Klaipėdos universitetas

Monte Carlo experiments are an efficient tool for investigation of the Laser-Induced Damage Threshold (LIDT) testing with pulsed lasers. In this study, the approach of sequential Monte Carlo search is developed for LIDT testing with bundle of laser pulses and compared with the approach of Sample Average Approximation (SAA). The likelihood ratio test is applied to accept or reject the hypothesis about the data distribution.

This paper presents a new algorithm for a batch of task makespan minimisation in heterogeneous multigrid computing. Heterogeneous grids are known to cause straggling task problem that increases task execution makespan. Existing task distribution algorithms solve this problem by using information about the compute node capacities or task sizes. However, such information may not always be available. Task stalling solves both problems. However, this method is described for queuing systems consisting of only two heterogeneous servers or grids. Our proposed algorithm is based on an improved task stalling method, allowing it to distribute tasks in systems consisting of two or more grids. Experiment results show reduced task execution makespan by up to 19,92% compared to FIFO. This allows us to conclude that the new algorithm is suitable for a batch of task makespan minimisation in heterogeneous multigrid computing.

This paper is focused on the Bayes approach to multiextremal optimization problems, based on modelling the objective function by Gaussian random field (GRF) and using the Euclidean distance matrices with fractional degrees for presenting GRF covariances. A recursive optimization algorithm has been developed aimed at maximizing the expected improvement of the objective function at each step, using the results of the optimization steps already performed. Conditional mean and conditional variance expressions, derived by modelling GRF with covariances expressed by fractional Euclidean distance matrices, are used to calculate the expected improvement in the objective function. The efficiency of the developed algorithm was investigated by computer modelling, solving the test tasks, and comparing the developed algorithm with the known heuristic multi-extremal optimization algorithms.

Šiame darbe sudarytas rekurentinis paslėptųjų Markovo modelių parametrų vertinimo algoritmas. Paslėptieji Markovo modeliai modeliuojami Gauso skirstiniu, kurio parametrai pasiskirstę pagal daugiamatį normalųjį dėsnį su nežinomais vidurkių vektoriumi ir kovariacijų matrica. Nežinomų parametrų įverčiai gaunami didžiausio tikėtinumo metodu. Rekurentinis algoritmas sudarytas remiantis didžiausio tikėtinumo metodu išvestomis formulėmis ir klasikiniu EM algoritmu. Kadangi rekurentinio algoritmo vykdymo laikas yra proporcingas apdorojamų stebėjimų skaičiui, tai jis gali būti naudojamas modelio parametrų vertinimui realiu laiku. Realizuoto rekurentinio EM algoritmo savybės buvo ištirtos kompiuteriniu eksperimentu klasterizuojant duomenis. Jis taip pat gali būti taikomas duomenų klasifikavimo ir atpažinimo realiu laiku uždaviniams spręsti.

This work contains Monte–Carlo Markov Chain algorithm for estimation of multi-dimensional rare events frequencies. Logits of rare event likelihood we are modeling with Poisson distribution, which parameters are distributed by multivariate normal law with unknown parameters – mean vector and covariance matrix. The estimations of unknown parameters are calculated by the maximum likelihood method. There are equations derived, those must be satisfied with model’s maximum likelihood parameters estimations. Positive definition of evaluated covariance matrixes are controlled by calculating ratio between matrix maximum and minimum eigenvalues.