JPH11102351A - Data sequential value predicting method, data sequential input determining method, and computer system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To predict the value of a future sequence by a communication network by inputting information regarding the sequence and times to a neural network and obtaining an output having the future predicted sequence from the neural network output. SOLUTION: An input unit 32 uses input data consisting of part values X(t-1) and X(t-2). Here, X represents a sequential value and (t) represents, for example, the current time. Those sequential values are sampled substantially at determined intervals. Input units 33 and 34 are used to input information regarding the current day of the week and the current time to the neural network. The neural network is trained by using a group of truing data. A connection 37 between units is a weighted connection. After the neural network is trained, the input data are provided for the input units 32 to 34 to generate an output X(t+1) as a predicted sequential value at time (t+1) at an output unit 36.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a data time series value prediction method for predicting a future time series value, and more particularly to a method for predicting a time series future value related to a traffic level in a communication network, and a neural network. The present invention relates to a data time series value prediction method for predicting at least one future data time series value, and a computer system realizing the method.

[0002]

2. Description of the Related Art As one of trend analysis and prediction methods,
Neural network technology. Neural networks have been used, for example, in many cases where it is necessary to predict financial market trends or to predict future evolution of a time series. A time series refers to continuous values measured at fixed time intervals of the equalizer over time. For example, the temperature in a specific building in a certain time, the number of births in a certain town in a certain time, the number of sunspots in a certain time, or the amount of water consumed in a certain area. In practice, time is measured as discrete time steps, such as the air temperature after each time interval.

[0003]

There are various problems when using neural network technology to predict the future development of a time series. The first problem is how to provide temporary information to the neural network. Because neural networks were originally designed to perform pattern recognition in static patterns, temporal dimensions must be provided in an appropriate manner. Another problem is that the training of neural networks and the need for a large database of information required to carefully evaluate the trained neural networks. Meeting these demands requires cost and time. Further, there are problems with limitations on the learning algorithms used to train neural networks. If the learning algorithm is not good, the training time will be long, and after the training, the performance of the neural network will not be very good. For example, neural networks overfit data, limiting the ability to generalize and address previously unseen data. Also, the neural network may simply learn noise detection instead of learning more significant and valuable information.

[0004] Neural networks that predict time series expansion are available in asynchronous transfer mode (ATM) communication networks. ATM technology offers great flexibility for transmission bandwidth allocation. With this technique, the amount of bandwidth allocated for a particular usage can be changed. To take full advantage of this capability, it is necessary to anticipate future bandwidth requirements in order to adjust the amount of bandwidth to meet future bandwidth requirements.
This prediction process must be able to guarantee sufficient bandwidth to provide quality for a particular task. At the same time, there is a need to minimize the overestimation of bandwidth requirements. In this way, the remaining maximum bandwidth can be used for other services. For example, one of the problems is predicting voice traffic in ATM communication networks. In this case, as much bandwidth as possible at any time should be left for video transmission or the like. The above is shown in FIG.

There are several specific problems when estimating traffic levels in ATM networks. A relatively short-term forecast must be possible, for example, providing an estimate of the traffic level 15 minutes in advance. There are also a number of telecommunications traffic characteristics that create particular problems in this area. For example, one of the characteristics of telecommunications traffic is that many cyclic effects of different periods are superimposed. For example, there are a trend related to a time corresponding to one business day, a trend on a day of the week (typically, traffic on a weekday is high and traffic on a weekend is very low), a trend in a month, and a trend in a season. In other words, the forecasting process needs to handle not only the cyclical effect but also the potential trends contained in the data. One known solution to this problem is to disarm by examining the periodicity of the cyclic effects and the average effect of each. There is. Eliminate such trends and make predictions from the resulting data. However, this takes time and complicates the process, leading to inaccurate predictions. Telecommunications is an area of rapid growth, with traffic behavior constantly evolving and changing. The prediction process must be able to handle such developments as well as the interaction between the various effects.

The next problem concerns early recognition of problems in telecommunications networks, especially ATM networks. ATM networks change continuously and generate many alarm or precursor information streams.
In this case, these streams need to be aware of when the sequence of events indicates an early significant cause of failure.

A further problem relates to customer network management. A "virtual private network" is provided to customers who make heavy use of the service provider network. This allows a portion of the service provider network to be
It can be controlled below the “level reference”. service·
The level criterion typically specifies the level of bandwidth available to the customer. If the customer utilizes bandwidth beyond this bandwidth, the data is effectively "discarded." However, it is difficult for customers to know such bandwidth requirements in advance and seek more bandwidth when needed.

Accordingly, it is an object of the present invention to predict future time series values and, in particular, to solve the above-mentioned problems,
It is an object of the present invention to provide a method and apparatus for predicting future time-series values in a communication network that at least reduces one or more of the above problems.

[0009]

According to a first aspect of the present invention, a method for predicting at least one future data time series value using a neural network comprises: (i) Inputting a plurality of time series values to a neural network, (ii) inputting information about time to the neural network, and (iii) obtaining an output having future predicted time series values from the neural network output. Be composed.

According to a second aspect of the present invention, the present invention provides
A computer system for predicting at least one future data time series value, comprising: (i) a neural network; and (ii) a first input configured to receive the plurality of time series values into the neural network;
(Iii) a second input configured to receive temporary information about the data time series values input to the neural network, and (iv) an output including future time series values from the neural network. And an output configured to provide.

This system has the advantage that a predicted future time series value can be obtained. These estimates are
It can be used for decisions, resource allocation, or other purposes. Because information about time is input into the neural network, the predicted values match the actual values well, which is particularly useful for applications that include many cyclic effects with different periods in the time series. Including time information eliminates the need for input data to be "disarmed" prior to use. This means that it is not necessary to remove information about circulation and other superficial trends in the data before entering the data into the neural network. Also, there is no need to combine this information with the output of the neural network.

Preferably, the information on the time includes information on a current time. By doing this,
In this way, the predicted values can be more closely matched to the actual values, which is particularly useful for time series applications involving many cyclical effects with different periodicities.

[0013] Further, preferably, the information on time is input to the neural network in the form of at least one pair of values on angle. This has the advantage that the number of inputs to the neural network required for time information can be relatively reduced. In addition, training of the neural network can be performed faster, and better generalization performance can be obtained. In addition, the cyclic characteristics of the time information can be complemented and can be well indicated by a pair of values relating to the angle. More accurate predictions can be made based on the information on the circulation characteristics of the time information.

Preferably, the pair of values includes a sine and a cosine of the angle. This makes it possible to indicate the time information at a certain position on the circumference of the circle. According to this display mode, the circulation characteristics of the time information can be easily accessed.

Preferably, the method further comprises the step of inputting at least some of the output from the neural network to the neural network.
The advantage of this is that the prediction can be made iteratively. For example, when predicting a time-series value in a 15-minute period by using this method, to perform prediction in a future 30-minute period,
The first prediction is used as input to the neutral network.

Preferably, the invention further comprises the step of inputting one or more auxiliary variable values into the neural network. This allows more information to be used, thus improving prediction. For example, if the time series is
In the case of the room temperature over time, the auxiliary variable value may be the outdoor temperature. By providing this auxiliary information to the neural network, better prediction results are obtained, especially if the auxiliary variable value and the predicted variable value have a strong correlation.

The invention is particularly useful when the time-series values of the data include information on bandwidth levels in an asynchronous transfer mode telecommunications network. This allows the amount of bandwidth to be adjusted and future bandwidth requirements to be predicted, thereby meeting future requirements. In this way, the remaining bandwidth available for other purposes can be maximized, while at the same time ensuring sufficient bandwidth for a particular task to maintain quality.

According to a third aspect of the present invention, in a computer system for predicting at least one future data time series value for a communication network, said communication network comprises a communication network management system, said computer system comprising: Input to a neural network configured to automatically receive time series values from a (i) neural network, (ii) a communication network management system, and (iii) future time series values. And an output from a neural network configured to supply the communication network management system. This allows the computer system to be embedded or integrated into the communication network management system. Using this method, the computer system automatically receives input and its output is processed by the communication network management system. For example, if a computer system predicts a bandwidth level for an asynchronous transfer mode telecommunications network, the output from the computer system can be used to automatically make bandwidth allocations, wherein a human operator can There is no need to intervene.

According to a fourth aspect of the present invention, in a method for predicting at least one future data time series value for a communication network, the communication network comprises a communication network management system, the method comprising: (i) Automatically inputting one or more time series values from the communication network management system to the neural network; (ii) obtaining an output including future time series values from the neural network; and automatically outputting the output to the communication network management system. Is provided. In this case, the method can be automatically executed together with the communication network management system. For example, if the bandwidth level was predicted for an asynchronous transfer mode telecommunications network in this manner, the output of the method could be used to automatically make bandwidth allocation adjustments, with the intervention of a human operator. No need.

According to a fifth aspect of the present invention, there is provided a data time series input value determining method for determining how many data time series input values are required to predict a future time series value, comprising: Forming a first set of vectors having the same magnitude and comprising a plurality of continuous time-series values; and (ii) forming a second vector comprising a plurality of consecutive time-series values having the same magnitude.
Where the first and second vectors differ in magnitude, and (iii) for each first vector, each first vector and its first neighbor vector Selecting other first vectors as first neighboring vectors such that a first measure of similarity between them is less than a threshold, and (iv) for each second vector,
And other second vectors are selected as second neighbor vectors such that a second measure of similarity between the second vector and its second neighbor vector is less than a threshold, where each second neighbor vector is The vectors correspond to each first neighborhood vector,
(V) determining the number of false neighbor vectors by comparing each first neighbor vector with its corresponding second neighbor vector; and (vi) determining a threshold of the number of false neighbor vectors of the first vector. It is configured to include a step of determining a required number of inputs according to the size.

According to this, the number of necessary input values is determined promptly, and there is no need for processing by trial and error. Further, the minimum number of input values for appropriately representing a time series can be determined. As a result, it is possible to reduce the computational complexity of the prediction process using the input value, and prevent prediction based on noise in the input data.

Preferably, the first similarity measure and the second similarity measure include a distance between two vectors. This simplifies the measure of similarity and speeds up the calculation. Also, the value of this distance well represents the similarity between the two vectors.

Preferably, the time-series value of the data includes information on a bandwidth level in an asynchronous transfer mode telecommunications network. This type of time series is particularly complicated because many cyclic effects with different periods are superimposed. In this case, it is particularly difficult to determine the minimum value of the input necessary for predicting the future value of the time series.

According to a sixth aspect of the invention, a computer system for determining how many data time series input values are required to predict future time series values comprises: A first processor forming a first set of vectors consisting of a plurality of consecutive time-series values; and (i)
i) a second processor having the same magnitude and forming a second set of vectors consisting of a plurality of consecutive time-series values, wherein the first vector and the second vector have different magnitudes; , (Iii) for each first vector, the other as a first neighbor vector, such that the first measure of similarity between each first vector and its first neighbor vector is less than a threshold. A selector for selecting the first vector; and (iv) for each second vector, a second for similarity between each first vector and its second neighbor vector.
A second selector that selects another second vector as the second neighbor vector such that the measure of
Where each second neighbor vector corresponds to each first neighbor vector, and (v) determining the number of false neighbor vectors by comparing each first neighbor vector with its corresponding second neighbor vector. And (vi) a fourth processor that determines the required number of inputs according to the magnitude of the first vector from which the threshold value of the number of false neighbor vectors is obtained.

According to this, the number of necessary input values is determined promptly, and there is no need for processing by trial and error. Further, the minimum number of input values for appropriately representing a time series can be determined. As a result, it is possible to reduce the computational complexity of the prediction process using the input value, and prevent prediction based on noise in the input data.

[0026]

DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below. Although these examples are the best known methods of putting the present invention into practice, the present invention is not limited to these examples.

As shown in FIG. 1, the trend analyzer 1 has a neural network 2. Input data 3
Is input to the trend analyzer 1 to generate a prediction 4. The prediction 4 is expressed in the form of a predicted future time-series value.

The input data 3 consists of time series values. These values are past and / or current values in the time series and also include time series predictions as described below. For example, the time series may be the room temperature with respect to the passage of time, and the input may be the current room temperature, the room temperature before 15 minutes, and the room temperature after 30 minutes. The time series value is usually a single variable, but a multivariable value may be used. As an example using a multivariable value, a time series includes a pair of temperature values and water consumption over time.

The input data 3 may include time information. For example, current time or future time. The term time is used to include date and time information. That is, the information on the time may include, for example, the day of the week. By including information about the time in the input data 3, the predicted value 4 generated by the trend analyzer is improved. This is especially the case when the time series includes a plurality of cyclic effects having various periodicities.

The time information included in the input data 3 is supplied in various formats. Stated another way, a time indication is provided to the input data 3. The term "display" means a description of a method and an item, and a set of rules that interpret that description. For example,
The time information may be indicated using a 24-hour clock system, or may be indicated as the number of seconds elapsed from a certain point. Thus, different display methods are suitable for different purposes. For example, when calculating the duration in seconds, it is more appropriate to indicate the number of seconds than to display the time with a 24-hour clock system. The time information included in the input data 3 is preferably displayed using sine / cosine coding. For this criterion,
The details are described below. By displaying such a time, there are several advantages. For example, the number of inputs to the neural network required for time information can be kept at a low level. Further, by displaying the time, the training of the neural network can be performed more quickly, and the generalization performance is improved. Further, by displaying the time, the circulation characteristic of the time information becomes clear, and more accurate prediction can be performed using the method of the present invention.

The time information may have one or more auxiliary variable values, but this is not essential. For example, when the time series relates to the indoor temperature, the auxiliary variable value may indicate the outdoor temperature.
By using the auxiliary variable value, the performance of the trend analyzer 1 is improved,
In particular, when the relationship between the auxiliary variable value and the time-series variable value is strong, the performance is improved.

The trend analyzer 1 predicts a time series value in the future. For example, a prediction of the room temperature after 15 minutes may be used as the output value. In addition, two or more output values are output, for example, 1
Temperatures after 5 minutes, 30 minutes and 1 hour may be predicted.

As shown in FIG. 2, the trend analyzer 1 includes a trend analysis engine 23 embedded in communication network management software 22. In this case, the input data 3 is supplied from the communication network 21, and the prediction 24 is generated from the trend analysis engine 23. By embedding the trend analyzer engine as described above, the engine 23
Automatically receives input from the communication network management system. The prediction 24 is output from the management system 22, and these predictions are used. For example, if the communication network 21 is an ATM telecommunications network, the trend analysis engine 23 predicts the bandwidth level for a particular service provided by the ATM network. The management system 22 automatically supplies information about the past and current bandwidth levels to the engine 23. Thereafter, the management system 22 utilizes the predicted bandwidth request 24 to adjust the bandwidth allocation to meet future requirements. The above processing does not require a human operator. By including the time information in the input data 3, the trend analysis engine 23 to be embedded in the host system can be made more appropriate.

The trend analysis engine 23 does not necessarily have to be embedded in the network management system. The trend analysis engine 23 may be an independent application like the trend analyzer 1 in FIG.

The term "communications network" refers to various types of systems in which information is passed between entities. For example, it may be a plurality of computers connected by cables, a mobile telephone network, or a telegraph system. "Telecommunications network" refers to a communication system suitable for telephone.

The trend analysis engine 23 is originally installed as a library of software elements. These software elements are only used to specifically embody the trend analysis engine 23 that is suitable for a particular use. The trend analysis engine 23 plays a general purpose role while taking the form of a library of software elements. That is,
The library of software components handles different input data and is suitable for many different analysis trends tasks with different output requirements and different numbers of auxiliary variables. A library of software elements is used to create a particular trend analyzer, where the configuration of the neural network 2 is configured to suit the task. With a general purpose engine, either an embedded or stand alone trend analyzer can be formed. General-purpose trend analyzer 23
Means "Cross Product" and "Cross Data
Layer. Cross-product refers to the trend analyzer being applicable to one or more telecommunications network types. Cross data layer refers to the ability of the trend analyzer to apply data collected from various layers of the telecommunications network. This is especially true for AT
Useful in M (Asynchronous Transfer Mode) networks and SDH (Synchronous Digital Hierarchy) networks.

As shown in FIG. 1, the trend analyzer uses a neural network 2. The neural network is preferably a feed-forward, multi-layer perceptual type network. In a feedforward neural network, each unit (including the input unit) feeds only units in the next layer. That is, there is no connection from one unit to the unit in the previous layer.

FIG. 3 is a schematic diagram illustrating this type of neural network. A layer of input unit 32 and hidden unit 35 is provided. Each input unit 32, 33, 34 is a hidden unit 3 that writes via a connection.
5 is connected. Each hidden unit 35 is connected to an output unit 36.

In the example shown in FIG. 3, the input unit 32
Is used for input data 3 consisting of past time-series values. X indicates a time-series value, and t indicates a specific time, for example,
Indicates the current time. In this example, three time-series values (t,
t-1, t-2) are provided as inputs 32 to the neural network, and these values correspond to the current time, the current time-1, and the current time-2. These time series values are substantially sampled at fixed intervals. Information about time is also input to the neural network. The input unit 33 is used to input information about the current day of the week, and the input unit 34 is used to input information about the current time.

The neural network shown in FIG.
First, training is performed using a set of training data. The inter-unit connections 37 are weighted connections, and the inputs to the neural network are modified by these weighted connections as they pass through the network and output to the output units. Connection 3 in the network during the training process
The weighting of 7 is modified to produce an output that approximates the expected output. These training processes are described below.

In the example shown in FIG. 3, after training of the neural network, the input data is
The output is provided to an output unit 36. In this example, the output is a predicted time series value for time (t + 1). However, this prediction may be at any time in the future. Also, a neural network 31 having one or more output units can be used. For example, using two outputs,
Predicted time series values for time (t + 1) and (t + 2) can be obtained. The input unit 32 shown in FIG.
33, 34, hidden unit 35, or output unit 3
It is not necessary to use the exact number of six. Also, the number of hidden layers may be different. Also, each of the input units 32, 33, 34 need not necessarily be connected to each of the hidden units.

FIG. 3 also shows that the neural network can be used for iterative prediction. In this case, information from the output unit 36 may be fed back to the neural network 31 as an input. This is indicated by arrow 38 in FIG. In this case, the time series values input to the neural network are fixed time intervals,
For example, it is sampled every 15 minutes. When making predictions iteratively, the output values used as inputs may be suitable for 15 minute intervals. For example, for this value to be a suitable input, then the output value of X must be used for time (t + 15).

The time information input to the neural networks 33 and 34 is displayed by using the sine / cosine coding device as described above. In this device, a specific type of time information is displayed using a pair of values that are sine and cosine values of a certain angle. For example, the day of the week, the time, or the month. Next, the input units 33 and 34
Are used for time information.

FIG. 4 shows how the sine / cosine encoder operates. In this example, the day of the week is displayed. The day of the week 50 is shown as a point on the circumference of the circle 45 as shown. For a particular day of the week, the radius from the center of the circle to the point indicating that day of the week can be determined. This is shown as Monday in FIG. Base line 46
Is calculated, and the angle 48 between the base line 46 and the radius 49 is calculated. The sine of this angle is the distance b in FIG.
And the cosine of this angle specifies the distance a. Considering these distances a and b as coordinates, the position of a point on a circle indicating a specific date is specified. In this way, the day of the week is specified using the pair of the sine value and the cosine value.
Similarly, other types of time can be indicated in this way by changing points on the circumference of the reference circle.

Another device for indicating time information is shown in FIG. Here, the seven units 42 are used for each day of the week 41. This is a one-bit encoding device,
Each of the two units 42 can be turned on or off. For example, to indicate Tuesday, the unit 43 is turned on as shown. It is disadvantageous in comparison with a sine / cosine coding apparatus in that the number of required units is not two but seven. This reduces the generalization ability of the neural network because it does not display the relationship between similar time information. For example, Monday is closer to Tuesday than Friday. It is also disadvantageous that the time required for network training is extended.

The process by which the neural network part 2 of the trend analyzer 1 is trained and evaluated is described in detail below. As the performance of the neural network deteriorates over time, it is necessary to maintain the neural network. This occurs when the characteristics of the input data change with time. This is a common occurrence,
In particular, it often occurs in telecommunications and the like in which usage patterns are continuously developed.

The initial training starts from an arbitrary starting position, that is, the weights in the neural network are arbitrarily set or all are set to the same value. Conversely, maintenance of the neural network occurs from the start position of the trained engine. In this case, the weights in the neural network are updated to reflect the new data.

In a preferred example, the trend analyzer 1 has C +
Writing is performed using an object-oriented programming language such as +. Successful training or maintenance returns a C ++ object called a neural network specification that contains information about the set of weights in the neural network.

The neural network 2 is trained by sending a set of training data to the neural network and, based on the algorithm learning rules, connects 37
Change the weight associated with. In a preferred example, a measured conjugate gradient learning algorithm is used, but a backpropagation rule may be used.

Ideally, the training data set should be made to show past data samples (ie, past time series). In the example of predicting voice traffic levels in a telecommunications network, the training data set ideally includes a spread of all traffic data that the user wants the predictor to make predictions. Abnormal data is included as well as typical data. However, at the same time, it is desirable to reduce the noise level in the training data set as much as possible. By doing so, the neural network does not learn to detect noise where the term noise means random fluctuations occurring in the data.

Data is collected to form a training data set. For example, FIG. 5 shows an example of a possible training data format. The first column 51 is a quantity Q indicating the number of audio circuits in the telecommunications network. The second column shows the time at which each quantity value was obtained, and the third column 53 contains auxiliary variable values. Data is collected in chronological order at fixed time intervals between samples. The size of this time interval is selected according to the requirements of the prediction task and the particular application to be used. The training performance is monitored such that the desired output is known for each training data input. During training, the trend analyzer 1 automatically validates its performance. This activation is performed during training by randomly selecting a portion of the training data to be checked. This has two effects. First, it prevents overtraining (in which case the engine learns too well for a particular data set and loses its ability to generalize), reducing the time it takes to train.

After training of the engine 23, predictions are made by sending further input data. During the prediction phase, engine 23 monitors performance and determines if retraining is needed. This is done by comparing recently sent input data with data from the training set. If the difference is significant according to a predetermined criterion or threshold, retraining is performed.

In the retraining, a copy of the trend analysis engine 23 is created using a neutral network, and the copy is retrained. After retraining is performed, the performance of the replica (or second generation engine) is enabled. If the activation is successful, the original engine will be replaced with a Nisei engine. By doing this, during retraining,
The original engine can be used. The second generation engine can be moved to a stationary node for retraining. Also, failure to retrain will not cause any damage to the original engine. The updated training data set is
Retraining uses the same algorithm as the first training, although one containing a more recent example is used.

The output from the neural network 4 includes, in addition to the predictions, a measure of the accuracy of each prediction. For example, FIG. 6 shows the information included in the output. This information includes the predicted amount 61 and the associated time at which the amount is predicted,
And associated accuracy. Accuracy is quantity 6
It exists in the same unit as 1 and indicates the range of the value of the predicted amount. In this example, the quantity 15320 is 15320+
Accuracy in the range of-/ 32.

This accuracy can be determined using conventional methods. For example, an average value of errors that occurred in a recent time interval can be used. Alternatively, the prediction can be treated as the maximum of the estimated probability density coefficient, and the error can be determined using the required level of confidence.

Since the trend analyzer 1 is based on the neural network technology, the following advantages are obtained. Accuracy-Prediction using a neural network engine is claimed to outperform multivariate discriminant analysis, autoregressive integral moving average and autoregressive moving average. Robustness-Neural networks are more resilient to noisy training data than standard statistical techniques. • Retentivity-With neural network technology, it is only necessary to retrain the engine periodically to keep performance at an acceptable level. Development time-Provide a library of software elements to minimize development time. Speed-With neural networks, the prediction mode takes only 0.01 seconds. Portability-The engine is the available cross-product and cross-data layer and can be used on a variety of platforms from PCs to workstations.

Like the neural network element 2, the trend analyzer 1 also has a management element. The management element has the ability to first create the engine and then maintain it. To retain the engine, train the neural network element 2 and retrain if necessary. As described above, the trend analysis engine 23 is initially set as a library of software elements. Once
Once these elements are put together, Trend Analysis Engine 2
Is a simple C ++ application programming
Interface (API) or standalone
Through application, it can be integrated with other system software.

The API includes a trend analysis interface object (called a TAI interface object) and provides a trend analysis interface object C ++ method (a method of communicating with the engine). This method usually has an associated "return event" method associated with it. The user inherits the return event method from this object, overloads it,
Add the behavior required for a specific application. See the attached appendix for this.

The steps required to implement the trend analysis engine 23 for a particular application are described below. In this example, the trend analyzer 23 is used to predict the reserved bandwidth for public network voice traffic. The objectives are: to achieve reliable voice transmission with predictable quality in an ATM environment; to release unused bandwidth for other services.

Prediction is required for interlocation traffic, ie, traffic transmitted from local exchange A to local exchange B. By predicting the traffic between each pair of locations, the traffic on each link can be calculated (by summation integration software), and the correct amount of bandwidth for each link is
Shortly before you need it, you can assign it. The bandwidth allocation is updated every 10 minutes.

The prediction is required to be made every 10 minutes. That is, data must be collected every 10 minutes (or less). The prediction accuracy is 0
Must be in the range between 10%.

FIG. 7 shows the actual required bandwidth 71 for a typical day for a telephone service, and the expected bandwidth 72 defined by the forecast of the trend analysis engine. The defined bandwidth 72 is calculated by adding a 5% safety margin to the prediction and 1
It is designed to be able to cope with slight fluctuations occurring in traffic at 0 minute intervals.

To determine the data prediction rate and any amount of noise, some samples need to be captured for analysis. Data must be collected at the same intervals as the predictions are made.

The details of the method of determining the number of past time-series values to be input to the engine will be described later in this specification.

By simplifying the visual inspection of analytical data capture and gaining experience with how the traffic profile changes, the audio profile can be determined by time of day, day of the week, date, and month. I understand.

In the next step, a trend analysis engine 23 including the neural network element 2 is created. To create the engine 23, of the six API methods,
Use one. In such a construction method, neural
A trend analysis specification that specifies the number of inputs to the network is required. In this example, number of ancillar
y variables is determined to be 0 and recall window siz
e is determined to be 4, and data log widnow size is set to 5. Once the user has determined the specification details, the trend analysis engine 23 is implemented by calling the components on the trend analysis specification object (see attached appendix).

The communication network management system 22 forms, updates, and holds training data sets. Then AP
Calling one of the six methods of I, the engine 23
Do training.

Once the training of the trend analyzer is completed, the preparation for prediction is completed. The first task is to provide the prediction buffer with enough data to make the prediction. At some point, a data item is added from the data source. It is the responsibility of the communication network management system 22 to extract this data from the source.

The AddInputPresentation method (refer to the attached appendix) is executed for the number of past time series values that need to be predicted. In the normal operation mode of the engine 23, prediction is performed. A new data item is entered, a prediction is made, and then the prediction data is deleted. It is the responsibility of the communication network management system 22 or integration software to make a copy of the prediction and delete the prediction from the engine. The prediction is generated by calling MakePrediction.

In this example, the number of repetitions is set to one.
This is because the prediction of the engine simply needs to make a prediction one step before. This time step is 10 minutes ahead, ie only 10 minutes are needed for this application.

The return event has prediction data. Because one or more data items may be sent, the prediction data is sent in pairs. One accuracy measure is transmitted. This accuracy measure is the mean squared error over the window of recent predictions kept in a log of recent predictions and actual values.

[0072]Iterative prediction  As mentioned above, trend analysis is used for further forecasting.
The output of the analyzer 1 may be used as an input to the analyzer 1.
However, to make further predictions, the actual data
When using prediction as input, use as actual data
Cannot be used. Auxiliary variable value
Can be used only when one prediction is performed. I
However, do not use auxiliary variable values to make multiple predictions.
If you must, the following options are available: The auxiliary variable for all future predictions is the last measurement
Is assumed to be constant.・ Specify multiple trend analyzers 1 and forecast 1, 2, 3, etc.
Progresses.・ Do not use auxiliary variable values for prediction.

By using a single trend analyzer and further processing in the integration layer, it is possible to perform variable length prediction. For example, in a trend analyzer setup, 15
Make predictions every minute. However, the user has the option to make multiple predictions beyond this time step. Thus, integration software can be created that aggregates multiple predictions into one value. This looks like one prediction with several time steps to the future, although it is actually multiple predictions.

[0074]The number of previous time series needed to make a prediction
Calculation  Voice traffic between two local exchanges on a telecommunications network
An example of estimating future time series values for
This is briefly described below. In this example, a new
A neural network was used. This system is
This is a trend analyzer 1 described later in a detailed book. This trend analyzer
Linked to host telecommunications network management system
Thus, 1339 time series points were used. Necessary for anticipation
To determine the number of values before
went.

1. Obtain a continuous value of the voice traffic volume in the equally-spaced period. For example, this value is as follows: x (0), x (1), x (2), x (3), x (4), x (5), ... x (1339)

2. A vector of magnitude 2 is formed from these values. For example, the vector looks like this: S (0) = [x (0), x (1)] S (1) = [x (1), x (2)] S (2) = [x (2), x (3)]・ S (1339) = [x (1338), x (1339)]

3. Compute all similarities between these possible vector pairs. For example, similarity can be calculated as the Euclidean distance between two vectors. As a measure of similarity, a measure other than the Euclidean distance can be used. Two vectors (1,1) and (4,5)
Has a Euclidean distance of 5 as shown in FIG. Vector (1,1) is shown as point 81, and vector (4,5) is shown as point 82. In FIG.
The distance between these two points, 81 and 82, is 83, and the length is 5 units. In this way, the distance is obtained for all the vector pairs. For example, vector pairs include S (0) and S (1), S (0) and S (2), and S (1) and S (2).

4. For each vector, find a neighboring vector. That is, another vector having the shortest Euclidean distance is selected for each vector. This other vector is called a neighbor vector.

5. For example, here, step 2 of the method is performed again for a vector of length 3.
In this example, the vectors are as follows: S (0) = [x (0), x (1), x (2)] S (1) = [x (1), x (2), x (3)] S (1) = [x ( 2), x (3), x (4)]

6. For a length 3 vector, the similarity is calculated for all possible vector pairs using the same method as the similarity calculation in step 3. Next, a second neighborhood vector is determined in the same manner as in step 4.

7. A predetermined vector (eg, S
(0)), one exists in step 4 and the other has a pair of corresponding neighborhood vectors obtained in step 6. Compare the two neighbor vectors of each pair. If the neighborhood vectors determined in step 6 are “worse” than the neighborhood vectors determined in step 4, these are considered as false neighborhood vectors. The quality of a neighbor vector is measured on the basis of how close it is to the associated vector. In this example, if the neighborhood vector criterion determined in step 6 is inferior to the neighborhood vector criterion determined in step 4, the neighborhood vector determined in step 6 is a false neighborhood vector. Typically, a predetermined threshold is used to determine whether the similarity criterion is bad.

8. Find the total count of false neighbor vectors. This process is repeated for large vectors. False neighbor vector 9 for vector magnitude 92
The total number graph of 1 is drawn as shown in FIG. The magnitude of the vector corresponds to the window size and the number of previous time series values input to the prediction system. FIG. 9 shows that the number of the false neighbor vectors 91 is rapidly reduced at the window size of 4. After this sharp decrease, there is little change in the graph. For a window size of 21, the number of false neighbor vectors is reduced to 5, and for a window size of 42, the number reaches 4. Consider a graph showing the vector size for the false neighborhood vector. First, as the number of inputs for prediction processing,
A relatively small value of vector size 92 was chosen. For example, FIG. 9 shows a vector or window size 4
It is shown. Using this previous number of time series, the trend analyzer 1 worked well as a predictor on the training data set and showed good generalization behavior on unobserved data.

The method or algorithm for determining the number of previous time series required for prediction is described in more detail below.

The algorithm analyzes successive input data values and determines the correct window size.
A one-dimensional sample Z (T) is collected, and continuous values are combined to generate a multidimensional vector s of dimension d.

For example, when the dimension d = 2, a vector S = {s (0), s (1),...} Is generated from continuous values as follows. s (0) = [z (0), z (1)] s (n) = [z (n), z (n + 1)] s (N-1) = [z (N-1), z (N)]

The theoretical result is that if the value of d is large enough, the path of these vectors in R ^d will show system variation. Here, z indicates the observed variable value. The goal is to find the minimum with this property. The minimum value of d is determined using the nearest neighbor. In this idea, for each of s (n), the nearest neighbor vector to S is found, and the distance between the vectors is NearestNeigh
Recorded as bourDistance (n, d). This distance is s
(N) and recalculated for the nearest neighbor vector, but this time with an incremental window size of NearestNeigh
bourDistance (n, d + 1). If the difference between these two values is greater than the original distance, it is determined to be the nearest false neighbor vector. Formally, | NearestNeighbourDistance (n, d) -NearestNeighbourDis
When tance (n, d + 1) | / NearestNeighbourDistance (n, d)> R, s (n) is determined to be the nearest false neighbor vector. A suitable value for the threshold value R is between 10 and 50. Value 1
It is preferable to use 0.

To find a suitable window size, calculate the nearest false neighborhood vector for the entire training set:
Gradually increase the window size. As the number approaches zero, the window size is fixed. At this stage, the variability of the system is quite reliable.

Various other applications are possible within the scope of the present invention. The present invention includes a case where a future time-series value is predicted. For example, financial forecasting in the stock market, electrical load forecasting in the power network, traffic forecasting in the transportation network, and failure forecasting in process control. Other applications include AT
There are call management control and link capacity allocation in the M network. In addition, from the service provider,
It can be used, for example, when a customer requests a predicted future bandwidth to negotiate an additional allocation of bandwidth.

[0089]Appendix TAPrediction  TAPrediction includes a predicted value and an associated time. TAPrediction :: GetPredictionValue floatGetPredictionValue () const; (Note) Returns prediction. TAPrediction :: GetTimePredictionIsFor Time GetPredictionIsFor () const; (Note) Returns the time related to prediction.DTDataSetSpecification  DTDataSetSpecification is used for data conversion performed in TA.
A place holder for the required structural information. DTDataSetSpecification :: DTDataSetSpecification DTDataSetSpecification (int no_of_ts input values,
int no of ancillary values, Bool month, Bool day o
f week, Bool hour, Bool minute, IncrementIntervalT
ype increment interval, int increment step, int no
of intervals to output, float normalisation upper
bound, float normalisation lower bound); no ts input values: This is the number of past values of the forecast quantity
It is. A typical value for this is 4. This value is TA
Recall in spec window must be equal to size
Absent. no of ancillary values: This is the prediction that affects the prediction
The quantity is the number of inputs other than time and past values. This value
Is the number in the TA specification of ancillary values
Have to go. Month: Boo indicating whether the data changes in a monthly cycle
Value. day of week: This is whether the data changes weekly
Is a Boolean value indicating hour: This is whether the data changes over time
Boolean value to indicate. minute: This is whether the data changes over time
Is a Boolean value indicating increment interval: This is the interval at which the
Or tell the engine (eg, seconds). increment step: This is how much the interval is incremented
Or let the engine know (eg; 30)
Ta and increment Combine intervals to get the engine
Indicates to the engine whether to increment by as much as normalisation upper bound: in the training / retraining phase
Set this value to 0.0 because it is set automatically
You. normalisation lower bound: in the training / retraining phase
Set this value to 0.0 because it is set automatically
You. DTDataSetSpecification :: IncrementIntervalType This is a countable type that takes the following value. DTDataSet  DTDataSet is the correct form to pass to TA training
Supply containers for data. This data set is shown in FIG.
Must have at least one row as shown in 0
Absent. DTDataSet :: DTDataSet DataSet (); DTDataSet (List of_ p <DTRow>^*rows); rows: list of row pointers. (Note) Generate a data set. DTDataSet :: LinkR18Has LinkR18Has (DTRRow^*rows id) rows id: Indicates a row pointer. (Note) Add a row to the data set.DTRow  DTRow provides a container for related information. Sand
In other words, as shown in FIG.
Can be combined with auxiliary variable values. Many rows are in the data set
Can be joined together. DTRow :: DTRow DTRow () DTRow (int row number) row number: Line number in the data set. (Note) Generate a row. DTRow :: LinkR5IscomposedOf LinkR5IscomposedOf (DTDataItem^*data item id) data item id: pointer to data item (Note) Add data item to row. Data item
System is added to the rows in a particular order. Date and time data
• Items always come at the very beginning of the line. Make a prediction
One data item comes. Finally, the user needs
You can add as many auxiliary variable data items
You. See DTDataItem.DTDataItem  DTDataItem is a data place holder. Data is day
Date and time information, or one data value
No. Combine multiple data items in one line
Can be. See DTRow. DTDataItem :: DTDataItem DTDataItem (Time^*time values, int column_number) DTDataItem (float numeric_value, int column_number) time values: Date and time information. numeric value: One data value. column number: position in the data item list. (Note) Generate data items.NNNeuralnetworkCreationSpec  NNNeuralnetworkCreationSpec is a neural network
This is a place holder that holds the information contained in the work element.
You. FIG. 12 shows two other components that need to be configured first.
Neural network students associated with different objects
Shows the specifications. These two objects are
Yard Network Specifications and Network
Training specifications. NeuralnetworkCreationSpec NNNeuralnetworkCreationSpec (NNNeuralnetworkCreatioSpec
nSpec^*  network spec id, NNNetworkTrainerSpec^*trainer spec
id) network spec id: Point to network specification
It is. trainer spec id: a pointer to the training specification.
(Note) Generate NNNeuralnetworkCreationSpec.NNNeuralnetworkSpec  NNNeuralnetworkSpec is another type of neural network.
Super type object for network support
Project. NNLayeredNetworkSpec is a sub-tie
And replaces the object NNNeuralnetworkSpec
You may.NNLayeredNetworkSpec  Layered network specifications consist of two configurations
Have. Provide an array of weights (trained
Trained) or without weights (trained
Not specs), you can make a call
You. NNLayeredNetworkSpec :: NNLayeredNetworkSpec NNLayeredNetworkSpec (Lis <int> &unit_mumbers); NNLayeredNetworkSpec (Lis <int> & unit_mumbers, SWAArr
ay &weights); unit_mumbers: list of 3 integer values • Number of units in input layer. Forecast amount of past
Depending on the number of values, changing time intervals, and the number of auxiliary
Thus, this number is determined. • The number of units in the hidden layer. Topology optimization
The number of levers is determined. -Number of units in the output layer. This is set to 1
It is. weights: this is in a neural network
It is the value of each weight between the connections. This value is
Set during training. Do not pass trained specifications
If so, it is necessary to weight. Untrained TA
If the specifications pass, no weighting is required.NNNetworkTrainerSpec  Network training specifications are based on neural network
Location holder for information contained in the training element. NNNetworkTrainerSpec :: NNNetworkTrainerSpec NNNetworkTrainerSpec (float target error, unsigned
int percentage validation, Bool is early stopping r
equired, unsigned int number of training cycles, l
ong random seed, unsigned int max number of steps,
 floatfractional tolerance); target error-this is the TA measured on the training data
This is a training stop condition. • A zero value aborts this test. 0 is a normal value
is there. A non-zero value gives an error value that stops training
(Unless training has previously stopped for other reasons). percentage validation: is early stopping required =
Only significant if TRUE. Training data percentage
Are randomly selected as valid data and therefore optimal
Not used for conversion. is early stopping required: early stopping new
Ral network technology should be used for generalization.
A Boolean value that indicates whether the In most cases,
They are set to TRUE. number of training cycles: Find the best solution
Number of times the TA is reinitialized and trained for • A zero value requires retraining. That is, the previous weight
One training cycle starting from the threshold. • A non-zero value indicates the number of training cycles to perform. each
At the beginning of the training cycle, weights are randomized. Return
The optimal network is the optimal training cycle. random seed: This initializes the weights and sets
Kind of pseudo-random number generator used to select
Control. • For values of -1, the generator is the system clock
Seeds the value derived from. This will generate
Maximize unpredicted numbers. With this parameter,
1 is a normal value. • Positive numbers are converted to unsigned int (eg, 32
Use this value as a seed. This option
Requires the same sequence of pseudo-random numbers each time.
Used for regression testing and debugging
You. max number of steps: This limits the number of TA updates
There are other conditions to stop training in order to. • A zero value aborts this test. This parameter
Is usually zero. • A non-zero value is the number of steps to stop the training cycle
Supply (if not previously stopped for other reasons)
). fractional tolerance: No more significant steps
If not, the optimizer stops (for other reasons,
If not previously stopped).・ Zero value is smaller than the accuracy of floating point calculation.
If so, this step is deemed insignificant. This
Conformance levels performed on a basis of
Not worth the extra time. Non-zero values are required for steps to be considered important
Shows a relative improvement. This is actually more
To reduce the time required for optimization without making significant differences
It may be used as a fairly simple method. This value
10^-2-10^-6Are recommended as starting points for the experiment.
You.

[Brief description of the drawings]

FIG. 1 is a general schematic diagram of a configuration for predicting a time series value in the future.

FIG. 2 embedded in communication network management software,
FIG. 3 is a diagram illustrating a configuration for predicting a future time-series value regarding a communication network.

FIG. 3 illustrates a neural network used in the configuration of FIG.
It is a general outline figure of a network.

FIG. 4 is a diagram illustrating a sine / cosine coding configuration used in the configuration of FIG. 2;

FIG. 5 is a diagram showing input data for the configuration of FIG. 2;

FIG. 6 is a diagram showing information included in the output of the configuration.

FIG. 7 is a graph showing a bandwidth required for a telephone service on a time axis.

FIG. 8 is a diagram illustrating a method of calculating a Euclidean distance between two vectors.

FIG. 9 is a diagram showing the number of false neighbors with respect to a window size.

FIG. 10 is a diagram showing a DTDataSet and a related DTRow.

FIG. 11 is a diagram showing a DTRow and related DTDatItems.

FIG. 12 is a diagram showing a neural network generation specification and related objects.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 1 ... Trend analyzer 2 ... Neural network 3 ... Input data 4 ... Forecast 21 ... Communication network 22 ... Communication network management system 23 ... Trend analysis engine 24 ... Forecast 32 ... Input unit 33 ... Input unit 34 ... Input unit 35 ... Hidden Unit 36 ... Output unit 37 ... Connection between units 41 ... Week 42 ... Unit 43 ... Unit 45 ... Yen 46 ... Base line 48 ... Angle 49 ... Radius 50 ... Day of the week

 ──────────────────────────────────────────────────続 き Continued on the front page (71) Applicant 390023157 THE WORLD TRADE CEN TRE OF MONTREAL, MON TREAL, QUEBEC H2Y3Y 4, CANADA (72) Inventor Jonason Coward UK, CB 10 2 Tiest, Essex, Ladd Winter , East View Closed 12 (72) Inventor Peter Hammer, C.M.234 DDS, Hartfordshire, Bishop's Stortford, Burleigh Hills 29 (72) Inventor Kevin John Tsuichen E.L. , Hertfordshire, Harpenden, Nyan Cross 7 (72) Inventor Philip Iriamu Hobson UK, GCM 23 4 Quality Assurance, Hertfordshire, Bishop's strike Tofodo, Woodpecker closed 13 (72) inventor Ray Frank UK, Aruemu 14 1 BN, Essex, Upminster, Ingure burn Gardens 40

Claims (29)

[Claims] 1. A data time series value prediction method for predicting at least one future data time series value using a neural network, comprising: (i) inputting a plurality of time series values to the neural network; A method for predicting a data time-series value, comprising the steps of: (i) inputting information about time into a neural network; 2. The method according to claim 1, wherein the information on the time includes information on a current time. 3. A method according to claim 1, wherein said information on time is input to the neural network in the form of at least one pair of values on angle. 4. The method according to claim 3, wherein said pair of values includes a sine and a cosine of said angle. 5. The method of claim 1, further comprising:
Inputting at least some of the output from the neural network to the neural network. 6. The method of claim 1, further comprising:
A method for predicting a data time series value, comprising inputting one or more auxiliary variable values to a neural network. 7. The method according to claim 1, wherein the time series value of the data is a univariate. 8. The method according to claim 1, wherein the time series of the data includes information related to a communication network. 9. The method according to claim 1, wherein the time series value of the data includes information on a traffic level. 10. The method of claim 1, wherein the time series value of the data includes information about a bandwidth level in an asynchronous transfer mode telecommunications network. 11. A computer system for predicting at least one future data time series value, the computer system comprising: (i) a neural network; and (ii) receiving a plurality of time series values to the neural network. (Iii) a second input configured to receive temporary information about the data time series values input to the neural network; and (iv) future time series from the neural network. An output configured to provide an output including a value. 12. At least one communication network
A computer system for predicting two future data time series values, wherein said communication network comprises a communication network management system, said computer system comprising: (i) a neural network; and (ii) automatically from said communication network management system. Neural network configured to receive time series values automatically
A computer system comprising: an input to a network; and (iii) an output from a neural network configured to provide future time series values to a communication network management system. 13. At least one communication network
A method for predicting two future data time series values, wherein said communication network comprises a communication network management system, said method comprising:
Automatically inputting one or more time series values to the network; and (ii) obtaining an output including future time series values from the neural network and automatically providing the output to a communication network management system. A data time-series value prediction method, characterized in that: 14. The method according to claim 13, further comprising the step of inputting temporary information relating to the input value of the data time series value into a neural network. 15. The method of claim 1, wherein the step of inputting a plurality of time series values to a neural network further comprises determining the number of time series values required to input to the neural network. A data time-series value prediction method, comprising the step of: 16. The method of claim 1, wherein the step (i) of inputting a plurality of time series values to a neural network further comprises determining a number of time series values required to input to the neural network. Determining the number of the time-series values, wherein (i)
Generating a first set of vectors, each first vector having the same magnitude, each first vector having a plurality of successive values in time series, and (ii) generating a second set of vectors. And the second
Are of the same magnitude, each second vector has a plurality of time-series consecutive values, where the first and second vectors are of different magnitudes, and (iii) each vector has For the first vector, select another first vector as a first neighbor vector, wherein a first measure of similarity between each first vector and its first neighbor vector is greater than a threshold. Small, and (iv) for each second vector, select the other second vector as the second neighbor vector, where the similarity between each first vector and its second neighbor vector is greater than a threshold Small, where each second neighborhood vector corresponds to a first neighborhood vector, and (v) comparing the number of false neighborhood vectors by comparing each first neighborhood vector with its corresponding second neighborhood vector And (vi) the false neighborhood vector According to the magnitude of the first vector to determine the number of values, the data time series value prediction method characterized by comprising the step of determining the number of required values for input to the neural network. 17. A communications network, comprising: (i) a communications network management system; and (ii) a computer system for predicting at least one future data time series value in the communications network, said computer system comprising: (ii) input to a neural network configured to automatically receive time series values from the communication network management system or (ii) providing future time series values to the communication network management system. A communication network comprising an output from a neural network configured to: 18. A data time series input value determination method for determining how many data time series input values are required to predict a future time series value, comprising: (i) a plurality of data time series input values having the same size; Forming a first set of vectors consisting of continuous time-series values; (ii) forming a second set of vectors of the same magnitude and consisting of a plurality of continuous time-series values, where The first vector and the second vector have different magnitudes; and (iii) for each first vector, a first measure of similarity between each first vector and its first neighboring vector is less than a threshold. Selecting the other first vectors as the first neighboring vectors such that: (iv) for each second vector, the similarity between each first vector and its second neighboring vector The second measure is such that the second measure is below the threshold. Selecting another second vector as a neighborhood vector, wherein each second neighborhood vector corresponds to each first neighborhood vector; and (v) each first neighborhood vector and its corresponding second neighborhood. Determining the number of false neighbor vectors by comparing the vectors; and (vi) determining the required number of inputs according to the magnitude of the first vector from which the threshold for the number of false neighbor vectors was obtained. Data time series input value determination method to be used. 19. The method of claim 18, wherein the number of consecutive time series values relates to substantially equal time intervals. 20. The method of claim 18, wherein the first
(I) obtaining a plurality of consecutive time-series values for substantially equal time intervals; and (ii) converting the continuous values into a plurality of equal-sized groups. And (iii) generating a vector from each of the groups having the same size. 21. The method of claim 18, wherein the first similarity measure and the second similarity measure include a distance between two vectors. Method. 22. The method of claim 18, wherein the first and second measures of similarity include a Euclidean distance between two vectors. 23. The method of claim 18, wherein the step (v) of determining a false neighbor vector further comprises: (i) a first vector between the first vector and the first neighbor vector.
(Ii) determining a second distance between the second vector and its second neighboring vector, wherein the first and second neighboring vectors correspond; (iii) the first And (iv) determining whether the difference is greater than a threshold value. 24. The method according to claim 18, wherein the time series value of the data includes information on a communication network. 25. The method of claim 18, wherein the time series value of the data includes information related to a telecommunications network. 26. The method according to claim 18, wherein the time series value of the data includes information about a bandwidth level in an asynchronous transfer mode telecommunications network. 27. The method of claim 18, further comprising using a neural network to perform at least one
A data time-series input value determination method characterized by predicting two future time-series values. 28. The method of claim 18, further comprising using a neural network to perform at least one
Predicting two future time series values, comprising: (i) inputting said number of input values into a neural network; (ii) inputting information about time into the neural network; (iii) A method for determining a data time-series input value, wherein said output including a future time-series value is obtained from a neural network output. 29. A computer system for determining how many data time series input values are required to predict future time series values, comprising: (i) a plurality of consecutive time series of the same magnitude; A first processor forming a first set of vectors of values; (ii) a second processor of the same magnitude and forming a second set of vectors of a plurality of consecutive time-series values; Where the first vector and the second vector are different in magnitude, and (iii) for each first vector, a second vector relating to the similarity between each first vector and its first neighbor vector. A selector for selecting another first vector as the first neighboring vector such that the measure of 1 is less than the threshold; and (iv) for each second vector, each first vector and its second Between similar neighborhood vectors A second selector that selects another second vector as a second neighbor vector such that a second measure for the second neighbor vector is less than a threshold, wherein each second neighbor vector is a respective first neighbor vector (V) a third processor for determining the number of false neighbor vectors by comparing each first neighbor vector with its corresponding second neighbor vector; and (vi) a threshold for the number of false neighbor vectors. Determine the required number of inputs according to the magnitude of the first vector obtained
And a computer comprising:
system.