J. A. Anochi

New learning strategy for supervised neural network: MPCA meta-heuristic approach

The problem of parameter optimization for a feedforward artificial neural network (ANN) to determined its best architecture is addressed. A new metaheuristic called Multiple Particle Collision Algorithm (MPCA), introduced by Luz et al. [12], was applied to design an optimum architecture for two models of supervised neural network: the Multilayer Perceptron (MLP), and recurrent Elman network. The NN obtained using this approach is said to be self-configurable. In addition, two strategies are employed for calculating the connection weights to the MLP and Elman networks: MPCA, and backpropagation algorithm. The resulting ANNs were applied to predict the monthly mesoscale climate for the precipitation field. The comparison is performed between the ANN configuration obtained by automatic process and another configuration proposed by a human specialist.

Data assimilation by artificial neural networks for the global FSU atmospheric model: Surface pressure

Data assimilation is the process by which measurements and model predictions are combined to obtain an accurate representation of the state of the modelled system as its initial condition. This paper shows the results of a data assimilation technique using artificial neural networks (NN) to obtain the initial condition to the atmospheric general circulation model (AGCM) for the Florida State University in USA. The Local Ensemble Transform Kalman filter (LETKF) is implemented with Florida State University Global Spectral Model (FSUGSM). LETKF is a version of Kalman filter with Monte-Carlo ensembles of short-term forecasts to solve the data assimilation problem. FSUGSM is a multilevel spectral primitive equation model with a vertical sigma coordinate, at resolution T63L27. The LETKF data assimilation experiments are based in simulated observations data. For the NN data assimilation scheme, we use Multilayer Perceptron (MLP-DA) with supervised training algorithm where NN receives input vectors with their corresponding response from LETKF scheme. The surface pressure results are presented. An self-configuration method finds the optimal NN and configures the MLP-DA in this experiment. The NNs were trained with data from each month of 2001, 2002 and 2003. A experiment for data assimilation cycle using MLP-DA was performed with simulated observations for January of 2004. The results demonstrate the effectiveness of the ANN technique for atmospheric data assimilation, with similar quality to LETKF analyses.

Optimization of feedforward neural network by Multiple Particle Collision Algorithm

Optimization of neural network topology, weights and neuron activation functions for given data set and problem is not an easy task. In this article, a technique for automatic configuration of parameters topology for feedforward artificial neural networks (ANN) is presented. The determination of optimal parameters is formulated as an optimization problem, solved with the use of meta-heuristic Multiple Particle Collision Algorithm (MPCA). The self-configuring networks are applied to predict the mesoscale climate for the precipitation field. The results obtained from the neural network using the method of data reduction by the Theory of Rough Sets and the self-configuring network by MPCA were compared.

CLIMATE PRECIPITATION PREDICTION FOR SOUTH REGION BY NEURAL NETWORK SELF-CONFIGURED

Climate prediction for precipitation field is a key issue, because such meteorological variable is the challenge for climate and weather forecasting due to the high spatial and temporal variability with strong impact on the society. A method based on the artificial neural network is applied to monthly and seasonal precipitation forecast in southern Brazil. The use of neural networks as a predictive model is widespread in different applications. The best configuration for the neural network is automatically calculated. The autoconfiguration scheme is described as an optimization problem.

Neural Network Models for Climate Forecasting based on Reanalysis Data

In this work a neural network model for climate forecasting is presented. The model is built by training a neural network with available reanalysis data. In order to assess the model, the development methodology considers the use of data reduction strategies that eliminate data redundancy thus reducing the complexity of the models. The results presented in this paper considered the use of Rough Sets Theory principles in extracting relevant information from the available data to achieve the reduction of redundancy among the variables used for forecasting purposes. The paper presents results of climate prediction made with the use of the neural network based model.

A Multi-objective Version of the Multiple Particle Collision Algorithm

A multi-objective version of a meta-heuristic, loosely inspired on the behaviour of particles inside a nuclear reactor, is presented. The Multi-objective Multiple Particle Collision Algorithm (MMPCA) uses the Pareto based fitness assignment, which uses the concept of dominance, to generate new solutions and build the Pareto set. The original population is duplicated, with purpose of classification and for applying the crowding distance approach. The latter procedures are also used in the NSGA-II.

MPCA META-HEURISTICS FOR AUTOMATIC ARCHITECTURE OPTIMIZATION OF A SUPERVISED ARTIFICIAL NEURAL NETWORK

Artificial neural networks (ANN) has been studied intensively, but there still are many unresolved issues. The search and definition of an optimal architecture remains a very relevant ANN research topic. The search space of neural network topology, each point rep- resents a possible architecture. Associating each point to a performance level relies on the a priori establishment of some optimality criterion. Here, a new meta-heuristics, multi-particle collision algorithm (MPCA) was applied to design an optimum architecture for a supervised ANN. The MPCA optimization algorithm emulates a collision process of multiple particles inspired in processes of a neutron traveling in a nuclear reactor. The multilayer perceptron (MLP) was the neural network adopted here, and backpropagation strategy was used for calculating of the weight of connections to the MLP-NN. The MLP-NN configured by this op- timal or inverse designs was applied to predict the seasonal mesoscale climate. The dataset for trainning is obtained from NCEP-NOAA reanalysis and from a metherological model. In order to reduce the dimension of the search space to find the optimized ANN, it is considered the following: three activation functions, up to three hidden layers, and up to 32 neurons per hidden layer. The comparison is performed between the ANN configuration obtained by automatic process and another configuration proposed by a human specialist.