Neural networks: practical application. What in Russia teach neural networks neural networks in economics and business

neural network intelligence artificial business

Neural networks can be implemented by software or hardware.

Options hardwareimplementation are neurocomputers, neuroplates and neurobis (large integrated circuits). One of the simplest and cheap neurobis is a Micro Devices MD 1220 model, which implements a network with 8 neurons and 120 synapses. Among promising developments, you can allocate Adaptive Solutions (USA) and Hitachi (Japan). The neurobis is one of the most high-speed: declared processing speed is 1.2 billion interneurone compounds per second (MHC / s). Schemes produced by Hitachi, make it possible to implement artificial neural networks containing up to 576 neurons.

Most modern neurocomputers are a personal computer or workstation, which includes additional neuroplates. These include, for example, FUJITSU FMR series computers. The possibilities of such systems are quite enough to solve a large number of applied problems with neuromathematics methods, as well as to develop new algorithms. Specialized neurocomputers are of greatest interest, in which the principles of the neural network architecture are implemented. Typical representatives of such systems are the computers of the MARK family of the company TRW (the first implementation of the perceptron, developed by F. Rosenblat, was called Mark I). The MARK III model of the TRW is a workstation containing up to 15 Motorola 68000 family processors with mathematical coprocessors. All VME bus combined processors. The system architecture that supports up to 65,000 virtual processor elements with more cell 1 million compounds allows you to process up to 450 thousand MNS / s.

Another example is NetSim's neurocomputer created by Texas Instruments based on the development of the University of Cambridge. Its topology is a three-dimensional grid of standard computing nodes based on processors 80188. Netsim is used to simulate Hopfield-Kohonen networks. Its productivity reaches 450 million MNS / s.

In cases where the development or implementation of hardware implementations of neural networks is too expensive, it is used cheaper software implementations.One of the most common software products is a program family. Brainmaker.cSS (CALIFORNIA SCIENTFIC SOFTWARE). Originally Developed by the company Loral Space Systems on request NASA and Johnson Space Center Package Brainmaker.the ball is soon adapted for commercial applications and today is used by several thousand financial and industrial companies, as well as US defense departments to solve the problems of forecasting, optimization and modeling situations.

Package assignment Brainmaker.- Solving tasks for which the formal methods and algorithms have not yet been found, and the input data is incomplete, roaming and contradictory. Such tasks include forecasting exchange rates and stocks on stock exchanges, modeling crisis situations, image recognition and many others. Brainmaker.solves the task, using the mathematical apparatus of the theory of neural networks (more specifically, the Hopfield network with training on the method of reverse dissemination of an error). A model of a multilayer neural network is built in RAM, which has a property of learning on a variety of examples, optimizing its internal structure. With the right choice of the network structure after its training on a sufficiently large number of examples, you can achieve high reliability of the results (97% and higher). There are versions Brainmaker.for MS DOS and MS Windows, as well as for Apple Macintosh. In addition to the basic version of the package in the family Brainmaker.the following additions include:

Brainmaker Student.- package version for universities. It is especially popular in small firms specializing in the creation of applications and for not very complex tasks.

Toolkit Option- A set of three additional programs increasing the possibilities Brainmaker, Binary,which translates the learning information into a binary format to accelerate training; Hypersonic Training,where high-speed learning algorithm is used; Plotting which displays facts, statistics and other data in graphical form.

Brainmaker Professional- Professional package version Brainmaker.with advanced functionality. Includes all options Toolkit.

GENETIC TRAINING OPTION(for package Brainmaker Pro) - The automatic optimization program of the neural network to solve the specified class of tasks using genetic algorithms for the presence of the best solutions.

Datmaker Editor- Specialized editor for automating data preparation when setting up and using a neural network.

TRAINING FINANCIAL DATA.- specialized data sets to configure the neural network to various types of analytical, commercial and financial operations, which include the real values \u200b\u200bof macroeconomic indicators NYSE, NADDAW, ASE, OEX, DOW, etc., inflation indices, statistical data of stock reports on various types of products, and Also information on futures contracts and much more.

Brainmaker Accelerator- Specialized neuroplated accelerator based on TMS320C25 signal processors of TEXAS INSTRUMENTS. Inserted into a personal computer, it accelerates the package work several times Brainmaker.

Brainmaker Accelerator Pro.- Professional multiprocessor neural board. It contains five TMS320C30 signaling processors and 32 MB of RAM.

Currently, the software market has a large number of diverse packages for designing neural networks and solving various tasks. Package Brainmaker.you can call the veteran of the market. In addition to representatives of this family, you can attribute to well-known and common software. Neurroshell.(WardSystem "S Group), Neuro Works.(Neural Ware Inc.) and NeurosolutionsNeurodimension Inc.). Object-oriented Family Environment Programs Neurosolutionsdesigned to simulate an artificial neural network of an arbitrary structure. User Systems Neurosolutionsthe possibilities of research and dialogue management are provided. All data on the network are available for viewing in the learning process through a variety of visualization tools. Designing an artificial neural network in the system Neurosolutionsbased on the modular principle, which allows you to simulate standard and new topologies. An important advantage of the system is the presence of special tools that allow modeling dynamic processes in an artificial neural network.

The use of neural network technologies is advisable when solving tasks with the following signs:

The absence of algorithms for solving problems in the presence of a sufficiently large number of parameters;

The presence of a large amount of input information characterizing the problem under study;

Noise, partial inconsistency, infertion or redundancy of the source data.

Neural technologies have been widely used in such directions as printing text recognition, product quality control in production, identification of events in particle accelerators, oil exploration, drug control, medical and military applications, management and optimization, financial analysis, prediction, etc.

In the economy, neural network technologies can be used to classify and analyze temporary series by approximation of complex nonlinear functions. It is experimentally established that the model of neural networks provide greater accuracy in identifying nonlinear patterns in the stock market compared with regression models.

Neural technologies are actively used in marketing to simulate customer behavior and distribution of market share. Neural technologies allow you to find hidden patterns in marketing databases.

Simulation of customer behavior allows you to determine the characteristics of people who will need to react to advertising and make purchases of a specific product or service.

Segmentation and modeling of markets based on neural technologies makes it possible to build flexible classification systems capable of segmenting markets, taking into account the variety of factors and features of each client.

Technologies of artificial neural networks have good prospects in solving problems of imitation and prediction of behavioral characteristics of managers and risk prediction tasks when issuing loans. No less important to the use of artificial neural networks when choosing customers for mortgage lending, predictions of bank customer bankruptcy, the definition of fraudulent transactions when using credit cards, drawing up customer ratings for loans with fixed payments, etc.

It should be remembered that the use of neural network technologies is not always possible and associated with certain problems and disadvantages.

1. It is necessary at least 50, and better 100 observations to create an acceptable model. This is a fairly large number of data, and they are not always available. For example, in the production of seasonal goods, the history of previous seasons is not enough to forecast for the current season due to changes in product style, sales policies, etc. Even in predicting demand for a fairly stable product based on monthly sales information, it is difficult to accumulate historical data from 50 to 100 months. For seasonal goods, the problem is even more difficult, as each season actually represents one observation. With the deficiency of the information model of artificial neural networks, they build in conditions of incomplete data, and then conduct their sequential refinement.

2. Building neural networks requires significant labor and time costs to obtain a satisfactory model. It must be borne in mind that there is no high accuracy obtained on the training sample, may result in the instability of the results on the test sample - in this case there is a "retraining" network. The better the system is adapted to specific conditions, the less it is capable of generalizing and extrapolation and, sooner it may be inoperable when changing these conditions. Expanding the volume of the training sample allows you to achieve greater stability, but by increasing learning time.

3. When teaching neural networks, traps may occur associated with entering local minima. The deterministic learning algorithm is unable to detect the global extremum or leave the local minimum. One of the techniques that allows you to bypass "traps", is the expansion of the dimension of the space by increasing the number of neurons of the hidden layers. Some opportunities for solving this problem open stochastic learning methods. When modifying the weighs of the network only on the basis of information on the direction of the gradient vector of the target function in the space of weights, it is possible to reach a local minimum, but it is impossible to get out of it, since at the extremum point "Driving force" (gradient) adds to zero and the cause of the movement disappears. To leave the local extremum and go to the search for global extremum, you need to create an additional force that will depend not on the gradient of the target function, but from some other factors. One simple method is to simply create a random force and add it to deterministic.

4. The sigmoidal nature of the gear ratio of the neuron is the reason that if in the learning process several weights become too large, the neuron enters the horizontal portion of the function in the saturation area. At the same time, changes in other scales, even large enough, practically does not affect the magnitude of the output signal of such a neuron, and hence the value of the target function.

5. A bad selection of the input variables range is a fairly elementary, but often error performed. If it is a binary variable with a value of 0 and 1, then at about half of the cases it will have a zero value: \u003d 0. Since it is included in the expression to modify the weight in the form of a factor, the effect will be the same as during saturation: modification of the respective scales will be blocked. The correct range for input variables should be symmetric, for example from +1 to -1.

6. The process of solving problems of the neural network is "opaque" for the user, which can cause distrust of the network predictive abilities on its side.

7. The predictive network ability is significantly reduced if facts incoming to the input (data) have significant differences from the examples on which the network was trained. This deficiency is clearly manifested in solving the tasks of economic forecasting, in particular when determining the trends of securities quotations and currency value in stock and financial markets.

8. There are no theoretically substantiated rules for designing and effective learning neural networks. This disadvantage leads, in particular, to the loss of neural networks of the ability to generalize the data of the subject area in the states of retraining (overralling).

Send your good work in the knowledge base is simple. Use the form below

Students, graduate students, young scientists who use the knowledge base in their studies and work will be very grateful to you.

• Introduction
• Conclusion
• Introduction
• The beneficial effect on the development of neural technologies was the creation of methods of parallel information processing.
• It is necessary to express appreciation to a wonderful surgeon, philosopher and cybernetic N.M. Amosov, together with students to systematizing an approach to the creation of artificial intelligence (AI). This approach is as follows.
• The basis of the Strategy of AI lies the concept of paradigm - a look (conceptual presentation) to the essence of the problem or the task and the principle of its decision. Consider two paradigms of artificial intelligence.
• 1. The paradigm of the expert assumes the following objects, as well as the stages of the development and functioning of the AI \u200b\u200bsystem:
• * Formatting knowledge - transformation by a problem knowledge expert into a form prescribed by the selected knowledge presentation model;
• * Knowledge Base Formation<БЗ) - вложение формализованных знаний в программную систему;
• * Deduction - solving the problem of logical output based on the BZ.
• This paradigm underlies the application of expert systems, logical output systems, including in the language of logical programming Prolog. It is believed that systems based on this paradigm are more studied.
• 2. Paradigm of the student, including the following provisions and sequence of actions:
• * Processing of observations, study of the experience of private examples - database formation<БД> Systems of the AI;
• * Inductive training - the transformation of the database in the BZ based on the generalization of knowledge accumulated in the database. and the rationale for the procedure for extracting knowledge from the BZ. This means that on the basis of data it is concluded about the generality of the dependence between the objects we observe. The main focus is on the study of approximating, probabilistic and logical mechanisms for obtaining general conclusions from private allegations. We can then substantiate, for example, the sufficiency of the generalized interpolation procedure (extrapolation), or the associative search procedure, with which we will satisfy requests to the BZ;
• * Deduction - According to the reasonable or alleged procedure, we select information from the BZ on request (for example, the optimal management strategy for the vector characterizing the current situation).
• Studies within the framework of this paradigm and its development was carried out so far, although they are at the heart of the construction of self-learning management systems (a remarkable example of a self-learning control system will be given below - shooting rules in artillery).
• What is the knowledge base, the overall and mandatory element of the AI \u200b\u200bsystem, differs from the database? The possibility of logical output!
• Now let's turn to the "natural" intelligence. Nature did not create anything better than the human brain. It means that the brain is both the carrier of the knowledge base, and the means of logical conclusion on its basis, regardless of which paradigm, we organized their thinking, that is, in which we fill in the knowledge base. - Learn!
• YES. Pospelov in a wonderful one in its kind, work illuminates the highest spheres of artificial intelligence - the logic of thinking. The purpose of this book is to at least partially dispersed to neurallet as a means of thinking, thereby attracting attention to the lowest, initial link of the entire chain of artificial intelligence methods.
• Throwing mysticism, we recognize that the brain is a neural network, neurosette, - neurons interconnected, with many inputs and the only way out each. Neuron realizes a fairly simple gear ratio that allows you to convert excitation at the inputs, taking into account the weights of the inputs, to the excitation value at the outlet of the neuron. A functionally finished fragment of the brain has an input layer of neurons - receptors excited from the outside, and the output layer, the neurons of which are excited depending on the configuration and the magnitude of the excitation of the input layer neurons. It is assumed to be neural. The imitating work of the brain, processes not the data itself, and their accuracy, or, in the generally accepted sense, weight, evaluation of this data. For most continuous or discrete data, their task is reduced to the indication of the probability of the ranges that belong to their values. For a large class of discrete data - sets of sets - it is advisable to fasten the neurons of the inlet layer.

1. Experience in the use of neural networks in economic tasks

With the help of neural networks, we solve the task of developing algorithms for finding an analytical description of the patterns of functioning of economic facilities (enterprise, industry, region). These algorithms apply to the prediction of some "output" objects of objects. The task of neural network implementation of algorithms is solved. The use of methods of recognition of images or corresponding neural methods allows to solve some of the ureranny problems of economic and statistical modeling, increase the adequacy of mathematical models, bring them to economic reality. The use of recognition of images in a combination with regression analysis led to new types of models - classification and piecewise linear. Finding hidden dependencies in databases is the basis for modeling and knowledge processing tasks, including for an object with difficult to formalizable patterns.

The choice of the most preferred model from some of their sets can be understood either as a reference task, or as a problem of choice based on a set of rules. Practice has shown that methods based on the use of priority scales of factors and search for a model that meets the maximum weighted amount of factors leads to a biased Results. Weight is something that needs to be determined in this and is the task. Moreover, the scales are local - each of them is suitable for this particular task and this object (group of objects).

Consider the task of choosing the desired model. Suppose there is some many objects M, whose activities are aimed at achieving a certain goal. The functioning of each object is characterized by n values \u200b\u200bof signs, that is, there is a mapping F: M -\u003e Rn. Consequently, our source point is the state of the state of the economic object: x \u003d. Indicators of the quality of the functioning of the economic object: F0 (x), F1 (x), ..., Fm (x). These indicators must be within certain limits, and some of them we strive to make either minimal or maximum.

Such a general formulation may be controversial, and it is necessary to apply the apparatus of the union of contradictions and bringing the setting of the problem to the correct form agreed with economic meaning.

We order objects from the point of view of some criteria function, but the criterion is usually poorly defined, blurred and possibly contradictory.

Consider the problem of modeling empirical patterns on a limited number of experimental and observed data. The mathematical model may be a regression equation or diagnostic rule, or forecasting rule. With a small sample, more efficient recognition metol. In this case, the influence of the management of factors is taken into account by variation of the values \u200b\u200bof factors in their substitution to the equation of regularities or to the decisive rule of diagnosis and forecasting. In addition, we apply the selection of essential features and generating useful features (secondary parameters). This mathematical apparatus is needed for predicting and diagnosing states of economic objects.

Consider the neural network from the point of view of the theory of compliance structures, as on the team of neurons (individuals. Neural network as a mechanism for optimizing neurons under collective solutions is a way of coordination of individual opinions, in which a collective opinion is the right reaction to the entrance, that is, the necessary empirical dependence.

Hence the justification of the application of the Committees in the tasks of choice and diagnostics. The idea is that instead of one decisive rule, to look for a team of decisive rules, this team develops a collective decision by virtue of the procedure processing individual solutions to members of the team. The selection and diagnostic models usually lead to incomplete systems of inequalities, for which instead of solutions it is necessary to describe the generalization of the concept of solution. Such a generalization is a collective solution.

For example, the committee of the inequality system is such a set of elements that most of the wings of this set satisfies each inequality. Committees - some class of generalizations of the concept of solutions for tasks that can be both joint and incomplete. This is a class of discrete approximations for conflicting tasks, they can also be correlated with blurred solutions. The method of committees currently determines one of the directions of analysis and solving problems of effective choice of options, optimization, diagnosis and classification. We give for example the definition of one of the main committees, namely: for 0< p < 1: p - комитетом системы включений называется такой набор элементов, что каждому включению удовлетворяет более чем р - я часть этого набора.

Committees can be considered as a certain class of generalizations of the concept of solving in case of non-promotional systems of equations, inequalities and inclusions, and as a means of parallelization in solving the tasks of selecting, diagnosing and forecasting. As a generalization of the concept of solving the problem, the committee structures are sets of elements with some (but, as a rule, not all) the properties of the solution, this is the type of blurred solutions.

As a means of parallelization, the committee designs directly act in multilayer neural networks. We have shown that for learning a neural network to accurately solve the task of classification, you can apply the method of building a committee of a system of affine inequalities.

Based on the above, it can be concluded that the method of committees is associated with one of the important areas of research and numerical solution of both diagnostic tasks and options and the tasks of setting up neural networks in order to obtain the required response to their input information on one or another person accepting their solutions.

In the process of operation of the Committees, such important properties for applied properties as heuristicity, interpretability, flexibility - the possibility of adherence and reconfiguration, the possibility of using the most natural class of functions - piecewise affine, and for setting the task of classification, diagnosis and forecasting, only correctness is required, There is so that the same object is not attributed to different classes.

Another side of the issue of compliant designs is associated with the concept of coalitions in the development of collective decisions, while the situation differ sharply in the case of collective preferences (here there are many pitfalls) and in the case of the rules of collective classification, in this case the procedures can strictly substantiate and they have wider opportunities . Therefore, it is important to be able to reduce decision-making tasks and objectives for the classification tasks.

2. Tabular method - the basis of artificial intelligence

In general, the principles of brain activity are known and are actively used. We use invisible tables in our memory, forcibly and freely fill out at the desk, driving, with a ministerial portfolio and without it, turning your head on a noisy street, for the book, at the machine and in Easel. We learn, learn all your life: and a schoolboy, conducting sleepless nights behind the letter, and the professor's impended experience. For with the same tables, we associate not only decision making, but also move, go, play the ball.

If you oppose associative thinking mathematical calculations, then what are their weight in a person's life? How was the development of a person when he didn't know how to count? Using associative thinking, able to interpolate and extrapolate, man accumulated experience. (By the way, remember the thesis of D. Mendeleev: Science begins when they begin to count.) You can ask the reader: How many times did you think today? You drove a car, played tennis, hurried to the bus, planning their actions. Imagine how much would you have to calculate (and where else to take the algorithm?) In order to lift the leg on the pavement, bypassing the border? No, we do not calculate anything every minute, and this is perhaps the main thing in our intellectual life, even in science and business. Mechanisms of sensations, intuitions, the automatism that we cannot explain, address subcortical thinking, in fact they are normal mechanisms of associative thinking using the knowledge base tables.

And most importantly, we do it quickly! How we will not think, trying to comprehend and reproduce the development of generic memory, the product of growth in the development process. We believe it is quite materially embodied and therefore implemented artificially, subject to modeling and reproduction.

We now formulate sufficient, today's principle of building a neural network, as an element of AI:

1. It should be recognized that the basis of imitation of the neuro-structure of the brain is a table interpolation method.

2. Tables are filled with or according to well-known calculation algorithms, or experimentally or experts.

3. Neurality provides high table processing speeds due to the possibility of avalanche-like parallelization.

4. In addition, the neural network allows input to a table with inaccurate and incomplete data, providing an approximate response on the principle of maximum or medium similarity.

5. The problem of neural network imitation of the brain is to transform not the most source information, but estimates of this information, in the substitution of information, the values \u200b\u200bof the excitation of receptors, skillfully distributed between the types, types, parameters, the ranges of their changes, or separate values.

6. Neurons of the output layer of each substructure by their excitation indicate the appropriate solutions. At the same time, these excitation signals for initial mediated information can be used in the next logical chain link without external intervention in operation.

3. Monitoring of the banking system

The example of the brilliant use of self-organizing Kochonen cards (SOM Self-Organizing Map) for the study of the Banking System of Russia in 1999-2005

The monitoring is based on the rating assessment based on the automatic execution of one procedure: on the multidimensional vector of the parameters of the banks on the computer screen highlights. It is drawn to the fact that neural network technologies make it possible to build visual functions of many variables, as if we transform a multi-dimensional space in one-, two- or three-dimensional. For each individual study of various factors, it is necessary to build your SOM. The forecast is possible only on the basis of an analysis of the temporary range of SOM ratings. New SOM is necessary to extend the conclusions chain, with the connection of data from outside, for example, a political nature.

Such an approach is undoubtedly effective and effective. But it seems that, compared to the potential of brain neurostructures, it holds back the scope and courage of thought, does not allow to pull long chains. Parking-investigation, combine the analysis with the forecast, to promptly take into account the emerging situation and introduce new factors and experience in experts into consideration. It should be agreed that all this is subject to the brain, and we again appeal to its structures, offering a project of software monitoring system.

The structure of neural networks and ways of learning. Logic functions underlying monitoring are mainly based on the conjunction of the logical values \u200b\u200bof variables displaying the ranges of changes in parameters or bank indicators.

The following indicators are presented:

* equity;

* Salded assets;

* liquid assets;

* Communication obligations;

* deposits of the population;

* liquidity coefficient;

* budget resources.

You can expand the system of indicators:

* The volume of investments in the era of a rapidly developing economy;

* profit volume;

* Last rating and migration value;

* Executions into the Support Fund of Science and Education;

* tax deductions;

* deductions to the Pension Fund;

* deductions to the charity and cultural fund;

* participation in UNESCO programs, etc.

Such a simple type of logical function in the transition to the region of real variables indicates the sufficiency of a single-layer neural network containing the inlet layer of receptors and the output layer on which the monitoring results are formed.

When building an input layer, it is necessary to take into account not only current indicators, but also the dynamics of the rating change in past periods of time. The output layer should reflect not only the rating, but also expert recommendations, as well as other solutions and conclusions.

It is advisable to the simplest type of training - building a knowledge base that corresponds to the concept of creating a neural network for a task: directly administering by connections by operator-researcher manually - from receptors to the neurons of the output layer in accordance with causal relations. Thus, the network is created already trained.

Then the gear ratio will also be simple and based on the summation of the excitation values \u200b\u200bat the inlet of the neuron multiplied by the weight of communication:

The task of the weight of the ties of ha compared with the gross task of all weights equal to one is more expedient due to the possible desire of an operator or an expert to take into account the influence of various indicators in different degrees.

Threshold H cuts off knowingly unacceptable conclusions, simplifying further processing (for example, the foundation of the average). The creation coefficient is due to the following considerations.

The maximum value of V can achieve p. In order for the rating value in some acceptable range, for example, the excitation values \u200b\u200bmust be converted by placing to \u003d pack.

The above assumptions allow you to promptly enter changes and clarification by the operator - by the user, develop the network, introducing new factors and considering the experience. To do this, the operator is sufficient, clicking the mouse, allocate the receptor, and then the neuron of the output layer and the connection is installed! It remains only to approximately assign the weight of the introduced communication from the range.

It should be done very important remark (OBS) regarding the entire material of the book and intended to be a very attentive reader.

Earlier, considering training, we clearly classified the initial reference situations, taking the accuracy of each component equal to one. Conducting the trace and paving dynamic excitation paths, we also believed weighing bonds equal to one (or some maximum constant value). But the teacher can immediately receive an additional degree of freedom, taking into account the factors to the extent and with the scales that he will task! We make assumption that different factors influence the result in varying degrees, and such an influence will be put on the training stage forcibly.

For example, it is known that on the eve of the war, the population in a huge number purchases soap, matches and salt. So, watching this factor, you can predict the emergence of the war.

Creating a neural network for analyzing historical or social events, one or more receptors should be selected, the excitation of which corresponds to the different levels of soap purchases, salt and matches at the same time. The excitation of these receptors should be transmitted, influence (along with other factors) to the degree of excitation of the niron of the output layer corresponding to the statement soon!

Nevertheless, the intensive purchase of soap, matches and salt is necessary, but not such a sufficient condition for the onset of war. It may indicate, for example, on the rapid revival of tourism in the region of the Chief Range of the Caucasus. In words, this is not the meaning of fuzzy logic, which allows you to take into account the irreparation of the event, not a Boolean variable yes - no, and some intermediate, an indefinite, weighted type of type "affects, but not so straight, which is necessarily ...". Therefore, communication (all or some), energized from this receptor, we set equal to some estimated value of a smaller unit and corrected subsequently, which reflects the effect of excitation of the receptor on the output.

Thus, the simultaneous purchase of soap, salt and matches is taken into account twice: the procurement level will be displayed as the degree of excitation of the corresponding receptors, and the nature of the influence of purchases on withdrawal will soon! - Using scales of synappsic ties.

Agree that, when constructing single-level networks, this approach suggests itself and is very simply implemented.

The structure of the receptor screen. The main part is the scroll window, in which you can view and set the state of the receptor layer, undoubtedly not able to fit on the static screen.

The scroll window indicates indicators and their estimated values \u200b\u200bin the range for the respective receptors. These are probabilistic values \u200b\u200bbased on reliability, intuition, expert estimates. Estimates involve the coverage of several receptors. For example, the evaluation of the fact that its own capital is not then 24, not then 34, not then 42 thousand y. e., But rather, all the time 24 may lead to an approximate estimate of the specified excitation values \u200b\u200bof 0.6.0.2 and 0.2 receptors corresponding to the ranges (20 - 25], (30 - 35], (40 - 45]. Statically asked indicators are displayed on the screen, Such as a rating resulting from past measurements, selective indicators previously found, as well as indicators of political, social and economic conditions. (their abundance and development can still require scrolling.)

You should also display the scrolling and the main operations menu:

* Go to the output layer screen;

* statistical processing of results (involves switching to the output screen);

* introduction of a new connection;

* introduction of a new receptor;

* Introduction of a new output layer neuron (involves switching screens);

* Introduction of a new indicator, etc.

The structure of the output layer screen. The output layer screen (Fig. 8.3) displays a system of concentric (embedded) rectangles or other flat figures reflecting the distribution of the rating of descending. In the center of the screen, bright dots marked the most prosperous banks or alleged ideal images. Each screen element rigidly corresponds to the neuron of the output layer. As a result of the monitoring, the neuron corresponding to the standard can most excite as possible, however, the point of the screen will be highlighted, which does not coincide with any standard, which is intermediate or averaged.

Fig. - 8.3. Output layer screen

Undoubtedly, a menu should be provided for an averaged rating assessment operation, demonstration of the success of the success, issuing warning signals, conclusion texts, recommended development strategies, data conservation for further development, etc.

Training neuralopet. To teach a neural network based on expert assessments, the ranges of permissible parameters should be specified, allowing the bank to be perfectly succeeding that has a maximum rating. Fixing several points whose coordinates (many parameter values) satisfy the values \u200b\u200bof the rating for well-known or alleged (taking into account the possible options) of banks, you can get several ideal representatives. The corresponding neurons, i.e. The elements of the output layer screen are isolated arbitrarily dispersed by the screens area. It is desirable that the standards with a large rating range closer to the center.

Next, go to the same filling of the covering rectangle, based on the following rating category, etc. to outsider banks.

To carry out such work by experts, the table is pre-formed (Table 1).

Neurons displaying banks, on the screen correspond to the magnitudes of their excitation - ratings.

Method of monitoring. The trained system that enters the user after a highly qualified expertise of economists and politicians is ready to use within CASE-technology CASE - Computer Aided Software Engineering.

Table 1 - Expert Estimates for Learning Neuraletas

At the same time, the user implements its right to additional training, clarification (for example, ties weights, to strengthen or weaken the influence of certain indicators based on their own experience), the introduction of additional indicators for the experiment to its risk, etc.

Suppose a user explores the situation that has developed around the bank "Invest-Tuda-and-back". Naturally, it does not have any satisfactory information on the feasibility of its own investments and therefore proceeds with a scrupulous data collection, resulting in approximate, probable, contradictory characteristics for modeling.

Using the receptor screen, the user specifies the values \u200b\u200bof their excitation based on quite reliable data, but sometimes considering options or - or (partially exciting different receptors), sometimes on TV, sometimes just skipping the indicators. Such indicators as the rating in the past and migration are still unknown, but the result obtained is supposed to be used in the future.

After entering data on the screen of the output layer, the bright point near the area of \u200b\u200boutsiders eloquently testifies to the protection of civil rights of the non-violent choice of a decision on the feasibility of investing the righteous accumulated capital.

The coordinates of this point on the screen are determined by the well-known formula of the middle coordinates of the coordinates of the neurons of those banks that are close to the controlled bank, and in terms of their initiation values. But on the same formulas based on the ratings of the banks of the banks there is a rating of the Bank's studied!

The user may decide on the addition of the knowledge base and, consequently, the neural network of information about the new bank, which is advisable if the Council of Experts subjected to a significant criticism of the resulting result and indicate a neural network error. Just use the option. To complement, as a result of which the computer dialog with the user is initiated:

- You want to change the rating - yes.

- new rating value --...

- Save!

Then the neuron of the output layer with the found coordinates is put in compliance with the new jar. His relationships are formed with the receptors that exciteed when entering information about the bank. The weight of each connection is supposed to be equal to the user entered by the excitation of the corresponding neuron receptor. Now the knowledge base is supplemented in the same way as the list of targeted installations of the artillery battery after the defeat of the next target.

However, a significant compulsory change in the rating may require the movement of the released point into the area of \u200b\u200bbanks with the corresponding level of rating, i.e. It is necessary for this bank to consolidate the other neuron of the output layer, in another area of \u200b\u200bthe screen. It is also established as a result of a computer dialogue with the user.

Adjustment and development. We have already mentioned the need and the possibility of constant clarification and development of neural networks. You can change the idea of \u200b\u200bthe advancement of the standard bank (real or ideal) and complement the knowledge base, i.e. This neurallet. We can adjust the weights of the links as a measure of the effect of individual indicators for the output result.

You can enter new indicators with their weights, consider new solutions and establish the degree of influence on them of the same or new indicators. You can adapt to neural network to solve adjacent tasks, taking into account the influence of individual indicators for the migration of banks (the transition from one rating level to another), etc.

Finally, you can, by purchasing this software product with a friendly interface and an excellent service, with a developed set of neural network conversion functions, remake it for a completely different task, for example, for a fascinating game in the railway roulette, on which we intend to stay below.

In conclusion, we note that in the economy and business, as well as in the management of complex objects, the decision-making systems are dominated, where each situation is formed on the basis of a constant number of factors. Each factor is represented by a variant or value of an exhaustive set, i.e. Each situation seems to be conjunction in which the statements regarding all the factors are necessarily involved in the neural network. Then all the conjunctions (situations) have the same number of statements. If in this case, two excellent situations lead to different solutions corresponding to the neural network is perfect. The attractiveness of such neural network lies in their depositibility to one-layer. If we reproduce the solutions (see subdaz. 5.2), we will get a perfect to neurallet (without feedback).

To build a perfect neural network, you can reduce the task of this section, subdrade. 6.2, as well as, for example, the task of assessing country risk and others.

Conclusion

The distribution of the excitation of the neurons of the output layer, and the bowl of all the neuron, which has the maximum excitation value, allows you to set the correspondence between the combination and the values \u200b\u200bof the excitations on the inlet layer (the image on the retina) and the response received (which is). Thus, this dependence and determines the possibility of logical output of the form "if something. . Those. serve the binding and memorization of relationships "Package - Corollary". The connection of the substructures of the neural network allows you to receive "long" logical chains based on such relationships.

From here it follows that the network works in two modes: in the training mode and in recognition mode (operating mode).

In the training mode, logical chains are generated.

In the recognition mode, the neural network on the presented image with high reliability determines how the type it refers, what actions should be taken, etc.

It is believed that in the human brain to 100 billion neurons. But now we are not interested in how neurons are arranged, in which up to 240 chemical reactions. We are interested in, neuron works on a logical level, as it performs logical functions. The implementation of only these functions should be the basis and means of artificial intelligence. Improving these logical functions, we are ready to violate the main laws of physics, such as the law of energy conservation. After all, we count not on physical modeling, but on an affordable, universal - computer.

So, we focus on the "(direct" use of neural networks in the tasks of artificial intelligence. However, their application applies to solutions and other tasks. For this, neural network models are built with a structure-oriented structure), use a special system of neuropow-like elements, a certain fork Transfer function (often use so-called sigmoillers based on the participation of the exhibitors in the formation of a gear ratio), specially selected and dynamically refined weights. In this case, the properties of the convergence of the excitation of neurons, self-optimization are used. When the input of the excitation vector is supplied through a certain number of clocks of the neural network, the values \u200b\u200bof the excitation of the output layer neurons (in some models, all the neurons of the inlet layer are neurons of the output layer and there are no other) converge to some values. They may indicate, for example, to which the standard is more like a "roaring". Unreliable input image, or that. How to find a solution to some task. For example, the well-known network of Hopfield. Although with restrictions, it can solve the task of the community - the task of exponential complexity. The chamming network successfully implements associative memory. Coonena network (Kohonen Maps), added 06/27/2011

The task of analyzing business activity, factors affecting decision-making. Modern information technologies and neural networks: principles of their work. Investigation of the use of neural networks in the objectives of forecasting financial situations and decision-making.

thesis, added 11/06/2011

Description of the technological process of paper science. Design paper machine. Justification of the use of neural networks in the management of molding paper web. Mathematical model of neuron. Simulation of two structures of neural networks.

coursework, added 15.10.2012

Ways to apply neural network technologies in intrusion detection systems. Expert system detection of network attacks. Artificial networks, genetic algorithms. Advantages and disadvantages of intrusion detection systems based on neural networks.

examination, added 30.11.2015

The concept of artificial intelligence as the properties of automatic systems to take on individual functions of human intelligence. Expert systems in the field of medicine. Various approaches to building artificial intelligence systems. Creating neural networks.

presentation, added 05/28/2015

Study of the task and prospects for using neural networks on radial-basic functions to predict the main economic indicators: a gross domestic product, the national income of Ukraine and the consumer price index. Assessment of results.

course work, added 14.12.2014

The concept and properties of artificial neural networks, their functional similarity with the human brain, the principle of their work, the area of \u200b\u200buse. Expert system and reliability of neural networks. Model of artificial neuron with activation function.

abstract, added 03/16/2011

The essence and functions of artificial neural networks (INS), their classification. Structural elements of artificial neuron. Differences between Ins and Machines with Neumanan architecture. Building and learning these networks, areas and prospects for their use.

presentation, added 14.10.2013

The use of neurocomputers in the Russian financial market. Prediction of temporary series based on neural processing methods. Determination of courses of bonds and shares of enterprises. Application of neural networks to the tasks of analysis of stock activity.

course work, added 05/28/2009

The history of the creation and main directions in modeling artificial intelligence. Problems of learning to visual perception and recognition. Development of elements of intelligence robots. Research in the field of neural networks. Principle of feedback Wiener.

UDC 004.38.032.26

O. V. Konyukhova, K. S. Lappochna

O. V. KONUKHOVA, K. S. LAPOCHKINA

The use of neural networks in the economy and the relevance of their use in the preparation of a short-term budget forecast

Application of Neural Networks in Economy and An Urgency of their Use by Drawing Up of A Short-Term Forecast Of The Budget

This article describes the use of neural networks in the economy. The process of predicting the budget of the Russian Federation and the relevance of the use of neural networks to compile a short-term budget is considered.

Keywords: economics, budget of the Russian Federation, budget prediction, neural networks, genetic algorithms.

In this article Application of Neural Networks in Economy is described. Process of Forecasting Of The Budget of The Russian Federation of Neural Networks for Drawing Up of The Short-Term Budget Is Considered.

Keywords: Economy, Budget of the Russian Federation, Budget forecasting, Neural Networks, Genetic Algorithms.

4) Automatic facilities grouping.

One of the interesting attempts to create a mechanism for rational depressive economy owned by English cybernetics Stafford Biiru. They were proposed to those who became well-known principles of management, based on neurophysiological mechanisms. Models of production systems were considered as very complex relationships between the inputs (resource threads) internal, invisible elements and outputs (results). The inputs of the models served sufficiently generalized indices, the main of which promptly reflect the amount of production of specific production, the necessary need for resources and performance. The solutions offered to effectively function by this kind of systems were taken after how all the options in this situation were found and discussed. The best solution was made by a majority vote participating in the discussion of managers and experts. To this end, the system has a situational room equipped with appropriate technical means. The approach to the creation of the management system proposed by S. Bir was effective for control not only by large manufacturing associations, such as the steel corporation, but also the economy of Chile 70s.

Similar principles were used in the group accounting method of arguments (Moscow State University) by Ukrainian cybernetic for modeling the economy of prosperous England. Together with economists (Parks, etc.), offered more than two hundred independent variables affecting the gross income, they were revealed by several (five to six) of the main factors, which with a high degree of accuracy determine the value of the output variable. Based on these models, various options for the economy were developed in order to increase economic growth in various savings standards, inflation levels and unemployment.

The proposed method of group accounting of arguments is based on the principle of self-organization of models of complex, in particular economic systems, and allows you to determine complex hidden dependencies in data that are not detected by standard statistical methods. This method was successfully used by A. and Ivakhnenko to assess the state of the economy and forecasting its development in countries such as the United States, United Kingdom, Bulgaria and Germany. Used a large number of independent variables (from fifty to two hundred), describing the state of the economy and affecting gross income in the studied countries. Based on the analysis of these variables using the method of group accounting of arguments, the main, significant factors were detected, with a high degree of accuracy determining the value of the output variable (gross income).

Studies in this direction have stimulating the impact on the development of neural network methods intensively used in recent times due to their ability to extract experience and knowledge from a small classified sequence. Neural networks after training on such sequences are able to solve complex informalizable tasks as experts are made based on their knowledge and intuition. These advantages become particularly significant in the conquering economy, for which the unevenness of the development rate is characterized, various inflation rates, a small duration, as well as incomplete and inconsistency of knowledge about the economic phenomena.

Work is widely known, which successfully applied the principles of self-organization of models of complex economic systems for building a neural network in solving problems of analyzing and modeling the development of the economy of Mordovia and the Penza region.

A characteristic example of the successful application of neural computing in the financial sector credit risk management. As you know, before issuing a loan, banks are carried out by complex statistical calculations on the financial reliability of the borrower to assess the likelihood of their own losses from the late refund of funds. Such calculations are usually based on the assessment of credit history, the dynamics of the company's development, the stability of its basic financial indicators and many other factors. One widely well-known US bank has tried the method of neural computing and concluded that the same task on the calculations of this kind is solved faster and more accurate. For example, in one of the evaluation cases of 100 thousand bank accounts, a new system, built on the basis of neural computing, has identified over 90% of potential non-payers.

Another very important area of \u200b\u200bapplication of neural computing in the financial sector prediction of the situation in the stock market. The standard approach to this task is based on a rigidly fixed set of "Rules of Games", which over time lose their effectiveness due to changes in the flow conditions on the stock exchange. In addition, systems built on the basis of this approach are too slow for situations requiring instant decision-making. That is why the main Japanese companies operating in the securities market decided to apply the method of neural computing. In the typical system based on the neural network, information was introduced by the total volume of 33 years of business activity of several organizations, including the turnover, the previous value of the shares, income levels, etc. Self-evidence at the real examples, the neural network system showed greater prediction accuracy and better speed: Compared with the statistical approach gave an improvement in performance as a whole by 19%.

One of the most advanced techniques of neural computation is genetic algorithms that imitate the evolution of living organisms. Therefore, they can be used as an optimizer of the neural network parameters. A similar system for predicting the results of contracts for long-term securities of increased reliability was developed and installed on the Sun workstation at Hill Samuel Investment Management. When modeling several bidders strategies, it achieved 57% accuracy in predicting the direction of market movement. In the insurance company TSB General Insurance (Newport) uses a similar method for predicting the level of risk when insuring private loans. This neural network is self-learning on statistical data on the state of unemployment in the country.

Despite the fact that the financial market in Russia is not yet stabilized and, arguing from a mathematical point of view, its model changes, which is due to the one hand, with the expectation of the gradual mining of the securities market and increase the share of the stock market associated with the flow of investments as domestic, so and foreign capital, and on the other - with the instability of the political course, after all, you can see the appearance of firms that need to use statistical methods other than the traditional, as well as the appearance of neuropackets on the software products and computing equipment to emulate neural networks on the IBM series computers and even specialized neuropilates on the basis of custom-made neurochipov.

In particular, in Russia, one of the first powerful neurocomputers for financial use is already successfully operating - CNAPS PC / 128 on the basis of the 4th neurobis of Alaptive Solutions. According to the company "Tora-Center", among organizations using neural networks to solve their tasks, the Central Bank, the Ministry of Emergency Situations, the Tax Inspectorate, more than 30 banks and more than 60 financial companies. Some of these organizations have already published the results of their activities in the field of neurocomputer use.

From the foregoing it follows that it is currently the use of neural networks in the preparation of a short-term budget prediction is an urgent topic for research.

In conclusion, it should be noted that the use of neural networks in all areas of human activity, including in the field of financial applications, is moving along the growing, partly as needed and due to wide opportunities for some, due to prestiges for others and due to interesting Applications for third.

BIBLIOGRAPHY

1. Federal Law of the Russian Federation of 01.01.2001 (with change of 01.01.2001) "On the state forecasting and programs of socio-economic development of the Russian Federation" [Text]

2. Bir S. Brain Firm [Text] / S. Bir. - M.: Radio and Communication, 1993. - 524 p.

3. Galyushkin, neurocomputers in financial activities [Text] /. - Novosibirsk: Science, 2002. - 215c.

4., Muller predictive models [Text] /, - Kiev: Technique, 1985. - 225 p.

5. Plogging, forecasting methods in the budget process [Text] / // Electronic magazine Corporate Finance, 2011. - № 3 (19) - P. 71 - 78.

6. Rutkovskaya M., Plinsky L. Neural networks, genetic algorithms and fuzzy systems: per. with Polish. [Text] / M. Rutkovskaya, L. Plinsky -: Hotline - Telecom, 20c.

7., Sostoris solutions on neural networks of optimal complexity [Text] /, // Automation and modern technologies, 1998. - No. 4. - P. 38-43.

Federal State Educational Institution of Higher Professional Education "State University - Training and Scientific and Production Complex", Orel

Candidate of Technical Sciences, Associate Professor, Associate Professor of the Department "Information Systems"

E-mail: ***** @ *** RU

Lapokhna Kristina Sergeevna

Federal State Educational Institution of Higher Professional Education "State University - Training and Scientific and Production Complex", Orel

Student group 11-PI (M)

But also to solve more important tasks - for example, to look for new drugs. The Village appealed to experts to find out what the features of technology are and how domestic companies and universities are used.

What is neural networks?

To understand what place the neural networks occupy in the world of artificial intelligence and how they are associated with other technologies for creating intelligent systems, let's start with definitions.

Neural networks - One of the methods of machine learning, the foundations of which originated in 1943, before the emergence of the term "artificial intelligence". Represent a mathematical model that remotely resembles the work of the nervous system of animals.

According to the Senior Researcher, Innopolis Stanislav Protasov, the closest analogue of the human brain, are convolutional neural networks, invented by mathematics, Yana Lekuna. "They underlie many applications applying for the title of artificial intelligence - for example, in FindFace or Prisma," he notes.

Machine learning - Subsection of artificial intelligence at the intersection of mathematics and computer sciences. It studies methods for building models and algorithms based on the principle of training. The machine analyzes the rain-time examples, allocates patterns, summarizes them and builds the rules with which various tasks are solved - for example, predicting the further development of events or recognition and generation of images, text and speech. In addition to the neural network, the methods of linear regression, trees of solutions and other approaches are also used here.

Artificial Intelligence - section of computer science on the creation of technological means for performing tasks machines, which were previously considered exclusively prerogative of a person, as well as the designation of such developments. The direction officially imposed in 1956.

Alexander Krinov

What can be called artificial intelligence, and what is not - the question of agreements. Humanity by and large did not come to the unequivocal wording, which is such an intelligence at all, not to mention artificially. But if we generalize what is happening, we can say that artificial intelligence is deep neural networks that decisive complex tasks at the level close to the level of a person and in one degree or another self-learning. At the same time, under self-learning, it means the ability to independently extract a beneficial signal from raw data.

What condition is the industry now?

According to the analytical agency Gartner, machine learning is now at the peak of overpriced expectations. The excitement characteristic of this stage around the new technology leads to an excessive enthusiasm, which turns into unsuccessful attempts to its ubiquitous use. It is assumed that it will be necessary to get rid of the illusions of the industry from two to five years. According to Russian experts, in a short time, neural networks will have to be tested for strength.

Sergey Nognayev

managing a portfolio of the Development Fund of the Internet Initiatives

Although scientists are engaged in formalization and development of neural networks for 70 years, two turning point in the development of this technology can be distinguished. The first - 2007, when at the University of Toronto, created algorithms for deep learning of multilayer neural networks. The second moment provoking today's boom is 2012, when researchers from the same university applied deep neural networks and won the ImageNet contest, learning to recognize objects in the photo and video with a minimum error.

Now computer facilities are enough to solve if not any, the overwhelming majority of tasks based on the neural network. Now the main obstacle is the lack of data marked. Conditionally saying that the system learned to recognize the sunset on video or photographs, she needs to rain a million shots of the sunset, indicating where it is in the frame. For example, when you upload a photo on Facebook, your friends recognize a cat in the rays of the sunset, and the social network sees a set of labels in it: "Animal", "Cat", "wooden", "floor", "evening", " Orange". Who has more data for learning, in order to neuralit and will be smarter.

Andrei Kalinin

manager "Search Mail.Ru"

Entertainment applications based on neural networks - for example, our Artisto or Vinci is just the vertex of the iceberg, and at the same time a great way to demonstrate their opportunities to a wide audience. In fact, neurosetics are able to solve a number of complex tasks. The most "hot" directions now is autopilots, voice helpers, chat bots and medicine.

Alexander Krinov

head of Computer Vision Service "Yandex"

We can say that the boom of the neural network has already come, but he has not come out on the peak. Further will only more interesting. The most promising directions today are, perhaps, computer vision, dialogue systems, text analysis, robotics, unmanned transport and generation of content - texts, images, music.

Perspective spheres for neural network

Transport

Robotics

Biotechnology

Agriculture

Internet things

Media and entertainment

Linguistics

Safety

Vlad Sershulsky

director of Microsoft Technological Cooperation Programs in Russia

Today, the neural revolution has already happened. Sometimes it is even difficult to distinguish fiction from reality. Imagine an automated combine with a variety of cameras. He makes 5 thousand pictures per minute and analyzes through the neural network, the weed before him or the pest infected with the pest, after which it decides how to do next. Fiction? No longer at all.

Boris Wolfson

headhunter Development Director

There is a certain High around the neural network and, in my opinion, a little overpriced expectations. We will also pass through the disappointment stage before you learn to effectively use them. Many breakthrough research results are not very applicable in business. In practice, it is often wiser to use other machine learning methods - for example, various algorithms based on trees of solutions. Probably, it looks not as exciting and futuristic, but these approaches are very common.

What do neural networks teach in Russia?

Market participants agree that many achievements of neural networks are still applicable only in the academic sphere. Bente, the technology is used mainly in entertainment applications, which are heated by the topic. Nevertheless, Russian developers teach a neural network and solving socially significant and business tasks. Let us dwell in detail in some directions.

Science and Medicine

The Yandex Data Analysis School participates in the Crayfis experiment in conjunction with representatives of Skolkovo, MIPT, HSE and US UCI and NYU universities. Its essence is to search for cosmic particles of ultra-high energy with smartphones. Data from cameras is transmitted by accelerated neural networks capable of fixing the traces of weakly interacting particles in the pictures.

This is not the only international experiment in which Russian specialists are involved. Scientists University Innopolis Manuel Matsar and Leonard Johard participate in the Biodynamo project. Having enlisted with the support of Intel and CERN, they want to create an experienced sample capable of reproducing a full-scale simulation of the brain bark. It is planned to improve the efficiency and efficiency of experiments in which the presence of a living human brain is required.

Professor Innopolis Yaroslav Kolodov participated in the development of a computer model capable of tens of times faster to predict the formation of protein ties. With this algorithm you can speed up the development of vaccines and medicines. In the same sphere, developers from Mail.Ru Group, Insilico Medicine and MFTi were noted. They used generative tune networks, trained to invent molecular structures, for finding substances that may be useful in various diseases - from oncology to cardiovascular diseases.

beauty and health

In 2015, the Russian company Youth Laboratories launched the first international beauty contest BEAUTY.AI. Photos of participants were evaluated by neural networks. When determining the winners, they took into account the floor, age, nationality, skin color, the symmetry of the face and the presence or absence of wrinkle users. The last factor also pushed the organizers to create a Rynkl service, allowing to track how the aging affects the skin and how different drugs are fighting with it.

Also neural networks are used in telemedicine. The Russian company "Mobile Medical Technologies", managing projects "Online Dr." and "Pediatrician 24/7", tests the bot diagnostic, which will be useful both to patients and doctors. He will be the first to tell, to which specialist to contact these or other symptoms, and the second will help determine what exactly the sick.

Optimization of business processes and advertising

The Russian Startup Leadza managed to apply a neural network for a more efficient budget distribution for advertising on Facebook and Instagram. The algorithm analyzes the results of past campaigns, builds a key metrics and based on them automatically redistributes costs so that online stores can get more customers for less.

The Guaranacam team involved machine learning technology to assess the efficiency of accommodation of goods and promotional materials in offline. The system operates on the basis of the Microsoft Azure cloud and analyzes the purchasing behavior of video surveillance cameras. Business owners receive a real-time trade status report. The project is already applied in the shopping center "Mega White Dacha".

At this successful domestic examples of using neural networks in business do not end. Logistix, experiments with technologies for creating artificial intelligence since 2006, has developed a warehouse optimization system. It is based on a student neural network that analyzes the data obtained from fitness trackers and redistributes the load between them. Now the team teaches a neural network to distinguish between marriage.

Holding "BelFingroup" went even further. His "daughter" BFG-SOFT has created a BFG-IS cloud platform, which allows you to manage an enterprise using its virtual model. The latter is automatically built on the basis of the production data system collected and not only shows how it is better to organize processes, taking into account the purposes specified, but also predicts the consequences of any changes - from replacing equipment before administering additional shifts. At the end of 2016, the development fund of the Internet initiatives decided to invest in a company of 125 million rubles.

Recruiting and personnel management

The Russian Agregator of Recruiters Stafory ends the training of a recurrent neural network that can not only give one-room answers to the questions of candidates, but also lead a full-fledged conversation with them about the vacancy stake. A team of the Superjob portal tests a service that predicts which of hundreds of the same type of the summary will be in demand by a specific employer.

Transport

The Russian developer of Cognitive Technologies intelligent systems applies neural networks to recognize vehicles, pedestrians, road signs, traffic lights and other objects entering the frame. The company also collects data for teaching a neural network for an unmanned car. We are talking about dozens of thousands of episodes, describing the reaction of drivers for certain critical situations on the roads. As a result, the system should formulate the optimal scenarios of the conduct of autobood. The same technologies are used to create smart agricultural transport.

In addition, neural networks can be used in the field of transport and in a different way. In the summer of 2016, Yandex added to him the ads of the automatic determination of the model of the machine according to its photo belonging to him. At that time, the system knew 100 brands.

Psychology and security

The Russian NTechlab startup, by Google in the International Competition of The MegaFace Benchmark Face Recognition Algorithms, used machine learning technology in the FinDFACE application. It allows you to find a person in social networks by photography. Often, users refer to the service to identify fakes, but it can be useful and law enforcement officers. With it, the identity of several criminals has already been established, including the Sitibank invaders in Moscow. Business version of FindFace.Pro is provided to companies interested in customer identification. Now the system is reassured to determine the gender, age and emotions of others, which can be useful not only when communicating with customers, but also when managing staff.

Similarly, neural networks also apply another Russian company - VisionLabs. It uses face recognition technologies to ensure security in banks and the formation of special offers for the most loyal customers of various retail points.

In the similar direction, the startup "Emotian" works. He is finalizing the system of determining the emotional state of cities. While neurallet calculates the happiest areas of publications on social networks, but in the future the company is going to take into account biometric data from cameras.

Media and creativity

One of the main players in the Russian neural network market is Yandex. The company uses machine learning not only in its search services, but also in other products. In 2015, she launched the Dzen's recommendation system, which forms a tape from news, articles, photos and video, based on the interests of a particular user. The more often it refers to the selected algorithm materials, the more accurately the neural network determines what else he may like.

In addition, Yandex is experimenting with creativity. Employees of the company have already managed to apply a neural network approach to poetry, and then

Department of Education of Moscow

GBOU Gymnasium №1503

"NEURAL NETWORKS. Their application, role and significance

In a modern and future economy "

(research)
Performed

grade 10 student

Bryzhenko Dmitry.

Leader:

Janikov Alexander Vasilyevich

Moscow

year 2013

Neural networks. Their application, role and significance in a modern and future economy
Plan:

Introduction ....................................................................................................................... Goals and objectives……………………………………………………………………………………… The concept of neural networks, their meaning ............................................................... Simplest analytical technology ........................................................ Nonlinear task ................................................................................. .. Advantages of using neural networks ............................................. The principle of operation of neural networks ................................................................. Software implementations .............................................................................. Application of neural networks ........................................................................ Forecasting changes in quotations ................................................... ... Price and production price and production ........................................................... Research factors in demand ................................................................. .. Property valuation…………………………………………………………………. Analysis of the consumer market ............................................................................. Fraud fight ...................................................................................... Text recognising…………………………………………………………………… Empirical part .................................................................................... Forecasting the course of the course USD / RUR .................................................. Assessment of the cost of real estate ............................................................ .. Disadvantages of using neural networks ........................................................ Conclusion ......................................................................................................... Bibliography……………………………………………………………………………… Applications .............................................................................................................. ...

3

The danger is not that the computer will once again begin to think like a person, but that man will try to think like a computer once.

(Sidney J. Harris)

Introduction

In the modern world, economic calculations must be very accurate, rely on the previous experience. Traditional methods, such as predicting demand for new products by public survey of the analysis of the obtained data manually, product quality analysis by testing individual instances, controlling potential risks to standard methods, slowly, but correctly departed to the background due to relatively low accuracy.

Neural networks are a new and very promising computing technology, which gives completely new approaches to the study of dynamic tasks in the economic field. Initially, neural networks opened new opportunities in the field of recognition of images, then statistical and based on the search for complex interrelationships (artificial intelligence) were added to this to find decision-making and solving problems in the field of economics.

The ability to model nonlinear processes, working with noisy data and adaptability allows you to apply neural networks to solve a wide class of tasks that cover a wide variety of areas of interest. Recognition of images, treatment of roaring or incomplete data, associative search, classification, optimization, forecast, diagnostics, process management, data segmentation, compression, complex mappings, modeling non-standard processes, speech recognition.

In the past few years, there were many software systems based on neural networks for applications in issues such as operations in the commodity market, assessment of the bankruptcy bankruptcy, creditworthiness assessment, investment control, loans.

The meaning of using neural networks in the economy is not at all to oust traditional methods or to invent a bicycle, and this is another possible means for solving problems.

The beneficial effect on the development of neural technologies was the creation of methods of parallel information processing.

The hypothesis is that neural networks are considered tool that can identify the most difficult dependencies. In my work, I want to check it out.

The practical significance of the research conducted by me is related to the fact that now not a very large number of companies use neural networks as the main tool. Therefore, with the "usual" calculation, they may make errors that can be revealed using a "neural network" approach.

I divided your work on 5 chapters. In the first chapter, I reveal the general concepts of neural networks, their meaning. In the second chapter, I cite software implementations, i.e. Programs created to work with neural networks. In chapter number 3, I cite detailed examples of using neural networks in practice. In the fourth chapter, I choose two examples and, using neural network technology, I spend the study, the results of which I describe in work.

The purpose of writing work:

• 
  Identify the need to use neural networks in the economy

Tasks:

1. 
   Figure out the system of neural networks, understand that they represent
2. 
   Determine the economic tasks that can be solved using neural networks
3. 
   Model the neural network using a software neuropacket and create a practical example using it.
4. 
   Evaluate the effectiveness of the use of neural networks in economic tasks.

1. The concept of neural networks, their meaning.

Neural networks are adaptive systems for processing and analyzing data, which are a mathematical structure that imitates some aspects of the human brain and demonstrating such capabilities such as the ability to informal learning, the ability to generalize and cluster non-classified information, the ability to independently build forecasts based on Already presented temporary series, the ability to find complex analytical dependencies.

Their essential differences from other methods, such as expert systems, is that neural networks do not need a previously known, specified model, and form it based on the information entered. Therefore, neural networks and genetic algorithms entered into practice everywhere, where it is necessary to solve problems of forecasting, classification, management - in other words, in the field of human activity, where there are poorly algorithmizable tasks, to solve whether or permanent operation of a group of qualified experts, or adaptive automation systems. What are the neural networks. Thus, neural networks can be considered complex analytical technology, i.e. A methodology that based on known algorithms allows for specified data to display the value of unknown parameters.

1.1. Simplest analytical technology

In order to be clearer, I will cite a classic example of the simplest analytical technology: the Pythagora theorem, which allows the length of the cathets to determine the length of the hypotenuse.

c 2 \u003d a 2 + b 2.

Knowing the parameters a and b, calculate C [hypotenuse] is not difficult.

1.2. "Nonlinear Task"

A completely different option of analytical technology is methods that are processed by the human brain. Examples of such an analytical technology are recognition of people known to us in a crowd or effective management of multiple muscles when serving. These tasks that can even solve the child's brain are still not as modern computers.

The uniqueness of the human brain is that he can learn to solve new tasks, such as drive, learn foreign languages, etc. Despite this, the brain is not adapted to the processing of large amounts of information - a person will not be able to calculate even a square root of a large number in the mind, without using paper or calculator. In practice, numerical tasks are very often found, much more complex, rather than the root extraction. To solve such tasks, additional tools are needed.

The neural network accepts input information and analyzes it in a manner similar to what is used by our brain. The network is capable of learning. Subsequent results are issued on the basis of previously obtained experience.

The main task of a specialist who uses neural networks to solve some problem is the need to choose the most effective neural network architecture, i.e. Correctly select the type of neural network, the algorithm for its training, the number of neurons and types of links between them. Unfortunately, this work does not have a strict algorithm, it requires a deep understanding of various types of neural network architectures, includes many studies and can take a long time.

Application of neural networks is advisable if:

Sufficient amounts of data on previous system behavior have been accumulated

There are no traditional methods or algorithms, satisfactorily solving the problem.

The data is partially distorted, not full or contradictory, as a result of which traditional methods give out unsatisfactory result.

Neural networks best manifest themselves where there is a large number of input data, between which there are implicit relationships and patterns. In this case, the neural networks will help automatically take into account various nonlinear dependencies hidden in the data. This is especially important in solutions to support decision-making and prediction systems.

1.3. Advantages of using neural networks

Neural networks are indispensable when analyzing data, for example, for preliminary analysis or selection, detecting coarse human errors. It is advisable to use neural network methods in incomplete information tasks, in tasks where the solution can be found intuitively, and at the same time traditional mathematical models do not give the desired result.

Methods of neural networks are an excellent addition to traditional methods of statistical analysis, most of which are associated with the construction of models based on certain assumptions and theoretical conclusions (for example, that the desired dependence is linear or that some variable has a normal distribution). The neural network approach is not related to such assumptions - it is equally suitable for linear and complex nonlinear dependencies, especially effective in exploration analysis, when the goal is to find out if there are actions between variables. At the same time, the data may be incomplete, contradictory and even knowingly distorted. If there is some connection between the input and output data, not even detected by traditional correlation methods, the neural network is capable of automatically tune into it with a given degree of accuracy. In addition, modern neural networks have additional opportunities: they allow us to evaluate the comparative importance of various types of input information, reduce its volume without losing significant data, recognize the symptoms of approximation of critical situations, etc.

1.4. The principle of operation of neural networks

The speed of modern computers is about 100 MFLOPS (10 ^ 8 FLOPS) (FLOPS - a unit that denotes the speed of the computer, with a floating semicolon) in the brain contains approximately 10 ^ 11 neurons. The time of passing one nervous impulse - 1 ms, it is believed that the productivity of one neuron is about 10 flops. The equivalent speed of the brain will be 10 ^ 11 * 10 \u003d 10 ^ 12 FLOPS. If we consider the challenges solved by the brain and calculate the required number of operations to solve them on ordinary computers, then we obtain an estimate of speed up to 10 ^ 12 FLOPS. The difference in performance between the ordinary computer and the brain is 4 orders! In many ways, this winnings is due to parallelism of information processing in the brain. Consequently, to increase computer performance, it is necessary to move from the principles of the background-neimane to parallel processing of information. However, parallel computers have not yet been distributed for several reasons, which are due to the technical difficulties of implementation.

Artificial neural network is a significant simplified model of a biological neural network, i.e. Element of the nervous system. Of biology, fundamental ideas and principles are borrowed:

• 
  Neuron is a switch receiving and transmitting pulses, or signals. If the neuron gets a sufficiently strong impulse, they say that neuron is activated, that is, transfers the pulses associated with it to neurons. Not activated neuron remains at rest and does not pass the impulse.
• 
  Neuron consists of several components: synapses connecting neurons with other neurons and receiving pulses from neighboring neurons, axon, transmitting the pulse to other neurons, and dendrite receiving signals from various sources, incl. from synapses.
• 
  When neuron receives a pulse exceeding a certain threshold, it transmits the pulse to subsequent neurons (activates the pulse).
• 
  The synapse consists of two parts: the premiputic, connected to the axon of the transmitting pulse of the cell, and the postsenphetic, connected to the dendrite of the receiving pulse of the cell. Both parts of the synapse connect the synaptic gap.

The signal from the neuron to other neurons is transmitted through axon, which is not directly related to the receiving pulse with neurons. The pulse varies several times in synapse: before departure - in the presynaptic part and upon receipt - in postsynaptic.

The transmission pulse is formed in neuron, depending on one or more pulses. In the case of several pulses, neurons accumulates them. It will give an impulse or not, depends on the nature of the impulses received, who they are transmitted, etc. Thus, the dependence between transmitted and obtained impulses is nonlinear. If the neuron transmits the pulse, then it is activated.

The mathematical model of the neuron is built as follows:

Fig. 1. Model of artificial neuron

• 
  The input of the neuron model X is a vector consisting of a large number (N) component. Each of the components of the input vector Xi is one of the pulses obtained by neuron.
• 
  The output of the neuron model is one number x *. This means that inside the model, the input vector must be converted and aggregated into the scalar. In the future, this impulse will be transferred to other neurons.
• 
  It is known that when receiving a pulse, the synaps of neuron changes it. Mathematically, this change process can be described as follows: For each of the Xi components, the weight is specified. The pulse passed through synaps takes the look of Wixi. Note that weights can be assigned to model initialization, and may be variables and measured during the calculations. Weight is the internal parameters of the network, which was discussed above. Speaking about networking, they mean the foundation of Sinapse scales.
• 
  The addition of the obtained pulses. The aggregation of the obtained pulses is the calculation of their sum of σwixi.

Fig. 2. An example of a neural network with one hidden level.

Typically, neurons are located on the network by levels. The illustration shows an example of a three-level neural network:

1. 
   At the first level - input neurons (marked in blue), which receive data from the outside and transmitting pulses to neurons at the next level through synapses.
2. 
   Neurons on the hidden (second, red) level are treated with the obtained pulses and transmit them to neurons at the output (third, green) level.
3. 
   Nero at the output level produce final analysis and output of data.

Of course, the network architecture may be more complex, for example, with a large number of hidden levels or with a changing number of neurons. Models of neural networks are classified by three main parameters:

• 
  Type of communication between neurons in the network
• 
  Type of gear ratio;
• 
  Used network training algorithm

Further, the most important step is to train the neural network. After the network is trained, we can assume that it is ready for use

Fig. 3. Process of learning neuraloet