Gursel Serpen

Performance of global–local hybrid ensemble versus boosting and bagging ensembles

This study compares the classification performance of a hybrid ensemble, which is called the global–local hybrid ensemble that employs both local and global learners against data manipulation ensembles including bagging and boosting variants. A comprehensive simulation study is performed on 46 UCI machine learning repository data sets using prediction accuracy and SAR performance metrics and along with rigorous statistical significance tests. Simulation results for comparison of classification performances indicate that global–local hybrid ensemble outperforms or ties with bagging and boosting ensemble variants in all cases. This suggests that the global–local ensemble has a more robust performance profile since its performance is less sensitive to variation with respect to the problem domain, or equivalently the data sets. This performance robustness is realized at the expense of increased complexity of the global–local ensemble since at least two types of learners, e.g. one global and another one local, must be trained. A complementary diversity analysis of global–local hybrid ensemble and base learners used for bagging and boosting ensembles on select data sets in the classifier projection space provides both an explanation and support for the performance related findings of this study.

Complexity Analysis of Multilayer Perceptron Neural Network Embedded into a Wireless Sensor Network

Abstract This paper presents computational and message complexity analysis for a multi-layer perceptron neural network, which is implemented in fully distributed and parallel form across a wireless sensor network. Wireless sensor networks offer a promising platform for parallel and distributed neurocomputing as well as potentially benefiting from artificial neural networks for enhancing their adaptation abilities and computational intelligence. Multilayer perceptron (MLP) neural networks are generic function approximators and classifiers with countless domain-specific applications as reported in the literature. Accordingly, embedding a multilayer perceptron neural network in a wireless sensor network in parallel and distributed mode offers synergy and is very promising. Accordingly, assessing the computational and communication complexity of such hybrid designs, namely an artificial neural network such as a multilayer perceptron network embedded within a wireless sensor network, of interest. This paper presents bounds and results of empirical study on the time, space and message complexity aspects of a wireless sensor network and multilayer perceptron neural network design.

Automated robotic parking systems: real-time, concurrent and multi-robot path planning in dynamic environments

This paper presents an integrated framework for a suite of dynamic path planning algorithms for real time and concurrent movement of multiple robotic carts across a parking garage floor layout without driving lanes. Planning and search algorithms were formulated to address major aspects of automation for storage and retrieval of cars loaded onto robotic mobile carts. Path planning algorithms including A*, D* Lite and Uniform Cost Search were implemented and integrated within a unified framework to guide the robotic carts from a starting point to their destination during the storage and the retrieval processes. For the circumstances where there is no clear path for a car being stored or retrieved, a procedure was developed to unblock the obstacles in the path. A policy that minimizes obstructions was defined for assigning the parking spots on a given floor for arriving cars. Performance evaluation of the overall proposed system was done using a multithreaded software application. A variety of rectangular parking lot layouts including those with 20 × 20, 20 × 40, 30 × 40, and 40 × 40 parking spaces were considered in the simulation study. Performance metrics of path length, planning or search time and memory space requirements were monitored. Simulation results demonstrate that the proposed design facilitates near optimal paths, and is able to handle tens of concurrent requests in real time and in the presence of immobilized carts.

Transformation based Score Fusion Algorithm for Multi-modal Biometric User Authentication through Ensemble Classification

Abstract This paper presents a simulation study for a fusion or combiner algorithm for an ensemble classifier, which facilitates multi-modal biometric user authentication. The study employs the transformation-based score fusion method with the voice and face recognition classifier outputs as inputs. This fusion technique is only dependent on the score generated from each biometric module. The simulation study was conceived to show the feasibility, utility and application of the transformation-based score fusion algorithm. The simulation scenario entails users registering, training, and authenticating through their voice recordings and face images. Performance of the authentication system based on the proposed fusion algorithm indicated that the true positive rate is 99.15%, and the true negative rate is 99.28%. Simulation results suggest that the proposed authentication system is feasible and its performance is promising for real-life applications.

A design heuristic for hybrid classification ensembles in machine learning

This paper presents a new design heuristic for hybrid classifier ensembles in machine learning. The heuristic entails inclusion of both global and local learners in the composition of base classifiers of a hybrid classifier ensemble, while also inducing both heterogeneous and homogenous diversity through the co-existence of global and local learners. Realization of the proposed heuristic is demonstrated within a hybrid ensemble classifier framework. The utility of proposed heuristic for enhancing the hybrid classifier ensemble performance is assessed and evaluated through a simulation study. Weka machine learning tool bench along with 46 datasets from the UCI machine learning repository are used. Simulation results indicate that the proposed heuristic enhances the performance of a hybrid classification ensemble.

Parallel and distributed neurocomputing with wireless sensor networks

This paper proposes a novel hardware computing platform for fully parallel and distributed computation of artificial neural network (ANN) algorithms. The proposed idea entails leveraging the existing wireless sensor networks (WSN) technology to serve as a parallel and distributed hardware platform to implement computations for artificial neural network algorithms. Feasibility of the proposed neurocomputing architecture is demonstrated through a simulation-based case study, which uses Kohonens self-organizing map as the neural network algorithm. MATLAB-based PROWLER, which is a protocol and application level simulator for wireless sensor networks, is employed for the simulation study. Findings demonstrated that the proposed neurocomputing architecture was able to train the SOM neural network with competitive accuracy values for the unsupervised clustering task. Conclusions of the simulation study suggest that the WSN-based neurocomputing architecture is a feasible alternative for realizing parallel and distributed computation of artificial neural network algorithms.

Real-Time Optimal Scheduling of a Group of Elevators in a Multi-Story Robotic Fully-Automated Parking Structure

Abstract This study presents a simulation-based feasibility study for development of a real-time scheduling algorithm for a multi-story and fully-automated parking structure with a group of elevators. Each elevator is conceived to carry one vehicle (car, small truck, SUV or minivan) between floors. Elevator count for a specific parking structure with number of floors in the range of 4 to 20, and 400 parking spaces on each floor is derived under an assumed customer arrival rate and mean service rate using the waiting line model of the queuing theory. A scheduling algorithm based on nested partitions and genetic algorithm is evaluated through the simulation study. The simulation environment models the mean arrival time of customers and elevator dynamics during morning rush hours for busy urban commercial multi-storied parking structures. Performance evaluation of the implemented elevator scheduling system was realized using the MATLAB environment. Performance metrics of mean customer waiting and elevator service times, and maximum customer waiting time were monitored. Simulation results demonstrate that the proposed design facilitates acceptable customer waiting and service times with good utilization rates for the elevators.

Hybrid random subsample classifier ensemble for high dimensional data sets

This paper presents a comparative performance evaluation of a random subsample classifier ensemble with leading machine learning classifiers on high dimensional datasets. Classification performance of the hybrid random subsample ensemble is compared to those of a comprehensive set of machine learning classification algorithms through both in-house simulations and the results published by others in the literature. Performance comparison is based on prediction accuracies on six datasets from the UCI Machine Learning repository, namely Dexter, Madelon, Isolet, Multiple Features, Internet Ads, and Citeseer, with feature counts of up to 105,000. Simulation results establish the competitive performance aspect of the hybrid random subsample ensemble for high dimensional datasets. Specifically, the study findings indicate that hybrid random subsample ensembles with a subsample rate of 15% and base classifier count of 25 or more can achieve a very competitive performance on high dimensional data sets when compared to leading machine learning classifier algorithms.

Exploiting mid-range DNA patterns for sequence classification: binary abstraction Markov models

Messenger RNA sequences possess specific nucleotide patterns distinguishing them from non-coding genomic sequences. In this study, we explore the utilization of modified Markov models to analyze sequences up to 44 bp, far beyond the 8-bp limit of conventional Markov models, for exon/intron discrimination. In order to analyze nucleotide sequences of this length, their information content is first reduced by conversion into shorter binary patterns via the application of numerous abstraction schemes. After the conversion of genomic sequences to binary strings, homogenous Markov models trained on the binary sequences are used to discriminate between exons and introns. We term this approach the Binary Abstraction Markov Model (BAMM). High-quality abstraction schemes for exon/intron discrimination are selected using optimization algorithms on supercomputers. The best MM classifiers are then combined using support vector machines into a single classifier. With this approach, over 95% classification accuracy is achieved without taking reading frame into account. With further development, the BAMM approach can be applied to sequences lacking the genetic code such as ncRNAs and 5′-untranslated regions.

Empirical Model Development for Message Delay and Drop in Wireless Sensor Networks.

Abstract Simulation of wireless sensor networks with very large number of motes poses significant challenges with respect to computational complexity. Application level code prototyping with reasonable accuracy and fidelity can however be accomplished through simulation that models only the effects of the wireless and distributed computations which materialize as delay and drop for the messages being exchanged among the motes. This study pursues that idea of empirical modelling of delay and drop and employs such a model to affect the reception times of wirelessly communicated messages. The delay and drop is therefore, modelled as random variables with probability distributions empirically approximated based on the data reported in the literature. The paper concludes with a case study that employs the proposed empirical delay and drop models for multilayer perceptron neural networks distributed across a wireless sensor network for a classification task on the Isolet dataset.