Search

Multi-agent pathfinding (MAPF) is the problem of finding $k$ non-colliding paths connecting $k$ given initial positions with $k$ given goal positions on a given map. In its sum-of-costs variant, the total number of moves and wait actions performed by agents before they definitely reach the goal is minimized. Not surprisingly, since MAPF is combinatorial, a number of compilations to Boolean Satisfiability (SAT) and Answer Set Programming (ASP) exist. In this article, we describe in detail the first family of compilations to ASP that solve sum-of-costs MAPF over 4-connected grids. Compared to existing ASP compilations, a distinguishing feature of our compilation is that the number of total clauses (after grounding) grow linearly with the number of agents, while existing compilations grow quadratically. In addition, the optimization objective is such that its size after grounding does not depend on the size of the grid. In our experimental evaluation, we show that our approach outperforms search-based sum-of-costs MAPF solvers when grids are congested with agents. We also show that our approach is competitive with a SAT-based approach when follow conflicts are taken into account. We also explore the potential of our solver when finding makespan-optimal solutions, in which makespan is minimized first and then cost is minimized. Our results show that makespan-optimal solutions are slightly suboptimal in most benchmarks. Moreover, our MAPF solver, when run in that mode, is faster and scales better.

We present a solution based on deep reinforcement learning (DRL) that jointly addresses spectrum allocation and latency constraint in EONs. The results show that using a simple network representation, this strategy outperforms typical K-Shortest Path heuristic approach and previous DRLbased approaches., The research leading to these results has received funding from MINECO Project AURORAS (RTI2018-099178) and Spanish Thematic Network Go2Edge (RED2018-102585-T).

Known attempts to build autonomous robots rely on complex control architectures, often implemented with the Robot Operating System platform (ROS). The implementation of adaptable architectures is very often ad hoc, quickly gets cumbersome and expensive. Reusable solutions that support complex, runtime reasoning for robot adaptation have been seen in the adoption of ontologies. While the usage of ontologies significantly increases system reuse and maintainability, it requires additional effort from the application developers to translate requirements into formal rules that can be used by an ontological reasoner. In this paper, we present a design tool that facilitates the specification of reconfigurable robot skills. Based on the specified skills, we generate corresponding runtime models for self-adaptation that can be directly deployed to a running robot that uses a reasoning approach based on ontologies. We demonstrate the applicability of the tool in a real robot performing a patrolling mission at a university campus.

Search is a universal problem-solving method in Artificial Intelligence. Specifically, Heuristic Search algorithms, such as A*, use a heuristic function to guide the search process. The heuristic function allows algorithms to explore only a part of the search space, resulting in an efficient search process. This paper presents a new heuristic function to solve the Generalized Covering Traveling Salesman Problem (GCTSP). The heuristic function is precalculated. The method to obtain the function is pre-calculating optimal solutions consecutively to small sub-problems of the GCTSP of increasing difficulty, using an incremental Best First Search algorithm, which reuses heuristics values previously precalculated. The resultating heuristic function can be used with different heuristic search algorithms. To the best of our knowledge, this problem has not been solved with Heuristic Search. This paper is the first to present a solution to the GCTSP using Heuristic Search algorithms, such as A* and Anytime search algorithms. We evaluated different Heuristic Search algorithms. The experimental evaluation shows results of the same quality, obtained orders-of-magnitude faster than the exact methods traditionally used in Operations Research.

Path planning has been considered as key part for expanding the intelligent properties because it faces the problem of finding collision-free motions for robot systems from one configuration to another. Resistive Grid Path Planning Methodology (RGPPM) is a novel and reduced CPU time modelling and simulation procedure for robotic motions planning. In this paper, robotical implementations of collision-free path planning using an emerging grid-based methodology is carried out by embedded systems applications and robot application framework.

Memristor and memristive circuits exhibit a hysteresis v-i characteristic, which is strongly dependent on the state variable function. In this way, controlling the state variable performance can give a more realistic description of the device. This work presents the use of a PWL window function to establish the behavior of the state variable function with the purpose of analyzing the influence of the novel window function in the characteristic curves of the memristive systems, such as: the pinched hysteresis loop and the memristance. The notion of asymmetrical window functions is introduced in order to catch sight of the advantages of the PWL window functions compared to the typical window functions. Moreover, a simple comparison regarding several window functions is achieved.

Multi-agent pathfinding, namely finding collision-free paths for several agents from their given start locations to their given goal locations on a known stationary map, is an important task for non-player characters in video games. A variety of heuristic search algorithms have been developed for this task. Non-real-time algorithms, such as Flow Annotated Replanning (FAR), first find complete paths for all agents and then move the agents along these paths. However, their searches are often too expensive. Real-time algorithms have the ability to produce the next moves for all agents without finding complete paths for them and thus allow the agents to move in real time. Real-time heuristic search algorithms have so far typically been developed for single-agent pathfinding. We, on the other hand, present a real-time heuristic search algorithm for multi-agent pathfinding, called Bounded Multi-Agent A* (BMAA*), that works as follows: Every agent runs an individual real-time heuristic search that updates heuristic values assigned to locations and treats the other agents as (moving) obstacles. Agents do not coordinate with each other, in particular, they neither share their paths nor heuristic values. We show how BMAA* can be enhanced by adding FAR-style flow annotations and allowing agents to push other agents temporarily off their goal locations, when necessary. In our experiments, BMAA* has higher completion rates and lower completion times than FAR.

In this paper, we propose a new selection scheme for Multi-Objective Evolutionary Algorithms (MOEAs) based on the Δ ρ indicator. Our new selection scheme is incorporated into a MOEA giving rise to the “Δρ-MOEA.” Perhaps, one of the most important disadvantages of MOEAs based on Δ ρ is the definition of the reference set. In this work, we propose to create a reference set at each generation using e-dominance and the set of nondominated solutions found so far. Our new selection scheme uses two different techniques to select solutions according to the modified generational distance indicator or the modified inverted generational distance indicator. Our proposed Δ p -MOEA is validated using standard test functions taken from the specialized literature, having three to six objective functions and it is compared with respect to two well-known MOEAs: MOEA/D using Penalty Boundary Intersection (PBI), which is based on decomposition, and SMS-EMOA-HYPE (a version of SMS-EMOA that uses a fitness assignment scheme based on the use of an approximation of the hypervolume indicator).

Given a stereo pair it is possible to recover a depth map and use that depth to render a synthetically defocused image. Though stereo algorithms are well-studied, rarely are those algorithms considered solely in the context of producing these defocused renderings. In this paper we present a technique for efficiently producing disparity maps using a novel optimization framework in which inference is performed in “bilateral-space”. Our approach produces higher-quality “defocus” results than other stereo algorithms while also being 10 – 100× faster than comparable techniques.

While prior depth from focus and defocus techniques operated on laboratory scenes, we introduce the first depth from focus (DfF) method capable of handling images from mobile phones and other hand-held cameras. Achieving this goal requires solving a novel uncalibrated DfF problem and aligning the frames to account for scene parallax. Our approach is demonstrated on a range of challenging cases and produces high quality results.

We address the problem of geo-registering ground-based multi-view stereo models by ground-to-aerial image matching. The main contribution is a fully automated geo-registration pipeline with a novel viewpoint-dependent matching method that handles ground to aerial viewpoint variation. We conduct large-scale experiments which consist of many popular outdoor landmarks in Rome. The proposed approach demonstrates a high success rate for the task, and dramatically outperforms state-of-the-art techniques, yielding geo-registration at pixel-level accuracy.

This paper leverages occluding contours (aka "internal silhouettes") to improve the performance of multi-view stereo methods. The contributions are 1) a new technique to identify free-space regions arising from occluding contours, and 2) a new approach for incorporating the resulting free-space constraints into Poisson surface reconstruction. The proposed approach outperforms state of the art MVS techniques for challenging Internet datasets, yielding dramatic quality improvements both around object contours and in surface detail.

Since the outcoming of the memristor, memristive systems and mem-elements in electronics, new features for analog and digital circuit design have been introduced, and as a result models for the memristor are strongly needed in order to incorporate the device to the design flow loop. In this paper, a model for the memristor is introduced, it is generated by solving the differential equation, that governs the physical functioning of the device, by using a homotopy formulation. The generated model is recast in fully symbolic form that can be used to carry out behavioral simulation of circuits containing memristors.

This paper describes an effort to automatically create ``tours'' of thousands of the world's landmarks from geo-tagged user-contributed photos on the Internet. These photo tours take you through each site's most popular viewpoints on a tour that maximizes visual quality and traversal efficiency. This planning problem is framed as a form of the Traveling Salesman Problem on a graph with photos as nodes and transition costs on edges and pairs of edges, permitting efficient solution even for large graphs containing thousands of photos. Our approach is highly scalable and is the basis for the Photo Tours feature in Google Maps, which can be viewed at http://maps.google.com/phototours.

We present a video-based system which interactively captures the geometry of a 3D object in the form of a point cloud, then recognizes and registers known objects in this point cloud in a matter of seconds (fig. 1). In order to achieve interactive speed, we exploit both efficient inference algorithms and parallel computation, often on a GPU. The system can be broken down into two distinct phases: geometry capture, and object inference. We now discuss these in further detail.

This paper addresses the problem of automatically obtaining the object/background segmentation of a rigid 3D object observed in a set of images that have been calibrated for camera pose and intrinsics. Such segmentations can be used to obtain a shape representation of a potentially texture-less object by computing a visual hull. We propose an automatic approach where the object to be segmented is identified by the pose of the cameras instead of user input such as 2D bounding rectangles or brush-strokes. The key behind our method is a pairwise MRF framework that combines (a) foreground/background appearance models, (b) epipolar constraints and (c) weak stereo correspondence into a single segmentation cost function that can be efficiently solved by Graph-cuts. The segmentation thus obtained is further improved using silhouette coherency and then used to update the foreground/background appearance models which are fed into the next Graph-cut computation. These two steps are iterated until segmentation convergences. Our method can automatically provide a 3D surface representation even in texture-less scenes where MVS methods might fail. Furthermore, it confers improved performance in images where the object is not readily separable from the background in colour space, an area that previous segmentation approaches have found challenging.

Real-time search is a standard approach to solving search problems in which agents have limited sensing capabilities and must act quickly. It is well known that real-time search algorithms like LRTA* and RTA* perform poorly in regions of the search space in which the heuristic function is very imprecise. Approaches that use look ahead or learning are used to overcome this drawback. They perform more computation in the planning phase compared to LRTA* andRTA* . In this paper we propose Path Real-Time A* (PRTA* ), an algorithm that, like LRTA*, performs little computation in the planning phase, but that, unlike LRTA*, terminates even if the problem does not have a solution. We show that our algorithm outperforms LRTA* and RTA* in standard real-time benchmark problems. Furthermore, we show that in some cases, PRTA* may also outperform lookahead-orlearning-enabled algorithms but carrying out significantly less computation.

This paper describes a work in progress on an ontology-based approach for autonomous system engineering. This is part of the ASys research programme which focuses on the development of a technology for autonomous systems product-line engineering. Such technology will comprehend from platform technology to designs of intelligent autonomous systems and support methodologies. Within this programme, a domain-ontology for autonomous systems (OASys) has been developed to describe the autonomous system's domain. OASys will be used as core software asset for the analysis, design, implementation, and run-time operation of any autonomous system. Two testbeds are being developed to test OASys's applicability: a mobile robot-based application, and the control system of a chemical reactor.

As shown in the bibliography, training an ensemble of networks is an interesting way to improve the performance with respect to a single network. The two key factors to design an ensemble are how to train the individual networks and how to combine them to give a single output. Boosting is a well known methodology to build an ensemble. Some boosting methods use an specific combiner (Boosting Combiner) based on the accuracy of the network. Although the Boosting combiner provides good results on boosting ensembles, the simple combiner Output Average worked better in three new boosting methods we successfully proposed in previouses papers. In this paper, we study the performance of sixteen different combination methods for ensembles previously trained with Adaptive Boosting and Average Boosting in order to see which combiner fits better on these ensembles. Finally, the results show that the accuracy of the ensembles trained with these original boosting methods can be improved by using the appropriate alternative combiner. In fact, the Output average and the Weighted average on low/medium sized ensembles provide the best results in most of the cases.

In previous researches we have analysed some methods to create committees of multilayer feedforward networks trained with the backpropagation algorithm. One of the most known methods that we have studied is Adaptive Boosting. In this paper we propose a variation of this method called weighted conservative boosting based on conservative boosting. In this case, a weight which depends on the database and on the ensemble is added to the equation used to update the sampling distribution. We have tested adaptive boosting, conservative boosting and weighted conservative boosting with seven databases from the UCI repository. We have used the mean Increase of Performance and the mean percentage of error reduction to compare both methods, the results show that weighted conservative boosting is the best performing method.

Adaptive boosting (Adaboost) is one of the most known methods to build an ensemble of neural networks. Adaboost has been studied and successfully improved by some authors like Breiman, Kuncheva or Oza. In this paper we briefly analyze and mix two of the most important variants of Adaboost in order to build a robuster ensemble of neural networks. The boosting methods we have studied are averaged boosting (Aveboost) and conservative boosting (Conserboost). We proposed the mixed method we have called averaged onservative boosting (ACE). In this method we apply the conservative equation used in Conserboost along with the averaged procedure used in Aveboost in order to update the sampling distrubution of Adaboost. We have tested the methods with seven databases from the UCI repository. We have used the mean increase of performance and the mean percentage of error reduction to compare both methods, the results show that the new proposed method performs better.

We present a new formulation to multi-view stereo that treats the problem as probabilistic 3D segmentation. Previous work has used the stereo photo-consistency criterion as a detector of the boundary between the 3D scene and the surrounding empty space. Here we show how the same criterion can also provide a foreground/background model that can predict if a 3D location is inside or outside the scene. This model replaces the commonly used naive foreground model based on ballooning which is known to perform poorly in concavities. We demonstrate how the probabilistic visibility is linked to previous work on depth-map fusion and we present a multi-resolution graph-cut implementation using the new ballooning term that is very efficient both in terms of computation time and memory requirements.

We present an algorithm and the associated capture methodology to acquire and track the detailed 3D shape, bends, and wrinkles of deforming surfaces. Moving 3D data has been difficult to obtain by methods that rely on known surface features, structured light, or silhouettes. Multispec- tral photometric stereo is an attractive alternative because it can recover a dense normal field from an un-textured surface. We show how to capture such data and register it over time to generate a single deforming surface. Experiments were performed on video sequences of un- textured cloth, filmed under spatially separated red, green, and blue light sources. Our first finding is that using zero- depth-silhouettes as the initial boundary condition already produces rather smoothly varying per-frame reconstructions with high detail. Second, when these 3D reconstructions are augmented with 2D optical flow, one can register the first frame's reconstruction to every subsequent frame.

There are two different ways to create a Multiple Classification System based on neural networks. The first one is the Ensemble approach; it consists on combining the outputs of different networks which solve the same problem in a suitable manner to give a single output. The second one is the Modular approach; it consists on decomposing the problem into subproblems, the final decision is taken with the information provided by the networks. One of the most known methods to build a Modular Neural Network is the Mixture of Neural Networks. In this paper we present a Mixture of Multilayer Feedforward Networks a modular method based on Multilayer Feedforward networks. Finally, we have included a comparison among Simple Ensemble, Mixture of Neural Networks and Mixture of Multilayer Feedforward Networks. We have tested the methods with eight databases from the UCI repository and the results show that Mixture of Multilayer Feedforward Networks is the best performing method.

This paper addresses the problem of obtaining complete, detailed reconstructions of shiny textureless objects. We present an algorithm which uses silhouettes of the object, as well as images obtained under varying illumination conditions. In contrast with previous photometric stereo techniques, ours is not limited to a single viewpoint and produces accurate reconstructions in full 3D. A number of images of the object are obtained from multiple viewpoints, under varying lighting conditions. Starting from the silhouettes, the algorithm recovers camera motion and constructs the object’s visual hull. This is then used to recover the illumination and initialise a multi-view photometric stereo scheme to obtain a closed surface reconstruction. The contributions of the paper are twofold: Firstly we describe a robust technique to estimate light directions and intensities and secondly, we introduce a novel formulation of photometric stereo which combines multiple viewpoints and hence allows closed surface reconstructions. The algorithm has been implemented as a practical model acquisition system. Here, a quantitative evaluation of the algorithm on synthetic data is presented together with a complete reconstruction of a challenging real object.

We enhance real-time search algorithms with bounded propagation of heuristic changes. When the heuristic of the current state is updated, this change is propagated consistently up to k states not necessarily distinct. Applying this idea to FALCONS, we have develop the new FALCONS(k), an algorithm that keeps the good theoretical properties of FALCONS and improves its performance. We provide experimental results on benchmarks for real-time search, showing the benefits of our approach.

Building an ensemble of classifiers is an useful way to improve the performance. In the case of neural networks the bibliography has centered on the use of multilayer feedforward (MF). However, there are other interesting networks like radial basis functions (RBF) that can be used as elements of the ensemble. In a previous paper we presented results of different methods to build the ensemble of RBF. The results showed that the best method is in general the simple ensemble. The combination methods used in that research was averaging. In this paper we present results of fourteen different combination methods for a simple ensemble of RBF. The best performing methods are Borda count, weighted average and majority voting.

As shown in the bibliography, training an ensemble of networks is an interesting way to improve the performance with respect to a single network. However there are several methods to construct the ensemble. In this paper we present some new results in a comparison of twenty different methods. We have trained ensembles of 3, 9, 20 and 40 networks to show results in a wide spectrum of values. The results show that the improvement in performance above 9 networks in the ensemble depends on the method but it is usually low. Also, the best method for an ensemble of 3 networks is called "decorrelated" and uses a penalty term in the usual backpropagation function to decorrelate the network outputs in the ensemble. For the case of 9 and 20 networks the best method is conservative boosting. And finally for 40 networks the best method is Cels.

As shown in the bibliography, training an ensemble of networks is an interesting way to improve the performance with respect to a single network. The two key factors to design an ensemble are how to train the individual networks and how to combine the different outputs of the networks to give a single output class. In this paper, we focus on the combination methods. We study the performance of fourteen different combination methods for ensembles of the type "simple ensemble" and "decorrelated". In the case of the "simple ensemble" and low number of networks in the ensemble, the method Zimmermann gets the best performance. When the number of networks is in the range of 9 and 20 the weighted average is the best alternative. Finally, in the case of the ensemble "decorrelated" the best performing method is averaging over a wide spectrum of the number of networks in the ensemble.

In previous researches we have analysed some methods to create committees of multilayer feedforward networks trained with the Back Propagation algorithm. One of the most known methods that we have studied is Adaptive Boosting. In this paper we present a variation of this method called Crossboost. In this version of AdaBoost, we have used k-cross-fold validation over the whole learning set to generate an specific training set and validation set for each network of the committee. In the new method, the data set used to train the ith-network is selectively sampled from its specific training set, the sampling distribution is calculated over the whole learning set. The diversity of a committee generated with our method increases respect the original method because each network has its specific validation set. We have tested Adaboost and Crossboost with ten databases from the UCI repository. We have used the mean percentage of error reduction to compare both methods, the results show that Crossboost performs better.

A hyperspectral image is used in remote sensing to identify different type of coverts on the Earth surface. It is composed of pixels and each pixel consists of spectral bands of the electromagnetic reflected spectrum. Neural networks and ensemble techniques have been applied to remote sensing images with a low number of spectral bands per pixel (less than 20). In this paper, we apply different ensemble methods of multilayer feedforward networks to images of 224 spectral bands per pixel, where the classification problem is clearly different. We conclude that in general, there is an improvement by the use of an ensemble. For databases with low number of classes and pixels, the improvement is lower and similar for all ensemble methods. However, for databases with a high number of classes and pixels, the improvement depends strongly on the ensemble method. We also present results of the classification of support vector machines (SVM) and see that a neural network is a useful alternative to SVM.

Training an ensemble of networks is an interesting way to improve the performance with respect to a single network. However there are several methods to construct the ensemble and there are no complete results showing which one could be the most appropriate. In this paper we present a comparison of eleven different methods. We have trained ensembles of a reduced number of networks (3 and 9) because in this case the computational cost is not high and the method is suitable for applications. The results show that the improvement in performance from three to nine networks is marginal. Also, the best method is called "Decorrelated" and uses a penalty term in the usual backpropagation function to decorrelate the network outputs in the ensemble.

In this paper, we present a new approach to high quality 3D object reconstruction. Starting from a calibrated sequence of color images, the algorithm is able to reconstruct both the 3D geometry and the texture. The core of the method is based on a deformable model, which defines the framework where texture and silhouette information can be fused. This is achieved by defining two external forces based on the images: a texture driven force and a silhouette driven force. The texture force is computed in two steps: a multi-stereo correlation voting approach and a gradient vector flow diffusion. Due to the high resolution of the voting approach, a multi-grid version of the gradient vector flow has been developed. Concerning the silhouette force, a new formulation of the silhouette constraint is derived. It provides a robust way to integrate the silhouettes in the evolution algorithm. As a consequence, we are able to recover the contour generators of the model at the end of the iteration process. Finally, a texture map is computed from the original images for the reconstructed 3D model.

In this paper we propose a new algorithm for rule extraction from a trained Multilayer Feedforward network. The algorithm is based on an interval arithmetic network inversion for particular target outputs. The types of rules extracted are N-dimensional intervals in the input space. We have performed experiments with four databases and the results are very interesting. One rule extracted by the algorithm can cover 86% of the neural network output and in other cases sixty four rules cover 100% of the neural network output.

The usual way to determine the number of hidden units in multilayer feedforward networks is by a trial and error procedure. An alternative is to use constructive algorithms which change the neural network structure during training. In this paper we propose a new constructive algorithm, explore several design alternatives and conclude that this new constructive algorithm is a better alternative than the procedure of trial and error for designing a neural network.

We present a review of feature selection methods based on an analysis of the training set. The focus is on the methods which have been applied to neural networks. We also present a methodology that allows evaluation and comparison of feature selection methods. This methodology is applied to the 7 reviewed methods in a total of 15 different real world classification problems. The result is an ordering of methods according to performance. From this ordering it is clearly concluded which method is the best and should be used. The best methods are based on information theory concepts like gd-distance and mutual information. We also discuss the applicability and computational complexity of the methods.

We present an experimental comparison of the generalization capability of one and two hidden layers multilayer feedforward neural networks. We used sixteen different real world problems in order to measure the generalization of both architectures. For each problem and architecture we carefully selected, by a trial and error procedure, the minimal network which solves the problem. Several runs with different initial conditions were obtained in order to get an average performance with an error. According to our results, the generalization capability of a one hidden layer network is better than the one of two hidden layers network. Furthermore, two hidden layer networks are more prone to fall into bad local minimum.

We review feature selection methods based on the analysis of a trained multilayer feedforward neural network. Furthermore, we present a methodology that allows experimentally evaluating and comparing feature selection methods. This methodology was applied to the 19 reviewed methods and we evaluated the usefulness of these methods for selecting the appropriate features in the case of using a multilayer feedforward as a pattern recognition method. We used a total number of 15 different real world classification problems in our experiments. From the result of the comparison, we conclude which methods perform better and should be used, and discuss their applicability.

Two generalizations of multilayer feedforward and backpropagation to interval arithmetic were proposed several years ago. These generalizations have several applications like the codification of rules in the training set, "don't care attributes", missing inputs, etc. There are two generalizations what means that there are two different training algorithms and there is no way, at first, to consider one of them better than the other. We present an in depth comparison between the two training algorithms. We have used a total number of 75 different problems and five different performance definitions for the comparison. The results are that the performance of the algorithms depends on the performance definition, but we can consider one of them as the "usual" one and in this sense one algorithm outperforms the other.

We present a review of feature selection methods, based on the analysis of a trained multilayer feedforward network, which have been applied to neural networks. Furthermore, a methodology that allows evaluating and comparing feature selection methods is carefully described. This methodology is applied to the 19 reviewed methods in a total of 15 different real world classification problems. We present an ordination of methods according to their performance and it is clearly concluded which method performs better and should be used. We also discuss the applicability and computational complexity of the methods.

We present the results of an experimental comparison among seven different weight initialization methods in twelve different problems. The comparison is performed by measuring the speed of convergence, the generalization capability and the probability of successful convergence. It is not usual to find an evaluation of the three properties in the literature on weight initialization. The training algorithm was backpropagation with a hyperbolic tangent transfer function. We found that the performance can be improved with respect to the usual initialization scheme.

We present an empirical comparison among four different schemes of coding the outputs of a multilayer feedforward networks. Results are obtained for eight different classification problems from the UCI repository machine learning databases. Our results show that the usual codification is superior to the rest in the case of using one output unit per class. However, if we use several output units per class we can obtain an improvement in the generalization performance depending on the problem and in this case the noisy codification seems to be more appropriate.

We present the results of a research on the combination of weight-decay and input selection methods based on the analysis of a trained multilayer feedforward network. This combination has been proposed and suggested by some other authors. The influence of weight-decay in seventeen different input selection methods is empirically analyzes with a total of eight classification problems. We show that the performance variation by introducing weight-decay strongly depends on the particular input selection method. The use of weight-decay can even deteriorate the efficiency of a method. Furthermore, it seems that weight-decay improves the performance of the worst input selection methods and deteriorate the performance of the best ones. In that sense, it diminishes the performance differences among different methods. We conclude that the combination of weight-decay and this type of input selection methods should be avoided.

We present a neural network application to the diagnosis of vocal and voice disorders, these disorders should be diagnosed in the early stage and normally cause changes in the voice signal. So we use acoustic parameters extracted from the voice as inputs for the neural network. In this paper, we focus our application on the distinction between pathologic and nonpathologic voices. The performance of the neural network is very good, 100% percent correct in the test. Furthermore, we have used neural network techniques to reduce the initial number of inputs (35), we conclude that only two acoustic parameters are needed for the classification between normal and pathological voices. The application can be a very useful diagnostic tool because it is noninvasive, makes possible to develop an automatic computer-based diagnosis system, is objective and can also be useful for evaluation of surgical, pharmacological and rehabilitation processes. Finally, we discuss the limitation of our work and possible future research.

In this paper we present the results of a comparison among six different weight initialization methods with two training algorithms and six databases. The comparison is performed by measuring the three following aspects: speed of convergence, generalization and probability of convergence. The two training algorithms are Backpropagation (BP) and another one that uses conjugate gradient and dynamical learning rate adaptation (NE). We found the best weight initialization scheme for the (BP) algorithm. The speed of convergence can be improved with respect to the usual initialization, but the two other aspects are similar. For the NE algorithm it is concluded that its performance depends on the initialization much more than BP. Its generalization and probability of convergence can be considered lower than BP and the different weight initialization schemes could not improve this drawback. On the other hand it is faster.