International Journal of Computational Intelligence Studies (31 papers in press)
Handling the Crowd Avoidance Problem in Job Recommendation Systems Integrating FoDRA
by Nikolaos Almalis
Abstract: In this article, we present the basic principles and approaches of the Job Recommender Systems (JRSs). Furthermore, we describe the four different relation types of the job seeking and recruiting problem, derived directly from the formal definition of the JRSs. We use our already published Four Dimensions Recommendation Algorithm (FoDRA) to calculate the suitability of person for a job and then we model a job seeking and recruiting problem with many candidates and many jobs (N-N case). Finally, we execute the algorithm and present the results proposing a solution -the minimum acceptable suitability level-for the crowd avoidance problem that occurred. Our study produces satisfying results and shows that this approach can be considered as an important asset in the domain of Job Seeking and Recruiting.
Keywords: Recommendation system; Job seeking and recruiting; Job recommender; Matching people and jobs; Constraint-based; Information filtering.
Creating classification rules using Grammatical Evolution
by Ioannis Tsoulos
Abstract: A genetic programming based method is introduced for data classification. The fundamental element of the method is the well - known technique of Grammatical Evolution. The method constructs classification programs in a C like programming language in order to classify the input data, producing simple if else rules. The paper introduces the method as well as the conducted experiments on a series of datasets against other well known classification methods.
Keywords: Genetic algorithm; Data classification; Grammatical evolution; Stochastic methods.
Special Issue on: Advances and Challenges in Nature Inspired Optimisations
Improving Bug Report Quality by Predicting Correct Component in Bug Reports
by Indu Chawla, sandeep Kumar Singh
Abstract: Bugs reported in bug tracking systems contain important information in the form of standard and mandatory fields like version, operating system, product, component and type of the bug etc. Unfortunately, the information provided in these fields is sometimes missing and inaccurate. Inaccurate information makes the program understanding difficult and inflicts delay in the process of bug fixing. Many times, these fields are reassigned during the bug fixing time even more than ones. This study explores the automatic identification of correct component field in a bug report. This study proposes using fuzzy similarity based approach for identifying the correct component as well as predicting the possibility of component reassignment. Experimentation is done using bug reports of three open source projects from Eclipse. The experimental results show that fuzzy similarity approach performs better as compared to state of art approaches on two out of three datasets in terms of precision score.
Keywords: Bug tracking system, Bug report, component prediction, component reassignment
The Connectivity and the Static-Cost-Effective Analysis of a Shifted Completely Connected Network
by MOHAMMED N. M. ALI, M. M. Hafizur Rahman, Adamu Abubakar Ibrahim, Dhiren K. Behera, Yasushi Inoguchi
Abstract: At the current time, finding an alternative computing device with extreme computation power became the main concern of the research community. Therefore, building a computer device able to execute extremely difficult calculations in a short period of time is required. Presently, the massively parallel computer (MPC) systems considered the highest computing devices, and the existence of these systems is important to execute many operations in many sectors such as engineering and science. These devices built based on an internal network called interconnection network which has a particular design represented by the network topology. The cost of these networks influenced highly by the price of the processing elements (PEs) and the communication links. Thus, the design of these topologies has a crucial impact on the network cost and performance. In this paper, we have proposed a new design of topology for the MPC systems; this topology has been evaluated statically in previous work, and it showed good results. Therefore, in this paper, we will focus on analysing the cost effectivity of this network to examine if it is a cost-beneficial network before going to the implementation step to assure that the profit of this network is deserving the cost.
Keywords: Network-on-Chip (NoC), Interconnection Networks, Hierarchical Interconnection Networks (HINs), Static Network Performance, Shifted Completely Connected Network (SCCN), Massively Parallel Computer (MPC) Systems, Conventional Interconnection Networks.
Nonlinear time series forecasting using a novel self-adaptive TLBO-MFLANN model
by Sibarama Panigrahi, H.S. Behera
Abstract: Time series forecasting (TSF) is a key field of research in several areas of study including engineering, finance, economics and management science. Conventionally, TSF has been predominantly performed using various linear statistical models. However, to cope up with nonlinear patterns exhibiting in most of the time series produced from real world phenomenon, recently, various artificial neural network (ANN) models have been used. Contrasting to traditional neural networks, higher order neural network (HONN) especially the functional link ANN (FLANN) has the capability to expand the input space with fewer trainable weights which makes it efficient to solve various complex problems. Motivated by this, in this paper, we have proposed a multiplicative FLANN (MFLANN) model for time series forecasting. The multiplicative unit in MFLANN model assists to capture the nonlinear patterns well. In addition, an improved version of teaching learning based optimization (TLBO), called self-adaptive TLBO (SATLBO) has been proposed to train the MFLANN model. The proposed SATLBO uses the gradient descent learning algorithm in the teacher phase while uses the past experience to adapt the learners parameters. This unique integration of SATLBO with gradient descent learning algorithm is used to determine the near optimal weight set of MFLANN. The proposed method is implemented in MATLAB environment and the obtained results are compared with other methods (DE based MFLANN, TLBO based MFLANN, CRO based MFLANN, Jaya based MFLANN and ETLBO-JPSNN) considering 11 benchmark univariate time series datasets. Extensive statistical analysis on obtained results indicated that the proposed SATLBO-MFLANN method is better and statistically significant in comparison to its counterparts.
Keywords: FLANN; Multiplicative FLANN; TLBO; Self-adaptive TLBO; Hybrid model; Time series forecasting.
Impact of C-Factor on PSO for Solar PV based BLDC Motor Drive Control
by Manoj Kumar Merugumalla, Prema Kumar Navuri
Abstract: The constriction factor (C-factor) based particle swarm optimization algorithm is proposed for the solar photovoltaic array powered brushless direct current (BLDC) motor drive control. The rotor position sensors are completely eliminated. Instead, it is determined by measuring the changes in back-emf. The CPSO algorithm is used for the tuning of proportional-integral-derivative (PID) controller parameters for the speed control of the drive. The BLDC motor drive is modelled in MATLAB/SIMULINK and trapezoidal back emf waveforms are modelled as a function of rotor position using matlab code. This paper deals with rise time, settling time, peak overshoot and steady-state error under varying conditions and examines the effectiveness of the BLDC motor drive with the proposed algorithm. The comparison of simulation results of proposed algorithm with other PSO algorithms demonstrates the prominence of the constriction factor for the drive controller.
Keywords: Brushless direct current motor ;Solar photo voltaic array;
Particle swarm optimization ; PID controller ; Constriction factor ; Inertia weight.
Robust Estimation of IIR System's Parameter using Modified Particle Swarm Optimization Algorithm
by Meera Dash, Trilochan Panigrahi, Renu Sharma
Abstract: This paper introduces a novel method of robust parameter estimation of infinite impulse response (IIR) system. When training signal contains strong outliers, the conventional squared error based cost function fails to provide desired performance. Thus a computationally efficient robust cost functions are used here. It is a fact that the IIR system falls in local minima. Thus the gradient based algorithm which is good for finite impulse response (FIR) system, can not be used for IIR. Therefore the parameters of the IIR system is estimated using modified particle swarm optimization algorithm. The most used and analyzed robust cost functions such as Hubers and saturation nonlinearity function are used in the optimization algorithm. The simulation results show that the proposed robust algorithms are providing better performance than the Wilcoxon norm based robust algorithm and conventional error squared based PSO algorithms.
Keywords: IIR system; impulsive noise; robust estimation; Wilcoxon norm; Hubers cost function; adaptive particle swarm optimization;saturation nonlinearity.
Design of Fractional Order PID Controller for Heat Flow System using Hybrid Particle Swarm Optimization and Gravitational Search Algorithm
by Rosy Pradhan, Santosh Kumar Majhi, Bibhuti Bhusan Pati
Abstract: This paper uses the hybrid Particle Swarm Optimization (PSO) and Gravitational Search Algorithm (GSA) for the design of fractional order proportional-integrator-derivative (FOPID) controller for a heat flow system. The social behavior of PSO is combined with the motion technique of GSA. The objective of the algorithm is to obtain the optimal controller parameters for the heat flow system. To obtain the optimal computation, different performance indices such as IAE (Integral Absolute Error), ISE (Integral Squared Error), ITAE (Integral Time Absolute Error), ITSE (Integral Time Squared Error) are considered for the optimization. The performance of the hybrid PSO-GSA is compared with the IMC-PID, Fractional Oder Filter-PID, PSO-FOPID. The proposed method performs comparatively better than the other published methods. Simulink/Matlab environment is used for simulation purpose.
Keywords: FOPID; performance indices; PSO; PSO-GSA.
Suitability and Importance of Deep Learning Feature Space in the Domain of Text Categorization
by Rajendra Kumar Roul
Abstract: One of the important features of Multilayer ELM (ML-ELM) is its capability of non-linearly mapping the features to an extended dimensional space and thereby builds the input features linearly separable. This paper studies the significance of deep learning feature space using ML-ELM for classification of text data which are the follow-up of my earlier work. The previous approach discusses a new feature selection technique named Combined Cohesion Separation and Silhouette Coefficient (CCSS) to generate a good feature vector and then used it for classification of text data using ML-ELM, which is a deep learning classifier. This approach has been extended here that has two main aspects. The first aspect is to compare the performance of CCSS approach with the traditional feature selection techniques and the second aspect is to test the performance of different conventional classification techniques on the higher dimensional feature space of ML-ELM. Results of the experiment on different benchmark datasets justify that the proposed CCSS technique is comparable with the existing feature selection techniques and the ML-ELM feature space is more promising compared to the traditional TF-IDF vector space for classification of text data.
Keywords: Classification; Cohesion; Deep learning; Extreme learning machine; Multilayer ELM; Separation; Silhouette coefficient.
Scalable Keyword-Based Search and Data Manipulation on Encrypted Data
by Prabhat Keshari Samantaray, Navjeet Kaur Randhawa, Swarna Lata Pati
Abstract: Cloud based services increases productivity and reduces infrastructure cost industries which attract researchers and individuals outsource data to the remote cloud servers. However, cloud servers pose several security issues and the most prominent issue is privacy. Although encrypting the data before outsourcing preserves the data privacy, it does not support data usability such as searching with keywords on encrypted data. Searchable Symmetric Encryption (SSE) is found to be an efficient solution and in this paper, we construct an SSE method which enables search operation on encrypted data and allows various data operations on encrypted data. We perform document clustering and construct index tree based secure search scheme. The index and queries are constructed with vector space model and are encrypted using secure kNN computation method. We defined two encrypted searchable schemes based on two security models. We perform thorough experiments to justify the efficiency of the proposed model.
Keywords: encrypted index construction;document clusters;encrypted search;encrypted data operations;searchable encryption;vector encryption;.
Special Issue on: Applications of Hybrid Bio Inspired Algorithms
Stock Price Trend Prediction with Long Short Term Memory Neural Networks
by Varun Gupta, Mujahid Ahmad
Abstract: Stock market is an immensely complex, chaotic and dynamic environment. Thus, the task of predicting changes in such an environment becomes challenging with regards to its accuracy. A number of approaches have been adopted to take on that challenge and machine learning has been as the crux in many of them. There are plenty of examples of algorithms based on machine learning yielding satisfactory results for such type of prediction. This paper presents the usage of Long Short Term Memory (LSTM) networks in this scenario, to predict future trends of stock market prices based on the patterns from price history, paired with technical analysis indicators. To achieve this, a model has been built, and a series of experiments have been conducted through a number of parameters and the results were analyzed against predefined metrics to assess if this algorithm presents any improvements in front of other machine learning methods and strategies. Also, a comparative study is presented which analyzes popularly used optimizers and error schemes to check which given optimizer yields the best results. The results obtained are promising and presented a reasonably accurate prediction for the rise or fall of a particular stock in the near future.
Keywords: Stock market prediction; LSTM; Recurrent neural networks; artificial neural networks; machine learning; deep learning; artificial intelligence; soft computing
PREDICTION OF AIR POLLUTION USING LSTM BASED RECURRENT NEURAL NETWORKS
by Varun Gupta, Akshat Jain, Ashim Bhasin
Abstract: This paper proposes a system that predicts the pollution level at some hour at a place. It also infers about the various parameters associated with the increasing pollution across the globe, its ill effects and the future scenario of the same. An air quality dataset reporting level of pollution and weather every hour for five years is taken and Long Short Term Memory network (LSTM) based Recurrent Neural Networks using keras library with Tensorflow as back-end were applied in a python environment. The paper studies all 13 parameters affecting the weather and air pollution conditions and forecasts the pollution for any hour given the weather conditions and pollution value for the previous hour.
Keywords: Air Pollution Prediction; LSTM; Recurrent Neural Networks; Artificial Neural Networks; deep learning; machine learning; soft computing; artificial intelligence
Special Issue on: CMDM 2017 Computational Intelligence and Data Mining
A Co-evolutionary Decomposition-based Algorithm for the Bi-level knapsack optimization problem
by Abir Chaabani, Lamjed Ben said
Abstract: Bi-level optimization problems (BOPs) are a class of challenging problems with two levels of optimization tasks. These problems allow to model a large number of real-life situations in which a first decision maker, hereafter the leader, optimizes his objective by taking the follower\'s response to his decisions explicitly into account. In this way, evaluating a solution in the upper level requires finding an optimal solution to the lower level problem. This fact makes BOPs difficult to handle and have kept researchers and practitioners busy alike. Recently, a new research field, called EBO (Evolutionary Bi-Level Optimization) has appeared thanks to the promising results obtained by the use of EAs (Evolutionary Algorithms) to solve such kind of problems. In this context,two recently proposed EBO called CODBA and CODBA-II were proposed to solve combinatorial BOPs. The proposed approaches were able to improve the quality of generated bi-level solutions regarding to the recently proposed methods within this research area. In fact, a wide range of applications fit the bi-level programming framework and real-life implementations still scarce. For this reason, we propose in this paper a Co-evolutionary Decomposition-based Bi-level Algorithm for the bi-level knapsack optimization problem. The computational performance of the proposed algorithm turned out to be quite efficient on both computation time and solution quality regarding to other competitive EAs.
Keywords: Bi-level combinatorial optimization; evolutionary methods; bi-level\r\nknapsack problem.
Web service selection based on QoS and user profile
by Ilhem Feddaoui, Faîçal Felhi, Jalel Akaichi
Abstract: The Web Services are from different sources, heterogeneous, and of large volume. The user is in a crucial situation to select the best Web services. The Web service selection process aims to discovery the desired Web services; as it allows to select the best Web services to users' query. In particular, various Web services have the same functionalities, so we need another factor to select the desired Web services, which is the Quality of Service (QoS). The QoS has an important role in the Web service selection process, it aims to classify the Web service that have same functionality. This paper focuses on different concepts of the QoS. We present a new approach that is composed by two services; its role is primarily the best Web service selection in relation with users' query and profile. In our approach, a better knowledge of user behavior is important because users can participate in research design and construction. The experiment shows that our method can accurately recommend the needed Web services in a faster time.
Keywords: Web service; Query; User profile; QoS.
An effective Genetic Algorithm for solving the Capacitated Vehicle Routing Problem with Two-dimensional Loading Constraint
by Ines Sbai, Olfa Limam, Saoussen Krichen
Abstract: In this article, we focus on the symmetric Capacitated Vehicle Routingrnproblem where customer demand is composed of two-dimensional weightedrnitems. The objective consists in designing a set of trips, starting and terminating at a central depot, that minimize the total transportation cost with a homogenous fleet of vehicles based on a depot node.rn Items in each vehicle trip must satisfy the two-dimensional orthogonal packing constraint. rnThe capacitated vehicle routing problem with two-dimensional loading constraint is an NP-hard problem of high complexity.rn Given the importance of this problem, many solution rnapproaches have been developed. So, to find better solution to this challenging problem, we propose to use a New Heuristic based on a Genetic Algorithm. rnOur algorithm is tested with 150 benchmark instances and compared with state-of-the-art approaches. Results shown that our proposed approach is competitive in terms of the quality of the solutions found.
Keywords: Capacitated Vehicle Routing Problem; Loading; Genetic Algorithm;rn2L-CVRP.
A Multi-Level Study for Trust Management Models Assessment in VANETs
by ilhem Souissi, Nadia Ben Azzouna, Lamjed Ben Said
Abstract: Nowadays, trust management is one of the key elements to ensure a high security level in ad hoc networks. Trust assessment can be perceived at three levels. First, the data perception trust need to be assessed in order to ensure a high quality of raw sensed data. Second, the trust relationship assessment is required to detect the selfish and malicious entities and to maintain the data integrity. Finally, the data fusion trust is essential to preserve the performance of the fusion process. In this paper, we intend to point out the need to integrate the data perception trust, the communication trust and the data fusion trust in order to preserve the information trustworthiness in VANETs. We further browse the literature to identify recent advancements with regard to each type of trust.
Keywords: Data Perception Trust; Communication Trust; Data Fusion Trust; VANETs
Special Issue on: BDCA'18 Data Science and Applications
Information Technology performance management by Artificial Intelligence in Microfinance Institutions: An overview
by Kaicer Mohammed
Abstract: This paper presents an overview of the use of new information technology to improve the management of microfinance institutions, experiencing a gap due to the growth of the microfinance sector and the diversity of products and services they offer to the target populations. We will show that artificial intelligence could play a role to ensure reliable management information systems in MFIs.
Keywords: Management of Informatics Technology; Artificial intelligence; Microfinance Institution; Central risk.
Special Issue on: CMDM 2017 Computational Intelligence and Data Mining
MC4.5 decision tree algorithm: An improved use of continuous attributes
by Anis Cherfi, Kaouther Nouira, Ahmed Ferchichi
Abstract: C4.5 is one of the top ten data mining algorithms, it is the most widely used decision trees construction techniques. Although effective, it suffer from the problem of complexity when it deals with continuous attributes. It also leads to a certain level of information loss. Therefore, minimizing such loss, and reducing the time complexity is one of the main goals in this paper. With the intention of alleviating these problems, this paper presents a novel algorithm namely MC4.5, which proposes the statistical mean as an alternative to the C4.5 threshold selection process. To demonstrate the effectiveness of the new algorithm, a complete evaluation was launched to prove that MC4.5 complies with the objectives previously mentioned. From the theoretical perspective, we develop an analysis of the complexity to compare algorithms. Empirically, we conduct an experimental study using 30 data sets to prove that, in most cases, the proposed algorithm leads to smaller decision trees with better accuracy comparing to the C4.5 algorithm.
Keywords: Decision tree; MC4.5; C4.5; Statistical mean; Continuous attributes;rnClassification; Information gain.
Solving flexible job-shop problem with sequence dependent setup time and learning effects using an adaptive genetic algorithm
by Ameni Azzouz, Meriem Ennigrou, Lamjed Ben SAID
Abstract: For the most scheduling problems studied in literature, job processing times are assumed to be known and constant over time. However, this assumption is not appropriate for many realistic situations where the employees and the machines execute the same task in a repetitive manner. They learn how to perform more efficiently. As a result, the processing time of a given job is shorter if it is scheduled later, rather than earlier in the sequence. In this paper, we consider the Flexible Job Shop Problem (FJSP) with two kinds of constraint, namely, the sequence-dependent setup times (SDST) and the learning effects. Makespan is specified as the objective function to be minimized. To solve this problem, an Adaptive Genetic Algorithm (AGA) is proposed. Our algorithm uses an adaptive strategy based on : (1) the current specificity of the search space, (2) the preceding results of already used operators and (3) their associated parameter settings. We adopt this strategy in order to maintain the balance between exploration and exploitation. Experimental studies are presented to assess and validate the benefit of the incorporation of the learning process to the SDST-FJSP over the original problem.
Keywords: scheduling problem; Genetic algorithm; Adaptive strategy; Learning effects.
Contributions to the Automatic Processing of the User-Generated Tunisian Dialect on the Social Web
by Jihene Younes, Hadhemi Achour, Emna Souissi, Ahmed Ferchichi
Abstract: With the growing use of social media in the Arab world, Arabic dialects are rapidly spreading on the web, leading to a growing interest from NLP researchers. These dialects are however, still under-resourced languages and the lack of available dialectal resources is a major obstacle to their study and processing. In this paper, we focus on the automatic processing of the user-generated Tunisian dialect (TD) on the social web and propose an approach that aids to automatically generate TD language resources (LRs), useful for any NLP research work dealing with this dialect. This approach exploits the large amounts of textual productions on the social web in order to extract and generate dialectal content. It is based on two main NLP components, namely the TD Identification and the TD transliteration. A machine learning approach using Conditional Random Fields (CRF), is proposed for implementing these two components and reached an accuracy of 87.45 for the TD identification and 90.49 for the automatic generation of dialectal contents by transliteration.
Keywords: Tunisian Dialect; language resources; corpora; lexica; identification; transliteration; natural language processing; machine learning.
Special Issue on: ISCSA2017 Computational Intelligence and Applications
Air pollution prediction through Internet of Things technology and Big Data Analytics
by Safae Sossi Alaoui, Brahim Aksasse, Yousef Farhaoui
Abstract: Air pollution is one of the biggest and serious challenges facing our planet nowadays. In fact, the need to develop models to predict this issue is considered so crucial. Indeed, our work aimed at building an accurate model to predict air quality of US country by using a dataset collected from connected devices of Internet of Things (IoT), namely from wireless sensor networks (WSN). Therefore, the huge amount of data captured by these sensors (approximately 1.4 million observations) brings about a highly complex data that necessitates new form of advanced analytic; its about Big Data Analytics. In this paper, we examine the possibility to make a fusion between the two new concepts Big Data and Internet of Things; in the context of predicting Air pollution that occurs when harmful substances; like NO2, SO2, CO and O3, are introduced into Earth's atmosphere.
Keywords: Internet of Things (IoT); Wireless sensor networks (WSN); Air pollution; Air Quality Index (AQI); Big Data Analytics; Apache Spark.
Big data: A distributed storage and processing for online learning systems
by Karim DAHDOUH, Ahmed DAKKAK, Lahcen OUGHDIR
Abstract: The new information and communication technologies have changed the way of teaching and learning. In particular, the big data technology that has recently been developed to overcome the limitations of traditional systems of storage, processing, and analysis. In fact, big data has been used in several fields including health care, public services, and online services such as social media and online learning. It offers a rich set of new technologies in terms of data integration, distributed storage, parallel processing, and data visualization. Furthermore, big data provides many techniques to bring solutions to various educational problems such as the courses recommendation engine, the prediction of learner behaviour, the exponential growth of the learners and pedagogical resources, etc. Today, thanks to the big data ecosystem, it is possible to greatly improve the effectiveness and performance of the online learning services. This article presents the big data paradigm, its components, technologies, and characteristics. It proposes an approach for incorporating big data, online learning systems, and cloud computing in order to enhance the efficiency of the distance learning environment. Also, it provides a methodology to store and process the data produced by online learning platforms using advanced big data technologies and tools. Moreover, It explores the advantages and benefits that big data offer to students, teachers and online learning professionals.
Keywords: computing environments for human learning; big data; cloud computing; e-learning; Online learning; Learner; Learning Management Systems (LMS); NoSQL database; Hadoop; MapReduce; Spark; Cassandra; Hive; Apache Flume; Apache Sqoop.
An Independent-domain Natural Language Interface for Multimodel Databases
by Bais Hanane
Abstract: Databases are gaining prime importance in the world of modern computing. Retrieving information stored in databases required the knowledge of the database Query languages such as Structured Query Language (SQL). However, learning this language can be difficult for non-expert users.Hence, the using of natural language is a very easy and convenient method that can provide powerful improvements to the use of data stored in databases. In this paper, we present the architecture of an intelligent natural language interface for a multimodel database. This interface functions independently of database domain, language and model. The using of machine learning approach helps our system to improve automatically its knowledge base through experience.
Keywords: Databases; Natural Language Processing (NLP); Intermediate XML Logical Query (IXLQ); Extended Context Free Grammar (ECFG); intelligent interface.
Special Issue on: IWCIA2017 Innovative Computational Intelligence for Knowledge Representation and Learning
Fast Training of Adaptive Structural Learning Method of Deep Learning for Multi Modal Data
by Shin Kamada, Takumi Ichimura
Abstract: Recently, deep learning has been applied in the techniques of artificial intelligence. Especially, their new architectures performed good results in the field of image recognition. However, the method is required to train not only image data, but also numerical data, text data, and other binary data. Multi modal data consists of two or more kinds of data such as a pair of image and text of giving an explanation of the image. The arrangement of multi modal data in the traditional method is formed in the squared array with no specification. In this paper, the method can modify the squared array of the multi modal data, according to the similarity of input-output pattern of adaptive structural learning method of Deep Belief Network. Some experimental results show that the computational time of deep learning decreases.
Keywords: Multi Modal Data; Automatically Data Arrangement Method; Deep Learning; Adaptive Learning Method; Restricted Boltzmann Machine; Deep Belief Network; Shorting Learning Time.
Characteristics of Contrastive Hebbian Learning with Pseudorehearsal for Multilayer Neural Networks on Reduction of Catastrophic Forgetting
by Motonobu Hattori, Shunta Nakano
Abstract: Neural networks encounter serious catastrophic forgetting or catastrophic interference when information is learned sequentially. One of the methods which can reduce catastrophic forgetting is pseudorehearsal, in which pseudopatterns are learned with training patterns. This method has shown superior performance for multilayer neural networks trained by the backpropagation algorithm. However, the backpropagation algorithm is biologically implausible because it requires passage of error signals backward from output neurons to input ones. That is, the learning cannot by executed locally. On the other hand, Contrastive Hebbian Learning (CHL) is a learning method using Hebbian rule for synaptic weight changes. Since Hebbian learning can be performed locally between two neurons and doesnt need to take into account error information computed at output neurons, it is much more biologically plausible than the backpropagation algorithm. In this paper, we examine characteristics of multilayer neural networks trained by CHL with pseudorehearsal when information is applied sequentially, and how catastrophic forgetting can be reduced.
Keywords: Contrastive Hebbian Learning; Pseudorehearsal; Multilayer Neural Networks; Catastrophic Forgetting; Pseudopatterns; Hebbian Rule; Additional Learning.
Search Performance Analysis of Qubit Convergence Measure for Quantum-Inspired Evolutionary Algorithm Introducing on Maximum Cut Problem
by Yoshifumi Moriyama, Ichiro Iimura, Shigeru Nakayama
Abstract: The quantum-inspired evolutionary algorithm (QEA) and QEA with a pair-swap strategy (QEAPS), where each gene is represented by a quantum bit (qubit), and the qubit is updated by a unitary transformation in both algorithms. QEA and QEAPS can automatically shift the evolution from a global search to a local search and have shown superior search performance to the classical genetic algorithm. However, the population get into a locally optimal solution and the solution search stagnates when the probability amplitudes of qubit excessively converge to '0> or '1>. In this study, we have proposed a measure that can confirm convergence state of qubits. From the results of the computational experiment in the maximum cut problem, we have clarified that the proposed measure can estimate the state of the qubit, and the quality of the obtained solution is improved by applying the method for maintenance of diversity.
Keywords: quantum-inspired evolutionary algorithm; QEA; QEA with pair-swap strategy; QEAPS; qubit convergence measure; Noah's ark strategy; population-based incremental learning; PBIL; univariate marginal distribution algorithm; UMDA; estimation of distribution algorithms; EDAs; maximum cut problem.
A Generative Model Approach for Visualising Convolutional Neural Networks
by Masayuki Kobayashi, Masanori Suganuma, Tomoharu Nagao
Abstract: Convolutional neural networks (CNN) have continued to achieve outstanding performance in a variety of computer vision tasks. CNNs have advanced significantly deeper and deeper, continuing to show substantial improvements for various tasks. Despite their successes, their models are often considered as black-box predictors, and their uninterpretable natures are major problems. In this paper, we introduce a new visualisation framework based on generative adversarial networks (GAN) to provide insight into how CNNs work. Following the standard GAN training, we train the generator and the discriminator to produce natural images that activate a particular unit in the pre-trained CNN. We apply our method to the AlexNet and CaffeNet and visualise the neuron activations. Our method is very simple, yet produces comparatively recognisable visualisations. We also attempt to use our visualisation as indications of models trust and verify the potential of our visualisations.
Keywords: Convolutional Neural Network; Generative Adversarial Networks; Visualisation; Activation Maximisation; Interpretability.
Time Series Classification using MACD-Histogram-based Recurrence Plot
by Keiichi Tamura, Takumi Ichimura
Abstract: Time series classification is one of the most active research topics in time series data mining, because they cover a broad range of applications in many different domains. Representation for time series is a technique that converts time series to feature vectors representing the characteristics of time series. The performance of classifying time series depends on this representation. Chaotic time series analyses have been well-studied. Moreover, recurrence plotting underlying chaos theory is one of the most robust time series representation for time series. In this study, we propose new time series representation utilizing the recurrence plot technique. Moving average convergence divergence (MACD) histogram is the acceleration of time that can represent local-variation in time series. Therefore, a recurrence plot that is made from MACD histogram, which is called a MACD-Histogram-based recurrence plot (MHRP), can handle time series very well. Recurrence plots are referred to as gray-scale images and we utilize stacked auto-encoders as a classifier for MHRPs. To evaluate the performance of the proposed classifier, experiments using the UCR time series classification archive was conducted. The experimental results showed that the proposed classifier outperforms other methods.
Keywords: Time series classification; Time series mining; Recurrence plot; Chaotic time series analysis; MACD histogram.
Mining Non-Redundant Recurrent Rules from a Sequence Database
by SeungYong Yoon, Hirohsia Seki
Abstract: Many methods have been studied for mining sequential patterns from arnsequence database. In particular, Lo et al. have proposed the notionrnof "recurrent rules" and an algorithm called NR^3 for mining them.rnRecurrent rules can express temporal constraints such as ``Whenever arnseries of precedent events occurs, eventually a series of consequentrnevents occurs,'' and they are useful in various domains, includingrnsoftware specification and verification. Although the algorithm NR^3rnand its successor BOB for mining non-redundant recurrent rules havernbeen given by Lo et al., mining recurrent rules still requiresrnconsiderable computational costs. In this paper, we propose a newrnalgorithm, called LF-NR^3, to make NR^3 more efficient, which is basedrnon a familiar program transformation "loop fusion". We apply the looprnfusion technique to NR^3, thereby simplifying the operations of thernoriginal mining algorithm. We also make use of a hash-based datarnstructure to reduce loads of the manipulation of sequences repeatedlyrnrequired in mining recurrent rules. We show the effectiveness of ourrnproposed method based on experimental results of some datasets used inrnthe literature.
Keywords: sequential pattern mining; recurrent rule; sequencerndatabase; program transformation; vertical data format.
Acquisition of Characteristic Sets of Block Preserving Outerplanar Graph Patterns by a Two-Stage Evolutionary Learning Method for Graph Pattern Sets
by Fumiya Tokuhara, Tetsuhiro Miyahara, Tetsuji Kuboyama, Yusuke Suzuki, Tomoyuki Uchida
Abstract: Knowledge acquisition from graph structured data is an important task in machine learning and data mining. Block preserving outerplanar graph patterns are graph structured patterns having structured variables and are suited to represent characteristic graph structures of graph data modeled as outerplanar graphs. We propose a learning method for acquiring characteristic sets of block preserving outerplanar graph patterns by a two-stage evolutionary learning method for graph pattern sets as individuals, from positive and negative outerplanar graph data, in order to represent characteristic graph structures more concretely.
Keywords: evolutionary method; genetic programming; outerplanar graphs; sets of graph structured patternsrn.
Special Issue on: RAACE17 Applications of Hybrid Bio Inspired Algorithms
Fuzzy Knowledge Based Fractional Order PID Control Implementation with Nature Inspired Algorithms
by Ambreesh Kumar, Rajneesh Sharma
Abstract: In this paper, we attempt to hybridize nature inspired optimization techniques with fuzzy knowledge based proportional integral derivative (PID) control for applications on fractional order systems. Two nature inspired approaches, namely, Genetic algorithm and Ant Colony algorithms have been employed for tuning the parameters of the fuzzy knowledge based fract-order PID controller offline. In the next stage, we fine tune the PID controller parameters using a fuzzy knowledge based formulation. In our proposed nature inspired fractional fuzzy PID (NIFFPID) framework, GA has been used for optimizing the inputs to the ANT controller. We illustrate effectiveness of our methodology by simulation results on four plants: one integer order and three fractional order ones having different orders. Simulation results and comparison of our approach against other approaches, viz., fractional order PID-ANT, fractional order PID-GA, fuzzy fractional PID-ANT and fuzzy fractional PID-GA, shows feasibility and effectiveness of our approach for fract order systems.
Keywords: Integer order plant; fract order plants; fuzzy knowledge based control; NIFFPID approach.