All issues

This article explores a method of machine learning based on the theory of random functions. One of the main problems of this method is that decision rule of a model becomes more complicated as the number of training dataset examples increases. The decision rule of the model is the most probable realization of a random function and it's represented as a polynomial with the number of terms equal to the number of training examples. In this article we will show the quick way of the number of training dataset examples reduction and, accordingly, the complexity of the decision rule. Reducing the number of examples of training dataset is due to the search and removal of weak elements that have little effect on the final form of the decision function, and noise sampling elements. For each $(x_i,y_i)$-th element sample was introduced the concept of value, which is expressed by the deviation of the estimated value of the decision function of the model at the point $x_i$, built without the $i$-th element, from the true value $y_i$. Also we show the possibility of indirect using weak elements in the process of training model without increasing the number of terms in the decision function. At the experimental part of the article, we show how changed amount of data affects to the ability of the method of generalizing in the classification task.

Nowadays the random search became a widespread and effective tool for solving different complex optimization and adaptation problems. In this work, the problem of an average duration of a random search for one object by another is regarded, depending on various factors on a square field. The problem solution was carried out by holding total experiment with 4 factors and orthogonal plan with 54 lines. Within each line, the initial conditions and the cellular automaton transition rules were simulated and the duration of the search for one object by another was measured. As a result, the regression model of average duration of a random search for an object depending on the four factors considered, specifying the initial positions of two objects, the conditions of their movement and detection is constructed. The most significant factors among the factors considered in the work that determine the average search time are determined. An interpretation is carried out in the problem of random search for an object from the constructed model. The important result of the work is that the qualitative and quantitative influence of initial positions of objects, the size of the lattice and the transition rules on the average duration of search is revealed by means of model obtained. It is shown that the initial neighborhood of objects on the lattice does not guarantee a quick search, if each of them moves. In addition, it is quantitatively estimated how many times the average time of searching for an object can increase or decrease with increasing the speed of the searching object by 1 unit, and also with increasing the field size by 1 unit, with different initial positions of the two objects. The exponential nature of the growth in the number of steps for searching for an object with an increase in the lattice size for other fixed factors is revealed. The conditions for the greatest increase in the average search duration are found: the maximum distance of objects in combination with the immobility of one of them when the field size is changed by 1 unit. (that is, for example, with $4 \times 4$ at $5 \times 5$) can increase the average search duration in $e^{1.69} \approx 5.42$. The task presented in the work may be relevant from the point of view of application both in the landmark for ensuring the security of the state, and, for example, in the theory of mass service.

Gradient-free optimization methods or zeroth-order methods are widely used in training neural networks, reinforcement learning, as well as in industrial tasks where only the values of a function at a point are available (working with non-analytical functions). In particular, the method of error back propagation in PyTorch works exactly on this principle. There is a well-known fact that computer calculations use heuristics of floating-point numbers, and because of this, the problem of finiteness of the mantissa arises.

In this paper, firstly, we reviewed the most popular methods of gradient approximation: Finite forward/central difference (FFD/FCD), Forward/Central wise component (FWC/CWC), Forward/Central randomization on $l_2$ sphere (FSSG2/CFFG2); secondly, we described current theoretical representations of the noise introduced by the inaccuracy of calculating the function at a point: adversarial noise, random noise; thirdly, we conducted a series of experiments on frequently encountered classes of problems, such as quadratic problem, logistic regression, SVM, to try to determine whether the real nature of machine noise corresponds to the existing theory. It turned out that in reality (at least for those classes of problems that were considered in this paper), machine noise turned out to be something between adversarial noise and random, and therefore the current theory about the influence of the mantissa limb on the search for the optimum in gradient-free optimization problems requires some adjustment.

We consider one of the problems of transport modelling — searching the equilibrium distribution of traffic flows in the network. We use the classic Beckman’s model to describe time costs and flow distribution in the network represented by directed graph. Meanwhile agents’ behavior is not completely rational, what is described by the introduction of Markov logit dynamics: any driver selects a route randomly according to the Gibbs’ distribution taking into account current time costs on the edges of the graph. Thus, the problem is reduced to searching of the stationary distribution for this dynamics which is a stochastic Nash – Wardrope equilibrium in the corresponding population congestion game in the transport network. Since the game is potential, this problem is equivalent to the problem of minimization of some functional over flows distribution. The stochasticity is reflected in the appearance of the entropy regularization, in contrast to non-stochastic case. The dual problem is constructed to obtain a solution of the optimization problem. The universal primal-dual gradient method is applied. A major specificity of this method lies in an adaptive adjustment to the local smoothness of the problem, what is most important in case of the complex structure of the objective function and an inability to obtain a prior smoothness bound with acceptable accuracy. Such a situation occurs in the considered problem since the properties of the function strongly depend on the transport graph, on which we do not impose strong restrictions. The article describes the algorithm including the numerical differentiation for calculation of the objective function value and gradient. In addition, the paper represents a theoretical estimate of time complexity of the algorithm and the results of numerical experiments conducted on a small American town.

In this paper, we propose an adaptive algorithm that simulates the process of forming the initial behavioral skills on the example of the system ‘eye-arm’ animat. The situation is the formation of the initial behavioral skills occurs, for example, when a child masters the management of their hands by understanding the relationship between baseline unidentified spots on the retina of his eye and the position of the real object. Since the body control skills are not ‘hardcoded’ initially in the brain and the spinal cord at the level of instincts, the human child, like most young of other mammals, it is necessary to develop these skills in search behavior mode. Exploratory behavior begins with trial and error and then its contribution is gradually reduced as the development of the body and its environment. Since the correct behavior patterns at this stage of development of the organism does not exist for now, then the only way to select the right skills is a positive reinforcement to achieve the objective. A key feature of the proposed algorithm is to fix in the imprinting mode, only the final action that led to success, and that is very important, led to the familiar imprinted situation clearly leads to success. Over time, the continuous chain is lengthened right action — maximum use of previous positive experiences and negative ‘forgotten’ and not used.

Thus there is the gradual replacement of the random search purposeful actions that observed in the real young. Thus, the algorithm is able to establish a correspondence between the laws of the world and the ‘inner feelings’, the internal state of the animat. The proposed animat model was used 2 types of neural networks: 1) neural network NET1 to the input current which is fed to the position of the brush arms and the target point, and the output of motor commands, directing ‘brush’ manipulator animat to the target point; 2) neural network NET2 is received at the input of target coordinates and the current coordinates of the ‘brush’ and the output value is formed likelihood that the animat already ‘know’ this situation, and he ‘knows’ how to react to it. With this architecture at the animat has to rely on the ‘experience’ of neural networks to recognize situations where the response from NET2 network of close to 1, and on the other hand, run a random search, when the experience of functioning in this area of the visual field in animat not (response NET2 close to 0).

The article proposes a method of joint analysis of multidimensional financial time series based on the evaluation of the set of properties of stock quotes in a sliding time window and the subsequent averaging of property values for all analyzed companies. The main purpose of the analysis is to construct measures of joint behavior of time series reacting to the occurrence of a synchronous or coherent component. The coherence of the behavior of the characteristics of a complex system is an important feature that makes it possible to evaluate the approach of the system to sharp changes in its state. The basis for the search for precursors of sharp changes is the general idea of increasing the correlation of random fluctuations of the system parameters as it approaches the critical state. The increments in time series of stock values have a pronounced chaotic character and have a large amplitude of individual noises, against which a weak common signal can be detected only on the basis of its correlation in different scalar components of a multidimensional time series. It is known that classical methods of analysis based on the use of correlations between neighboring samples are ineffective in the processing of financial time series, since from the point of view of the correlation theory of random processes, increments in the value of shares formally have all the attributes of white noise (in particular, the “flat spectrum” and “delta-shaped” autocorrelation function). In connection with this, it is proposed to go from analyzing the initial signals to examining the sequences of their nonlinear properties calculated in time fragments of small length. As such properties, the entropy of the wavelet coefficients is used in the decomposition into the Daubechies basis, the multifractal parameters and the autoregressive measure of signal nonstationarity. Measures of synchronous behavior of time series properties in a sliding time window are constructed using the principal component method, moduli values of all pairwise correlation coefficients, and a multiple spectral coherence measure that is a generalization of the quadratic coherence spectrum between two signals. The shares of 16 large Russian companies from the beginning of 2010 to the end of 2016 were studied. Using the proposed method, two synchronization time intervals of the Russian stock market were identified: from mid-December 2013 to mid- March 2014 and from mid-October 2014 to mid-January 2016.

In this paper, we discuss the procedure for finding nominal trajectories of the planar five-link bipedal robot with point contact. To this end we use a virtual constraints method that transforms robot’s dynamics to a lowdimensional zero manifold; we also use a nonlinear optimization algorithms to find virtual constraints parameters that minimize robot’s cost of transportation. We analyzed the effect of the degree of Bezier polynomials that approximate the virtual constraints and continuity of the torques on the cost of transportation. Based on numerical results we found that it is sufficient to consider polynomials with degrees between five and six, as further increase in the degree of polynomial results in increased computation time while it does not guarantee reduction of the cost of transportation. Moreover, it was shown that introduction of torque continuity constraints does not lead to significant increase of the objective function and makes the gait more implementable on a real robot.

We propose a two step procedure for finding minimum of the considered optimization problem with objective function in the form of cost of transportation and with high number of constraints. During the first step we solve a feasibility problem: remove cost function (set it to zero) and search for feasible solution in the parameter space. During the second step we introduce the objective function and use the solution found in the first step as initial guess. For the first step we put forward an algorithm for finding initial guess that considerably reduced optimization time of the first step (down to 3–4 seconds) compared to random initialization. Comparison of the objective function of the solutions found during the first and second steps showed that on average during the second step objective function was reduced twofold, even though overall computation time increased significantly.

Finding PageRank vector is of great scientific and practical interest due to its applicability to modern search engines. Despite the fact that this problem is reduced to finding the eigenvector of the stochastic matrix $P$, the need for new algorithms is justified by a large size of the input data. To achieve no more than linear execution time, various randomized methods have been proposed, returning the expected result only with some probability close enough to one. We will consider two of them by reducing the problem of calculating the PageRank vector to the problem of finding equilibrium in an antagonistic matrix game, which is then solved using the Grigoriadis – Khachiyan algorithm. This implementation works effectively under the assumption of sparsity of the input matrix. As far as we know, there are no successful implementations of neither the Grigoriadis – Khachiyan algorithm nor its application to the task of calculating the PageRank vector. The purpose of this paper is to fill this gap. The article describes an algorithm giving pseudocode and some details of the implementation. In addition, it discusses another randomized method of calculating the PageRank vector, namely, Markov chain Monte Carlo (MCMC), in order to compare the results of these algorithms on matrices with different values of the spectral gap. The latter is of particular interest, since the magnitude of the spectral gap strongly affects the convergence rate of MCMC and does not affect the other two approaches at all. The comparison was carried out on two types of generated graphs: chains and $d$-dimensional cubes. The experiments, as predicted by the theory, demonstrated the effectiveness of the Grigoriadis – Khachiyan algorithm in comparison with MCMC for sparse graphs with a small spectral gap value. The written code is publicly available, so everyone can reproduce the results themselves or use this implementation for their own needs. The work has a purely practical orientation, no theoretical results were obtained.

Technical diagnosis’ current trends are connected to application of FEM computer simulation, which allows, to some extent, replace real experiments, reduce costs for investigation and minimize risks. Computer simulation, just at the stage of research and development, allows carrying out of diagnostics of equipment to detect permissible fluctuations of parameters of equipment’s work. Peculiarity of diagnosis of rolling equipment is that functioning of rolling equipment is directly tied with manufacturing of product with required quality, including accuracy. At that design of techniques of technical diagnosis and diagnostical modelling is very important. Computer simulation of cold rolling of strip was carried out. At that upper working roll was doing vibrations in horizontal direction according with published data of experiments on continuous 1700 rolling mill. Vibration of working roll in a stand appeared due to gap between roll’s craft and guide in a stand and led to periodical fluctuations of strip’s thickness. After computer simulation with the help of DEFORM software strip with longitudinal and transversal thickness variation was gotten. Visualization of strip’s geometrical parameters, according with simulation data, corresponded to type of inhomogeneity of surface of strip rolled in real. Further analysis of thickness variation was done in order to identify, on the basis of simulation, sources of periodical components of strip’s thickness, whose reasons are malfunctions of equipment. Advantage of computer simulation while searching the sources of forming of thickness variation is that different hypothesis concerning thickness formations may be tested without conducting real experiments and costs of different types may be reduced. Moreover, while simulation, initial strip’s thickness will not have fluctuations as opposed to industrial or laboratorial experiments. On the basis of spectral analysis of random process, it was established that frequency of changing of strip’s thickness after rolling in one stand coincides with frequency of working roll’s vibration. Results of computer simulation correlate with results of the researches for 1700 mill. Therefore, opportunity to apply computer simulation to find reasons of formation of thickness variation of strip on the industrial rolling mill is shown.

Social media is a crucial indicator of the position of assets in the financial market. The paper describes the rigid solution for the classification problem to determine the influence of social media activity on financial market movements. Reputable crypto traders influencers are selected. Twitter posts packages are used as data. The methods of text, which are characterized by the numerous use of slang words and abbreviations, and preprocessing consist in lemmatization of Stanza and the use of regular expressions. A word is considered as an element of a vector of a data unit in the course of solving the problem of binary classification. The best markup parameters for processing Binance candles are searched for. Methods of feature selection, which is necessary for a precise description of text data and the subsequent process of establishing dependence, are represented by machine learning and statistical analysis. First, the feature selection is used based on the information criterion. This approach is implemented in a random forest model and is relevant for the task of feature selection for splitting nodes in a decision tree. The second one is based on the rigid compilation of a binary vector during a rough check of the presence or absence of a word in the package and counting the sum of the elements of this vector. Then a decision is made depending on the superiority of this sum over the threshold value that is predetermined previously by analyzing the frequency distribution of mentions of the word. The algorithm used to solve the problem was named benchmark and analyzed as a tool. Similar algorithms are often used in automated trading strategies. In the course of the study, observations of the influence of frequently occurring words, which are used as a basis of dimension 2 and 3 in vectorization, are described as well.