Luis M. Torres-Treviño

Multi-objective optimization of a welding process by the estimation of the Pareto optimal set

The joining of Advanced High Strength Steel (AHSS) Martensitic type is being introduced in automotive industry; however, the optimization of the welding process is required to meet customer quality requirements. Two neural networks are built for modeling the relationship between the welding parameters and the output response of the process. An evolutionary algorithm is used for multi-objective optimization considering the neural networks as objective functions. The results consist of a set of solutions that approximate the Pareto optimal set. The related response of this set is known as the Pareto front. The set of solutions are validated in the real process satisfying the security and quality requirements.

Analysis and evaluation in a welding process applying a Redesigned Radial Basis Function

Highlights? We made a RBFNN Redesigned that showed good performance in a real case. ? The Hybrid Learning Process presented applies a GA to calculate the matrix of centers. ? The coefficient of determination becomes the statistical evaluation function of GA. ? The Hybrid Learning Process does not need aggregate the widths in the hidden layers. The Hybrid Learning Process method proposed in this work, is applied to a Genetic Algorithm and Mahalanobis distance, instead of computing the centers matrix by Genetic Algorithm. It is determined in such a way as to maximize the coefficient of determination R2 and the Fitness Function depends on the prediction accuracy fitted by the Hybrid Learning approach, where the coefficient of determination R2 is a global metric evaluation. The Mahalanobis distance is a measurement of distance which uses the correlation between variables and takes into account the covariance and variance matrix in the input variables; this distance helps to reduce the variance into variables. The purpose of this work is to show a methodology to modify the Radial Basis Function and also improve the parameters and variables that are associated with Radial Basis Function learning processes; since the Radial Basis Function has mainly two problems, the Euclidean distance and the calculation of centroids. The results indicated that the statistical methods such as Residual Analysis are good alternatives and excellent methods for validation of the effectiveness of the Neural Network models. The principal conclusion of this work is that the Radial Basis Function Redesigned improved the accuracy of the model using a Hybrid Learning Process and the Radial Basis showed very good performance in a real case, considering the prediction of specific responses in a laser welding process.

On-Line Learning in an Embedded Maximum Sensibility Neural Network

A maximum sensibility neural networks was implemented in an embedded system to make on-line learning. This neural network has advantages like easy implementation and a quick learning based on manage information in place of a gradient algorithm. The embedded maximum sensibility neural network was used to learn non linear functions on-line using potentiometers and a push button giving the function of activation and learning. The results give us a platform to apply on-line learning using neural networks.

Identification and prediction using symbolic regression alpha-beta: preliminary results

A novel approach is proposed for generating equations from measured data of dynamic processes. A composition of unary (alpha) and binary (beta) functions is represented by a real vector and adapted by an evolutionary algorithm to build mathematical equations. The equations can be used for identification and prediction considering a mathematical model with specific number of inputs and outputs. Three cases are used for illustration of the approach where mathematical models and plots of theirs performance are presented with promising results.

An Embedded Fuzzy Agent for Online Tuning of a PID Controller for Position Control of a DC Motor

The aim of the presented work is to illustrate the performance of embedded fuzzy agent for online tuning of a PID controller. A DC motor is used for illustration of its position control. An agent is built using an embedded fuzzy system type 1. Performance of the PID controller is evaluated on line and this performance is improved in low cycles of tuning by the agent.

Control by Online Learning Using a Maximum Sensibility Neural Network

In this paper, a maximum sensibility neural network is proposed to make an online learning system of a inverse controller of a plant. This neural network is trained to learn the response of the plant to different random inputs. Once the network is trained, it can be used to control the plant to a desired output.

Statistical inference in a redesigned Radial Basis Function neural network

A Hybrid Learning Process method was fitted into a RBF. The resulting redesigned RBF intends to show how to test if the statistical assumptions are fulfilled and to apply statistical inference to the redesigned RBFNN bearing in mind that it allows to determine the relationship between a response (to a process) and one or more independent variables, testing how much each factor contributes to the total variation of the response is also feasible. The results show that statistical methods such as inference, Residual Analysis, and statistical metrics are all good alternatives and excellent methods for validation of the effectiveness of the Neural Network models. The foremost conclusion is that the resulting redesigned Radial Basis Function improved the accuracy of the model after using a Hybrid Learning Process; moreover, the new model also validates the statistical assumptions for using statistical inference and statistical analysis, satisfying the assumptions required for ANOVA to determine the statistical significance and the relationship between variables.

Generation of a Reconfigurable Logical Cell Using Evolutionary Computation

In nature, an interesting topic is about how a cell can be reconfigured in order to achieve a different task. Another interesting topic is about the learning process that seems to be a trial and error process. In this work, we present mechanisms about how to produce a reconfigurable logical cell based on the tent map. The reconfiguration is realized by modifying its internal parameters generating several logical functions in the same structure. The logical cell is built with three blocks: the initial condition generating function, the tent map, and the output function. Furthermore, we propose a reconfigurable structure based on a chaotic system and an evolutionary algorithm is used in order to tune the parameters of the cell via trial and error process.

Symbolic regression using α- β operators and estimation of distribution algorithms: preliminary results

Modeling processes is an important task in engineering; however, the generation of models using only experimental data is not a straightforward problem. Linear regression, neural networks, and other approaches have been used for this purpose; nevertheless, a mathematical description is desirable specially when an optimization is required. Symbolic regression has been used for generating equations considering only experimental data. In this paper, two new operators are proposed to represent a mathematical model of a process. These operators simplified the way for representing equations making possible its use as a symbolic regression. The correct model is generated selecting the appropriate operators and parameters using an evolutionary algorithm like the estimation of distribution algorithms. As a preliminary results, three cases are used to illustrated the performance of the proposed approach. The results indicates that the use of these α, β operators are a promising way to apply symbolic regression to model complex process.

Positional Synthesis of RRRR Linkage Using Evolutionary Computation

In this work, an algorithm based on evolutionary computation is used to solve positional synthesis for RRRR linkage. The main result allows to synthesize a linkage that satisfy higher amount of design pairs [input,output] than obtained by algebraic methods. The design pairs are important in every mechanic synthesis because specify a desired trajectory of the mechanism; however, algebraic methods avoid to use a high number of pairs. In the proposed approach, more design pairs can be involved generating more equations so more solutions can be considered. Numerical results and simulations are shown in order to illustrate the performance of the proposed approach. Finally, some discussions and conclusions are given.