The structure of the neuron represented in Figure 1.
 Figure 1 - Neuron |
Neuron implements the function provided below.
 |
(1) |
Neurons are united in a group or layer. One layer or more form the neural network. Input layer neurons receive signals, convert them and then transmit the hidden layer neurons. Further work next layer until the weekend. The output layer gives the signals to the user. Each conclusion of neurons of any layer is fed to the input of all neurons of the next layer. The number of neurons in the layer can be anything. Scheme of the neural network is presented in Figure 2.
 Figure 2 - The transfer function |
Training artificial neural networks
ANNs can be trained, that is, to improve their work under the influence of environmental environment, modifying its parameters. There are many definitions of "training", but for the most suitable ANN follows, the Mendel and Maklarenom [5]: Learning - a process in which free parameters of the neural network being adapted as a result of its continuous stimulation external environment. Type of training is determined by the way, which made changes.
In addition to the recent literature the term "training" is also used equitable concept of "Training Network" and "network settings".
In addressing control problems commonly used supervised learning ANN which implies the existence of "teacher" who is watching the reaction of the network and sends a change in its parameters.
There are two types of supervised learning: a direct-controlled teaching and learning driven. Since the first appearance came before the second and more common, they usually refer to it simply as a controlled teaching.
ANN originally did not have any knowledge. In the process of learning "teacher" and network are exposed from the external environment, ie, their inputs received training signal, coinciding with one of the input patterns. "Teacher" tells the network what should be the correct (desired) response to standing the impact of issuing the corresponding output pattern. On the basis of magnitude error between the actual and desired outputs of the network by a particular rule a setting of its parameters. Repeat this process can be iteratively adjust the ANN so that it will emulate the "teacher", that is, his knowledge of the external environment will shift to it.
Typical tasks using direct instruction are approximation of unknown function, described by a set of data, and identification dynamic object. These problems are known input signals and the correct reaction to them, that is a training set of patterns.
The best known method of directly supervised learning layer pryamonapravlennyh ANN algorithm is back propagation of errors (backpropagation algorithm), which is a generalization of the method least squares.
Driven instruction does not use the knowledge of "teacher" on the desired output of ANN instead, training is conducted on the evaluation of the network transformation of input-output. Score is the external environment after filing for login network of training impact. At the same settings ANN conducted so as to maximize the scalar index of this evaluation, called stimulus (reinforcement signal).
The idea of this method is based on a real learning process, passing from living beings. In psychology it is known as a law action Thorndike. With respect to the stimulated train the ANN, this law can be rephrased as follows: if the action taken by the system of education, leads a satisfactory result, the tendency of the system to carry out this operation increases (the system is stimulated). Otherwise, the tendency to produce such action is reduced.
The most typical example of a system driven learning is adaptive control system. It is the learning part of the controller, and facility management, external influences and tones of his foreign assignments were environment. As a result, the impact of that environment, the controller generates a certain control signal, which translates into a new facility management state. In doing so, the quality of governance can be assessed only on the signal output object. As required by the reaction of the controller to achieve state of the object is not known in advance, you can not create a training set of templates, and, therefore, to apply a direct-controlled study. In this case can only be stimulated by learning controller for the quality of work whole management system as a whole, ie to assess the state of the environment.
There are two implementation driven learning: the direct incentives and encouragement to the prisoner. In the first case, the assessment work and corresponding settings INS conducted at each step of the learning systems. In the second case, when detained stimulating setting network is performed so as to maximize a cumulative assessment of the system in a certain sequence of steps.
Although direct stimulation of a typical classical schemes adaptive management, in recent years much attention is being paid to detainee stimulation. For the convergence of the method of direct incentive needed for each subsequent position of the external environment identifies only its previous position and previous exposure rendered ANN. When detained teaching such a restriction is removed. On the other hand, is much easier to formulate a criterion for determining the optimal behavior management system as a whole, the results of any actions, than criteria guide its movement at each step. In general, you will notice that system, learning only as a result of its interaction with the external environment, is more intellectual than using additional information "teacher".
There are many versions of a classical algorithm common mistake to use it as a method of direct or driven learning ANN. However, significant problems with its use and other methods of setting parameters ANN-based method of least squares and steepest descent, is their locality. In However, the target function (total error on the training set of templates or a cumulative assessment of the effectiveness of the learning system) is not unimodal. The number of local optimum for most practical problems of education millions in the dimension of the search space of the order of 100-1000. Consequently, the outcome of studies depends on the correct start point, and it is necessary to often-repeated procedure of setting ANN parameters.
These problems can be solved with the use of global optimization. The most effective of these is a genetic algorithm (GA). Considering ANNs as a single set of parameters, the GA is able to carry out its optimum setting when the dimension of the search space enough to most of the practical tasks. In this range of consideration applications far exceeds the ability of algorithm back propagation of error.
In the past ten years, developed many ways to controlled
training ANN using GA. The results obtained prove the great potential
this symbiosis. Sharing the ANN and the GA algorithm has
ideological advantage because they relate to the methods of evolutionary
model and develop within a paradigm of drawing techniques
natural methods and tools as the most optimal.
Conclusion
As a result of the work, after the task was to analyze existing software packages that work in this direction, methods for forecasting in the context of the enterprise, their weaknesses. After a detailed analysis has been chosen method prognozmrovaniya based on neural networks, which gives the most accurate and optimal forecasts. Development subsystem forecasting the demand for equipment for the cable TV networks should take into account the set of the following factors:
- Seasonality;
- The needs of consumers;
- Loading equipment;
- Material resources;
- Human resources;
- Financial resources;
- Instability in the country's economy.
After a detailed analysis is the approximate structure of the neural network to solve the task. Thus providing the basic input and output variables were the neural network in Figure 3.
 Figure 3 - Structure of neural networks (Animation: The volume of -17.5 in size, size - 854x317, number of frames - 4, the delay between frames - 1000 ms, the delay between the last and first frames - 1000 ms, number of cycles of recurrence - a continuous cycle of recurrence) |
References
- Ramon y Cajal S. Histologie du systeme nerveux de l'homme et des vertebres. — Paris: Maloine, 1911. — 714p.
- Shepherd G. M., Koch C. Introduction to synaptic circuits // The Synaptic Organization of the Brain (G. M. Shepherd, ed.). — New York: Oxford University Press, 1990. — 3—31p.
- Churchland P. S. Neurophilosophy: Toward a Unified Science of the Mind/Brain. — Cambridge, MA: MIT Press, 1986. — 127p.
- Aleksander I., Morton H. An Introduction to Neural Computing. — London: Chapman & Hall, 1990. — 218p.
- Бокун И.А., Темичев А.М. "Прогнозирование и планирование экономики". - М.: Наука, 2002.
- Бухалков М.И. Внутрифирменное планирование: Учебник. - 2-е издание., испр. и доп. - М.: ИНФРА-М, 2000.
- Фогель Л., Оуэнс А., Уолш М. Искусственный интеллект и эволюционное моделирование. - М.: Мир, 1969.
- Д. Рутковская, М. Пилиньский, Л. Рутковский "Нейронные сети, генетические алгоритмы и нечеткие системы " [Internet resource]: книга / Д. Рутковская - Киев: Горячая линия -Телеком, 2008. - Access mode: http://goods.marketgid.com/
- Саймон Хайкин "Нейронные сети. Полный курс" [Internet resource]: книга / Саймон Хайкин - Киев: Вильямс, 2006 г. . - Access mode: http://www.ozon.ru/
- Кричевский М.Л. "Интеллектуальные методы в менеджменте: Нейронные сети; Нечеткая логика; Генетические алгоритмы; Динамические системы" [Internet resource]: книга / СПб: Питер, 2005 г. . - Access mode: http://www.char.ru/
At writing of this abstract of thesis master's degree work is not yet completed. Final completion: December 1, 2009. Complete text of work and materials on the topic can be got for an author or his leader after the indicated date.
DonNTU > Master's portal Autobiography