|
Foreword Lucian Vinţan
Abstract. The main aim of this work is to synthetically point out a state of the art in developing automatic design space exploration methods dedicated to Computer Architecture multi-objective optimizations, as they were developed during the time by the author and his research group. Also the paper points out some of the author's envisaged further work directions that would hopefully advance this important research field. So, in order to achieve better convergence speed and solutions' quality, specific domain-knowledge for each of the target computing system is necessary to be developed and integrated into the DSE tools. Adequately representing domain-knowledge is a significant scientific challenge. Also, automatically calculating the degrees of logical contradiction for a certain domain-knowledge represented using fuzzy logic rules would be of certain scientific interest. Effective meta-optimization methods are needed. Integrating some advanced Response Surface Models to accelerate the automatic DSE processes represents another fertile research challenge, too. Read the pdf
Gheorghe Pop, Dragoş Drăghicescu, Dragoş Burileanu
Abstract. The performance of a speaker recognition systems based on Gaussian mixture models is often impaired both by the low quality and by short duration of test speech samples. A best material selection criterion is described in this paper, especially suitable for forensic automatic speaker recognition systems, where even enrollment speech quality might be important in some cases. The material selection is performed by checking well known short-time measures of input speech that carry quality information, such as linear cepstral peak, spectral autocorrelation peak to valley ratio, and windowed autocorrelation lag energy. Our tests show that the proposed approach outperforms reported speaker recognition solutions that consider quality of the input speech, at least in co-channel speech conditions. Read the pdf.
Călin Bîra, Liviu Gugu, Mihaela Maliţa, Gheorghe M. Ştefan
Abstract. The MapReduce architecture and the associated programming model defined for the one-chip Connex organization gets an important additional feature: the associated data vectors can be organized as a two-dimension arrays of vertical and horizontal vectors which are managed by the transpose operation in order to maximize the degree of parallel execution. Thus, a one-chip Map-Reduce engine is able to perform easy additional intra vector operations by transforming the horizontal vectors, involved in Map operations, in vertical vectors using the transpose operation. We propose the Transpose-MapReduce high-level architecture supported by a fine grain cellular structure -- the Connex system already implemented in silicon. It is structured in three sub-architectures: the data processing, data transfer and inter-cell communication architectures. A simulator written in SCHEME is used to write and evaluate few meaningful algorithms: AES encryption, FFT, Batcher's merge-sort. Read the pdf.
Tao Song, Linqiang Pan
Abstract. Spiking neural P systems (shortly called SN P systems) are a class of distributed and parallel neural-like computing devices, which are inspired by the way of biological neurons communicating with each other by means of impulses/spikes. SN P systems with cooperating rules are a new variant of SN P systems, where each neuron has the same number of components and some components of a neuron can be empty. In a step of a computation, one component from each neuron is used, with the same label in all neurons; from these components, one rule is applied, in the way usual in SN P systems. In the terminating mode, adopted in this paper, after choosing a component of the neurons, this component is applied until no rule from this component, in any neuron, is enabled (we switch from a component to another one, nondeterministically chosen, when no rule of the component can be used, in any neuron of the system). In this work, we investigate how many neurons are needed to construct a Turing universal SN P system with cooperating rules as a number generator in terminating mode. Specifically, we construct a Turing universal SN P system having 8 neurons, which can generate/compute any set of Turing computable natural numbers. This result gives an answer to an open problem formulated in [V.P. Metta, S. Raghuraman, K. Krithivasan, CMC15, 267--282, 2014]. Read the pdf.
Paulina Grzegorek, Janusz Januszewski
Abstract. In 2-dimensional bin packing problem each item is a rectangle of side lengths not greater than 1. The items are packed online into square bins of size 1×1 and 90°-rotations are allowed. In t-space bounded model of online bin packing each item can be packed only into one of t active bins. If it is impossible to pack an item into any active bin, we close one of the current active bins and open a new active bin to pack that item. In this paper a 3.577-competitive 3-space bounded online packing algorithm is presented. Furthermore, an online algorithm for packing squares with the competitive ratio 2.8 is described. Read the pdf.
Ricardo Soto, Broderick Crawford, Eric Monfroy, Fernando Paredes
Abstract. A Constraint Satisfaction Problem (CSP) consists in a sequence of variables holding a domain of possible values and relations among these variables called constraints. A meta-CSP can be seen as a metaproblem whose decomposition leads to a set of CSPs. The meta-variables correspond to subproblems of the original problem, and a meta-constraint is a relation among those meta-variables. Meta-CSPs find many applications in industry, usually in processes that involve time and actions such as the control of a robot, a manufacturing process, or the scheduling of any common activity. In this paper, we introduce the notion of Sequentially Dependent Meta-CSP (SD Meta-CSP), which extends the meta-CSP in order to support applications where a dependency between sub-problems is mandatory. In this case, the meta-CSP is decomposed into a set of sub-problems {Pi, Pi+1, … , Pn}, but the instance of the sub-problem Pi+1 sequentially depends on the solution of the sub-problem Pi. In this work we provide a formal definition for the SD Meta-CSPs, a framework to handle it, and we illustrate its applicability to video games. In particular, we model and implement agents as SD Meta- |