithm is fast.

ESTIMATING GREENHOUSE GAS EMISSIONS USING COMPUTATIONAL INTELLIGENCE

This work proposes a Neuro-Fuzzy Intelligent System – ANFIS (Adaptive Network based Fuzzy Inference System) for the annual forecast of greenhouse gases emissions (GHG) into the atmosphere. The purpose of this work is to apply a Neuro-Fuzzy System for annual GHG forecasting based on existing emissions data including the last 37 years in Brazil. Such emissions concern tCO2 (tons of carbon dioxide) resulting from fossil fuels consumption for energetic purposes, as well as those related to changes in the use of land, obtained from deforestation indexes. Economical and population growth index have been considered too. The system modeling took into account the definition of the input parameters for the forecast of the GHG measured in terms of tons of CO2. Three input variables have been used to estimate the total tCO2 one year ahead emissions. The ANFIS Neuro-Fuzzy Intelligent System is a hybrid system that enables learning capability in a Fuzzy inference system to model non-linear and complex processes in a vague information environment. The results indicate the Neural-Fuzzy System produces consistent estimates validated by actual test data.

Complex Kalman Filter Trained Recurrent Neural Network Based Equalizer for Mobile Channels

This paper presents a two state neural equalizer for mobile channels. The channel is modeled by the wide sense stationary - uncorrelated scattering (WSS-US) channel which is known to be an adequate model for wireless applications. The neural equalizer is trained by an extended Kalman filter in order to speed up the equalizer training. Simulation results are also shown in the paper for several scenarios indicating a good trade-off in performance and computational complexity. Comparisons involving traditional equalizers such as decision feedback equalizers (DFE) are also shown indicating that the proposed equalizer outperforms DFE equalizers. On the other hand, the proposed neural equalizer is outperformed by per survivor processing the (PSP) class of equalizers which are much more computational complex than the neural class of equalizers proposed in this paper.

Further Developments on Router Nodes Positioning for Wireless Networks using Artificial Immune Systems

This paper shows further developments on the positioning of intermediate router nodes using artificial immune systems for use in industrial wireless sensor networks. These nodes are responsible for the transmission of data from sensors to the gateway in order to meet criteria, especially those that lead to a low degree of failure and reducing the number of retransmissions by routers. In the present paper positioning configurations on environments in presence of obstacles is included. Affinity functions which roles are similar to optimization functions are explained in details and case studies are included to illustrate the procedure. As was done in previous papers, positioning is performed in two stages, the first uses elements of two types of immune networks, SSAIS (Self-Stabilising Artificial Immune System) and AINET (Artificial Immune Network), and the second uses potential fields for positioning the routers such that the critical sensors attract them while obstacles and other routers repel them.

Applying Artificial Immune Systems for Deploying Router Nodes in Wireless Networks in the Presence of Obstacles

This work deals with the deploying of router nodes using artificial immune systems techniques, particularly for industrial applications of wireless sensor networks. Possible scenarios include configurations with sensors blocked by obstacles. These nodes make possible the transmission of data from sensors to the gateway in order to meet criteria especially those that lead to a low degree of failure and reducing the number of retransmissions by the routers. These criteria can be set individually or in groups, associated with weights. Router nodes deploying is accomplished in two phases, the first uses immune networks concepts and the second employs potential fields ideas for deploying the routers in such way that critical sensors attract them while obstacles and other routers repel them. A large number of case studies were considered from which some representative ones were selected to illustrate the method, for different configurations in the presence of obstacles.

Use of Genetic Algorithm for Spatial Layout of Indoor Light Sources

People spend many hours inside buildings that are naturally and artificially illuminated. Since mankind has been able to tame the fire and use it to illuminate, the natural condition of nighttime darkness has been modified. With the advent of electric lighting this has been intensified. The problem of indoor lighting presents several options according to the specific purpose of the lighting. There is room for some heuristic choices and genetic algorithms have been chosen as a computational intelligence technique that allows multi-objective solutions and the inclusion of heuristics and versatility in specific situations that occur in many particular applications. In this way, the main objective of this article is to optimize the number of light sources in indoor environments with the aid of genetic algorithms to obtain a suitable light intensity with the smallest number of light sources. One of the paramount reasons for using such algorithm is that it returns an acceptable solution to an optimization problem with infinite possibilities in a finite number of trials. A case study is presented in which the applicability of genetic algorithms to the problem is discussed, and the results indicate the viability of the method.