Catálogo de publicaciones - libros

Hoeffding trees are state-of-the-art for processing high-speed data streams. Their ingenuity stems from updating sufficient statistics, only addressing growth when decisions can be made that are guaranteed to be almost identical to those that would be made by conventional batch learning methods. Despite this guarantee, decisions are still subject to limited lookahead and stability issues. In this paper we explore Hoeffding Option Trees, a regular Hoeffding tree containing additional nodes that allow several tests to be applied, leading to multiple Hoeffding trees as separate paths. We show how to control tree growth in order to generate a mixture of paths, and empirically determine a reasonable number of paths. We then empirically evaluate a spectrum of Hoeffding tree variations: single trees, option trees and bagged trees. Finally, we investigate pruning. We show that on some datasets a pruned option tree can be smaller and more accurate than a single tree.

Feedforward neural networks are particularly useful in learning a training dataset without prior knowledge. However, weight adjusting with a gradient descent may result in the local minimum problem. Repeated training with random starting weights is among the popular methods to avoid this problem, but it requires extensive computational time. This paper proposes a simultaneous training method with removal criteria to eliminate less promising neural networks, which can decrease the probability of achieving a local minimum while efficiently utilizing resources. The experimental results demonstrate the effectiveness and efficiency of the proposed training method in comparison with conventional training.

There are numerous families of Neural Networks (NN) used in the study and development of the field of Artificial Intelligence (AI). One of the more recent NNs involves the Bursting neuron, pioneered by Rulkov. The latter has the desirable property that it can also be used to model a system (for example, the “brain”) which switches between modes in which the activity is excessive (“bursty”), to the alternate case when the system is “dormant”. This paper, which we believe is a of pioneering sort, derives some of the analytic properties of the Bursting neuron, and the associated NN.

To be more specific, the model proposed by Rulkov [11] explains the so-called “bursting” phenomenon in the system (brain), in which a low frequency pulse output serves as an envelope of high frequency spikes. Although various models for bursting have been proposed, Rulkov’s model seems to be the one that is both analytically tractable and experimentally meaningful. In this paper, we show that a “small” scale network consisting of Bursting neurons rapidly converges to a synchronized behavior implying that increasing the number of neurons does not contribute significantly to the synchronization of the individual Bursting neurons. The consequences of such a conclusion lead to a phenomenon that we call “”.

This paper sumarises a comparative study of multiple neural networks as applied for the identification of the dynamics of an Unmanned Aerial Vehicle (UAV). Each of the networks are based on nonlinear autoregressive technique and are trained online. Variations in the architecture, batch size and the initial weights of the multi-network are analysed. A dynamic selection mechanism optimally chooses the most suitable output from the host of networks based on a selection criteria.

During polysomnography, multivariate physiological measurements are recorded, and analysed to identify episodes of breathing disorders occur during patients sleep for the diagnosis of sleep disordered breathing disorders. Measurement distortions, such as signal losses that may occur due to loosening of a sensor, are often present in these measurements. Reliability and accuracy of automated diagnostic procedures using polysomnographic data can be improved through automated identification and recovery of such measurement distortions. In this study is an attempt towards that focusing on the respiratory measurements. Respiratory measurements are a main criterion in assessing sleep disordered breathing episodes. Treating respiratory system as a deterministic dynamic system, functional mapping that exists between two state space embeddings are approximated using artificial neural networks. Performance of the trained neural networks in identification of measurement distortions and measurement recovery are reported.

Web service composition is the process of integrating existing web services. It is a prospective method to build an application system. The current approaches, however, only take service function aspect into consideration. With the rapid growth of web service applications and the abundance of service providers, the consumer is facing the inevitability of selecting the “maximum satisfied” service providers due to the dynamic nature of web services. This requirement brings us some research challenges including web service quality model, the design of web service framework monitoring service real time quality. The further challenge is to find the algorithm which can handle customized service quality parameters and has good performance to solve NP-hard web services global selection problem. In this paper, we propose an adaptive web service framework using an extensible service quality model. Evolutionary algorithms are adopted to accelerate service global selection. We report on the comparison between Cultural Algorithms with Genetic Algorithms and random service selection.

We propose a genetic programming approach to extraction of glycan motifs by using tag tree patterns, which are tree structured patterns with structured variables. A structured variable in a tag tree pattern can be substituted by an arbitrary tree. Our experiments show that we have obtained tag tree patterns as the best individuals including similar substructures of glycan motifs obtained by the previous works.

This paper describes a new approach to the use of genetic programming (GP) for feature construction in classification problems. Rather than wrapping a particular classifier for single feature construction as in most of the existing methods, this approach uses GP to construct multiple (high-level) features from the original features. These constructed features are then used by decision trees for classification. As feature construction is independent of classification, the fitness function is designed based on the class dispersion and entropy. This approach is examined and compared with the standard decision tree method, using the original features, and using a combination of the original features and constructed features, on 12 benchmark classification problems. The results show that the new approach outperforms the standard way of using decision trees on these problems in terms of the classification performance, dimension reduction and the learned decision tree size.

A dual-population genetic algorithm (DPGA) is a new multi-population genetic algorithm that solves problems using two populations with different evolutionary objectives. The main population is similar to that of an ordinary genetic algorithm, and it evolves in order to obtain suitable solutions. The reserve population evolves to maintain and offer diversity to the main population. The two populations exchange genetic materials using interpopulation crossbreeding. This paper proposes a new fitness function of the reserve population based on the distance to the main populations. The experimental results have shown that the performance of DPGA is highly related to the distance between the populations and that the best distance differs for each problem. Generally, it is difficult to decide the best distance between the populations without prior knowledge about the problem. Therefore, this paper also proposes a method to dynamically adjust the distance between the populations using the distance between good parents, i.e., the parents that generated good offspring.

As an important area in AI, Distributed Constraint Satisfaction Problems (DisCSPs) can be used to model and solve many problems in multi-agent systems. Concurrent search, newly proposed, is an efficient technique for solving DisCSPs. In this paper, a novel concurrent search algorithm is presented. Dynamic Variable Ordering (DVO) is used in concurrent backtrack search instead of random variable ordering. In order to make DVO effective, domain sizes of unfixed variables are evaluated approximately according to current partial assignments after a variable is assigned. This method can be performed by a single agent and there is no need to send messages during heuristic computation. In addition, a simple look-ahead strategy inspired from centralized constraint programming techniques is added to the improved algorithm. Experiments on randomly generated DisCSPs demonstrate that the algorithm with DVO heuristic and look-ahead strategy can drastically improve performance of concurrent search.