Vol. 33 · No. 2

This paper introduces an unsupervised learning algorithm for optimal training of competitive neural networks. The learning rule of this algorithm is derived from the minimization of a new objective criterion using the gradient descent technique. Its learning rate and competition difficulty are dynamically adjusted throughout iterations. Numerical results that illustrate the performance of this algorithm in unsupervised pattern classification and image compression are also presented, discussed, and compared to those provided by other well-known algorithms for several examples of real test data.

Latent Semantic Indexing (LSI) has been widely used in information retrieval due to its efficiency in solving the problems of polysemy and synonymy. However, LSI is notably a computationally intensive process because of the computing complexities of singular value decomposition and filtering operations involved in the process. This paper presents MR-LSI, a MapReduce based distributed LSI algorithm for scalable information retrieval. The performance of MR-LSI is first evaluated in a small scale experimental cluster environment, and subsequently evaluated in large scale simulation environments. By partitioning the dataset into smaller subsets and optimizing the partitioned subsets across a cluster of computing nodes, the overhead of the MR-LSI algorithm is reduced significantly while maintaining a high level of accuracy in retrieving documents of user interest. A genetic algorithm based load balancing scheme is designed to optimize the performance of MR-LSI in heterogeneous computing environments in which the computing nodes have varied resources.

A Personal DataSpace Management System is a platform to manage personal data with heterogeneous data types, in which keyword query is a primary query form for users who know little about the structure of the dataspace. Unlike exploratory queries in web search, a user in a personal dataspace usually has a specific search target and wants to find some known items in mind. To improve result quality in terms of query relevance in a personal dataspace, we propose the concept of compact index in this paper. We refer to the most important and representative semantics from documents as core content, and build compact index on it. We propose algorithm for selecting core content from a document based on semantic analysis, which can process English and Chinese documents uniformly. Furthermore, a software platform named Versatile is introduced for flexible personal data management, in which core content is extracted for building compact indexes and generating query-biased snippet efficiently and accurately. Finally, extensive experiments have been conducted to show the effectiveness and feasibility of compact indexes in personal dataspace management system.

This paper presents a number of security and VO management capabilities in a large-scale distributed Grid operating system. The capabilities formed the basis of the design and implementation of a number of security and VO management services in the system. The main aim of the paper is to provide some idea of the various functionality cases that need to be considered when designing similar large-scale systems in the future.

Several methods have been proposed to analyse trajectory data. However, a few of these methods consider trajectory relations with relevant features of the geographic space. One of the best-known methods that take into account the geographical regions crossed by a trajectory is the SMoT algorithm. Nevertheless, SMoT considers only disjoint geographic regions that a trajectory may traverse, while many regions of interest are contained in other regions. In this article, we extend the SMoT algorithm for discovering stops in nested regions. The proposed algorithm, called SMoT+, takes advantage of information about the hierarchy of nested regions to efficiently discover the stops in regions at different levels of this hierarchy. Experiments with real data show that SMoT+ detects stops in nested regions, which are not detected by the original SMoT algorithm, with minor growth of processing time.

This paper describes and analyses the possibilities of a quantitative assessment of message transmission between safety-related equipment for control and communication systems with a guarantee of a higher safety integrity level (SIL). The theoretical methods and standards recommended for industrial safety-related control, information and communication systems with SIL3 are described. The main part of the contribution covers theoretical methods and practical procedures used within a safety analysis of safety-related message transmission with the requirement of SIL4 for the area of railway interlocking systems. The theoretical analysis of these methods is compared with the knowledge and experience the authors gained within several safety verifications of such systems in practice in Slovakia as well as abroad. Also, based on a case study in the mentioned application area, the authors proposed their own quantitative mathematical model for assessing the safety integrity level of message transmission within a closed transmission system with the failure tree analysis (FTA) and Markov's analysis (MA). The designed model, which evaluates the effect of electromagnetic interferences (EMI) and random hardware failures of a safety-related communication system (SRComS) to its safety, was realised and verified. The verification in available conditions confirmed the contribution of the model to the process of objectification of the safety assessment of message transmission via SRComS.

Devices that form a wireless sensor network (WSN) system are usually remotely deployed in large numbers in a sensing field. WSNs have enabled numerous applications, in which location awareness is usually required. Therefore, numerous localization systems are provided to assign geographic coordinates to each node in a network. In this paper, we describe and evaluate an integrated software framework WSNLS (Wireless Sensor Network Localization System) that provides tools for network nodes localization and the environment for tuning and testing various localization schemes. Simulation experiments can be performed on parallel and multi-core computers or computer clusters. The main component of the WSNLS framework is the library of solvers for calculating the geographic coordinates of nodes in a network. Our original solution implemented in WSNLS is the localization system that combines simple geometry of triangles and stochastic optimization to determine the position of nodes with unknown location in the sensing field. We describe and discuss the performance of our system due to the accuracy of location estimation and computation time. Numerical results presented in the paper confirm that our hybrid scheme gives accurate location estimates of network nodes in sensible computing time, and the WSNLS framework can be successfully used for efficient tuning and verification of different localization techniques.

Recent industry requires efficient fault discovering and isolation solutions in process equipment service. This problem is a real-world problem of typically ill-defined systems, hard to model, with large-scale solution spaces. Design of precise models is impractical, too expensive, or often non-existent. Support service of equipment requires generating models that can analyze the equipment data, interpreting the past behavior and predicting the future one. These problems pose a challenge to traditional modeling techniques and represent a great opportunity for the application of AI-based methodologies, which enable us to deal with imprecise, uncertain data and incomplete domain knowledge typically encountered in real-world applications. In this paper the state of the art, theoretical background of conventional and AI-based techniques in support of service tasks and illustration of some applications to process equipment service on bio-ethanol production process are shortly described.

Online social networks provide large amount of valuable data and may serve as research platforms for various social network analysis tools. In this study, we propose a mathematical model for efficient exploration of an online social network. The goal is to spend minimal amount of time searching for characteristics which define a sub-network of users sharing the same interest or having certain common property. We further develop an efficient hybrid method (HEA), based on the combination of an Evolutionary Algorithm (EA) with Local Search procedure (LS). The proposed mathematical model and hybrid method are benchmarked on real-size data set with up to 10 000 users in a considered social network. We provide optimal solutions obtained by CPLEX solver on problem instances with up to 100 users, while larger instances that were out of reach of the CPLEX were efficiently solved by the proposed hybrid method. Presented computational results show that the HEA approach quickly reaches all optimal solutions obtained by CPLEX solver and gives solutions for the largest considered instance in very short CPU time.

This paper investigates the optimization of transmission spatial layer selection with unreliable rank indicator (RI) in downlink LTE systems. Taking the block error rate (BLER) into consideration, we propose an accurate throughput calculation (ATC) algorithm at user equipment (UE) side as well as at evolved NodeB (eNB) side. On the basis of ATC algorithm, we propose an accurate RI selection algorithm to periodically choose the preferred number of transmission spatial layers at UE side. Further based on acknowledgement (ACK)/ negative acknowledgement (NACK) history, channel quality indicator (CQI) is adjusted at eNB side to achieve the throughput optimal target BLER. By substituting the derived BLER into ATC algorithm, the optimal number of transmitted spatial layers in current downlink channel is derived at eNB side. Simulation results show that both the proposed CQI adjustment scheme for spatial layer selection and RI selection algorithm yield up significant throughput improvement for different evaluation scenarios in LTE systems.

The popularity of group recommender systems has increased in the last years. Ever more social activity is generated by users over the web and thus not only domains such as TV, music or holiday resorts are used and researched for group recommendations, but also collaborative learning support, digital libraries and other domains seem to be promising for group recommendations. Moreover, principles of group recommenders can be used to overcome some single-user recommendation shortcomings, such as cold start problem. Numerous group recommenders have been proposed; they differ in application domains (which in the group recommender context are specific in various group characteristics). Today's group recommenders do not include and use the power of social aspects (a group structure, social status etc.), which can be extracted and derived from the group. We provide a survey of group recommendation principles for the Web domain and discuss trends and perspectives for further research.