JPH075235B2 - Elevator group management control device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to an elevator group supervisory control device effective for assigning a hall call to an optimal elevator when a plurality of elevators are installed in parallel. It is a thing.

[Problems to be Solved by Prior Art and Invention]

In the current elevators, the mainstream of group management control is allocation control using an evaluation function.

This is to calculate, for example, which car is best assigned to each call when a hall call occurs, by numerically calculating for each car using a predetermined evaluation function for various evaluation indexes such as estimated waiting time. , Is assigned to the car with the largest value or the car with the smallest value, and advanced control can be performed by appropriately selecting the evaluation index and devising the evaluation function.

On the other hand, in recent years, call assignment control by an expert system using fuzzy theory has been proposed in order to perform more advanced control.

This is a method of recognizing various evaluation indexes as fuzzy quantities, using a rule group that describes an appropriate allocation method in IF-THEN format, and selecting and assigning an optimal car from the goodness of fit to the rule group. It becomes easy to incorporate the expert's knowledge into the control, and it is possible to realize detailed control according to the characteristics of the building.

By the way, as a matter of course, in any of the above-described methods, the assignment algorithm (evaluation formula or evaluation rule) is created by a human being. That is, these methods
Examine what kind of index should be used as an evaluation index, such as predicted waiting time, long wait occurrence probability, forecast outage rate, etc.
Furthermore, the evaluation formula and evaluation rules are determined in consideration of the priority (importance) of each index. For this reason, when trying to cope with complicated traffic patterns and fluctuating traffic demands, evaluation formulas and evaluation rules are becoming more complex, and more complex evaluation formulas and evaluation rules are being aimed at in order to further improve accuracy. Even if we try to develop evaluation rules, human ability is inevitably limited, and if possible, it is very difficult.

The present application has been made in view of these points, and it is not necessary for humans to consider the allocation algorithm at all, and moreover, a determination system for automatically determining an optimal allocation car according to various traffic situations is automatically generated. It is an object of the present invention to provide a new elevator group management device that is completely different from the conventional one.

[Means for Solving the Problems]

A feature of the present invention is that a neural net (described later in detail) is applied to allocation control.

That is, the present invention is a means for calculating an input pattern to the neural net from the hall call and the state of each car,
It is characterized in that it is provided with a means for calculating a neural net according to the input pattern and a means for judging an allocation based on the calculation result of the neural net.

Furthermore, the neural net is composed of an input layer, an intermediate first layer, an intermediate second layer, and an output layer, and the units of the input layer and the intermediate first layer are connected to each other only for each unit related to its own unit. The units of the intermediate first layer and the intermediate second layer are connected to each other so that the units related to the own machine and the units related to other machines are also connected to each other, and the connection structure and weight setting of each machine Is symmetrical to improve the efficiency and accuracy of arithmetic processing.

〔Example〕

An embodiment of the present invention will be described below. First, a neural network will be briefly described.

A neural network is a network that imitates the human brain, in which multiple units corresponding to neurons (nerve cells) in the brain are joined intricately, and the operation of each unit and the connection form between units are By making good decisions,
A pattern recognition function and a knowledge processing function can be embedded, and they are introduced, for example, in "Nikkei Electronics," August 10, 1987 issue (No. 427), P115 to P124.

First, FIG. 7 shows the structure of a unit that models a neuron. The unit Ui transforms the total sum of the input Qj from other units according to a certain rule to obtain Qi, and the joints with other units have variable weights Wij respectively. This weight is for expressing the strength of the coupling between each unit, and if this value is changed, the structure of the network will be substantially changed without changing the connection. The learning of the network described later is to change this value, and the weight Wij takes positive, zero, and negative values. Zero means no bond.

When a unit receives inputs from multiple units, if the sum of the inputs is represented by NET, the sum of the inputs of unit Ui is Is.

Each unit applies the sum NET of this input to the function f and converts it into the output Qi as shown in the following equation.

This function f may be different for each unit, but generally the threshold function shown in FIG. 8 (a) or the sigmoid function shown in FIG. 8 (b) is used.

This sigmoid function is a differentiable pseudo-linear function, Can be expressed as The value range is 0 to 1, and approaches 1 as the input value increases and approaches 0 as the input value decreases. When the input is 0
It becomes 0.5. Add a threshold θ (bias), In some cases.

FIG. 9 is a diagram showing an example of the structure of the network, and the illustration of the weight of the connecting portion between the units is omitted.

The neural network is classified into a pattern associative type and an automatic associative type according to the structure of the network, and the pattern associative type is used in the present invention. The pattern associative type is a network that converts an input pattern into a certain output pattern.
As shown in the figure, each unit is hierarchized into an input layer, middle layer, and output layer. Each unit is connected from the input layer to the output layer, but the units in each layer are not connected. Moreover, the input unit and the output unit are independent.

In such a neural net, when input data is given to each unit of the input layer, this signal is converted by each unit, transmitted to the intermediate layer, and finally comes out from the output layer. It is necessary to set the strength of coupling between the units, that is, the weight, to an appropriate value. The weight is set by learning the network as follows.

First of all, set all weights randomly,
Input data for learning (data in which desired output is known in advance) is given to each unit in the input layer. Then, at this time, the output value output from each unit of the output layer is compared with the desired output value, and the value of each weight is corrected so as to reduce the difference (error). Then, this is repeated using a large number of learning data until the error converges. A learning algorithm for correcting each weight value from this error value will be described later.

When the learning is completed in this way, the knowledge processing function is automatically embedded in the neural net, and the desired output can be always obtained not only for the learning data but also for the unknown input data.

The present invention utilizes this neural net for elevator allocation control. Note that the neural net can be configured as an LSI by configuring each unit using an amplifier and a resistor, but it is also possible to configure a virtual neural net by software and process all by arithmetic operations. In the embodiment, a case of realizing with software will be described.

FIG. 1 is a diagram showing an embodiment of the entire configuration of the present invention. Here, for convenience of explanation, the elevators to be controlled are Nos. 1 to 3
The number of units is three, but the present invention can be similarly applied to any number of units.

In FIG. 1, 1 is a hall call button (1
Only one floor is shown, the others are omitted) 2 is a landing call signal, 3A is an operation control device that manages the operation of Unit 1,
Similarly, 3B and 3C are operation control devices that manage the operation of Units 2 and 3, respectively, and 4 is the state of each car (car position, direction,
A car information signal indicating stop, traveling, door open / closed state, car call, load, etc.) 5 is a microcomputer for performing a function of allocation as a county management device, and a hall call signal 2 read through the input / output interface 6 Based on the data of the car information signal 4, the function as the input pattern calculation means 7 for converting the traffic situation at the time when a new hall call is generated as an input pattern to the neural network, and the input pattern calculation means 7 for each layer. It has a function as a neural net computing means 8 for sequentially computing the outputs of the units, and a function as an assignment determining means 9 for deciding and allocating which machine is optimal from the output values of the units in the output layer. The result is output as the allocation signal 10 via the input / output interface 6. Each of the operation control devices 3A to 3C controls the operation of the car so as to sequentially respond to the hall call assigned by the assignment signal 10 and the car call registered in its own machine.

FIG. 2 is a flow chart showing the processing procedure of allocation according to the present invention.

In Fig. 2, when the program starts, the procedure is
Determine whether there is a new hall call on M11,
If so, proceed to step M12.

In step M12, information about the state of each car (position, direction, etc.), floor, direction of hall call, etc. is read. Then, in step M13, each data of the input pattern to the neural network is calculated based on these information.

An example of each data of the input pattern is shown in FIG. As shown in FIG. 3, each data of the input pattern indicates the floor and direction of the hall call newly generated from the hall call signal,
In addition, the car position, the traveling direction, the load, etc. are input to the input pattern calculating means 7 from the car information signals of the respective units, and are calculated as follows. For example, each data ▲ P ⁽¹⁾ ₁ ▼ to ▲ P ⁽¹⁾ ₄ ▼ of the input pattern for Unit ¹ is ▲ P ⁽¹⁾ ₁ ▼: Floor and direction of new landing call and car position and driving direction of Unit 1. The floor difference (the number of floors) is obtained from and, and it is normalized in the range of 0 to 1.

▲ P ⁽¹⁾ ₂ ▼: Of the hall calls assigned to Unit 1 and the car calls of Unit 1, count the number of calls that exist between the car position of Unit 1 and the floor of the new hall call, and count it. Normalize to the range of 0-1.

▲ P ⁽¹⁾ ₃ ▼: Count the number of calls farther from the floor of the new hall call among the hall calls assigned to Unit 1 and set it to 0.
Normalize to a range of ~ 1.

▲ P ⁽¹⁾ ₄ ▼: Normalize the load in the car of Unit 1 to the range of 0 to 1.

To ask.

Here, the normalization to the range of 0 to ¹ is performed by, for example, ▲ P ⁽¹⁾ ₁ ▼
If, ▲ P ⁽¹⁾ ₄ ▼, It can be obtained by dividing by the maximum possible value.

The input patterns ▲ P ⁽²⁾ ₁ ▼ to ▲ P ⁽²⁾ ₄ ▼ of the No. 2 machine and the input patterns ▲ P ⁽³⁾ ₁ ▼ to ▲ P ⁽³⁾ ₄ ▼ of the No. 3 machine are similarly obtained.

From the input pattern thus obtained,
Calculate the net.

An example of the structure of the neural network applied to the present invention is shown in FIG. Where ▲ U ⁽¹⁾ ₁ ▼ to ▲ U ⁽¹⁾ ₄ ▼, ▲ U ⁽²⁾ ₁
▼ to ▲ U ⁽²⁾ ₄ ▼, ▲ U ⁽³⁾ ₁ ▼ to ▲ U ⁽³⁾ ₄ ▼ are each data of the input pattern in each unit of the input layer ▲ P ⁽¹⁾ ₁ ▼
~ ▲ P ⁽¹⁾ ₄ ▼, ▲ P ⁽²⁾ ₁ ▼ ~ ▲ P ⁽²⁾ ₄ ▼, ▲ P ⁽³⁾ ₁ ▼ ~
It corresponds to ▲ P ⁽³⁾ ₄ ▼. Also, ▲ U ⁽¹⁾ ₅ ▼, ▲ U
⁽¹⁾ ₆ ▼ ~ ▲ U ⁽³⁾ ₅ ▼, ▲ U ⁽³⁾ ₆ ▼ are the middle and first layer units, ▲ U ⁽¹⁾ ₇ ▼, ▲ U ⁽¹⁾ ₈ ▼ ~ ▲ U ^{(3 )} ₇ ▼, ▲ U ⁽³⁾ ₈
▼ is the middle second layer unit, ▲ U ⁽¹⁾ ₉ ▼, ▲ U ⁽³⁾ ₉ ▼
Is the unit of the output layer, and the output of the unit ▲ U ⁽¹⁾ ₉ ▼ Q
⁽¹⁾ ₉ ▼ is the first unit, unit ▲ U ⁽²⁾ ₉ ▼ output ▲ Q ⁽²⁾ ₉
▼ corresponds to Unit 2, and unit ▲ U ⁽³⁾ ₉ ▼ output ▲ Q ⁽³⁾ ₉ ▼ corresponds to Unit 3. In this example, since each data of the input pattern is normalized to the range of 0 to 1 as described above, each unit of the input layer does not perform any processing,
Although the input data is output as it is, it is possible to use the number of floors and the number of calls as input data as they are and normalize them with the input / output characteristics of each unit in the input layer. It is assumed that each unit in each intermediate layer and each output layer has the input / output characteristics shown in FIG. 8 (b) described above. Of course, it may be a characteristic in which a threshold value (bias) is added to this as in the former case. The value of the weight that indicates the degree of coupling between units is
It is obtained in advance by learning as described later. Therefore, if each value of the input data is determined, the output of the neural net shown in FIG. 4 can be calculated.

First, in this example, since each unit of the input layer outputs the input data as it is as described above and is input to each unit of the intermediate first layer, the output of each unit of the intermediate first layer is sequentially output from this input. Calculate and obtain (procedure of Fig. 2
M14). For example, the total sum of the inputs of the unit ▲ U ⁽¹⁾ ₅ ▼ is NE ▲
If T ⁽¹⁾ ₅ ▼ and the output is ▲ Q ⁽¹⁾ ₅ ▼, Can be calculated as However, ▲ W ⁽¹⁾ _i5 ▼ is the unit of the input layer ▲ U ⁽¹⁾ _i ▼ and the unit of the intermediate first layer ▲
It represents the weight between U ⁽¹⁾ ₅ ▼.

When the calculation is completed for each unit of the intermediate first layer, each output becomes an input to each unit of the intermediate second layer, and therefore, the output is calculated for each unit of the intermediate second layer (procedure M15). ).

Similarly, the output is calculated for each unit in the output layer (procedure M16). Then, the optimum number is selected from the output of each output unit.Here, the value of each weight is set in advance in the learning process, and the output unit corresponding to the assigned number is assigned to the learning input pattern. If it is considered optimal to allocate the desired output (output target) of No. 1 to the desired output of another output unit and assign it to the No. 1 machine for an input pattern for learning, the unit ▲ U ^{( 1)} Set the output target of ₉ ▼ to 1 and set the output target of units ▲ U ⁽²⁾ ₉ ▼ and ▲ U ⁽³⁾ ₉ ▼ to 0, and learn until the error becomes sufficiently small for many learning patterns. Since the weight after repeating is set, the output of the output unit corresponding to the optimal allocation machine is closest to 1 as a result of the calculation in step M16.
Therefore, of the outputs ▲ Q ⁽¹⁾ ₉ ▼ and ▲ Q ⁽³⁾ ₉ ▼, you may decide to assign the one with the highest output value, but the output of any output unit is below a predetermined value (eg 0.5). In other words, it may be the case that none of the units are appropriate for allocation.Therefore, first, in step M17, it is judged whether or not there is a predetermined value or more in each output unit, and if there is, in step M19. The allocation is determined for the machine with the largest value, and if there is no more than the specified value, recalculation is performed by the backup allocation algorithm (for example, allocation by the conventional evaluation function) in step M18, and the most of the evaluation value is determined in step M19. The allocation is decided to be the largest or the smallest unit.

In the above embodiment, only static elements are used as each data of the input pattern, but dynamic (temporal) elements such as predicted waiting time of each hall call and elapsed time after the hall call occurs. You may make it take in etc. Further, in the above-mentioned embodiment, each data of the input pattern is data for each machine, but, for example, data common to the entire group such as the number of waiting passengers on each floor and the occurrence frequency of hall calls is added as one of the input patterns, You may make it connect it from the unit of an input layer to each unit of an intermediate | middle layer.

The network structure of the neural net is not limited to that shown in FIG. 4 (both connection and number of units),
Instead of connecting all the units between layers as in the general network structure shown in FIG.
As shown in the figure, first, between the input layer and the intermediate first layer, only the units corresponding to the input data of the own machine are mutually connected, and then the first and second intermediate layers are not connected to each other. If the units corresponding to the units are connected to each other and the units are connected symmetrically, the number of calculations can be reduced and the accuracy can be improved. That is, in the case of an elevator, since data having completely different properties are used as the input pattern data, for example, data regarding the floor difference between the new call and Unit ¹ ▲ P ⁽¹⁾ ₁ ▼ and data regarding the load of Unit 2 ▲ Q ^{(2) This} is because not only is it meaningless to connect data with completely different properties from other units, such as ₄ ▼, but it is also possible that noise will occur rather as the number of calculations increases. . In addition, the connection as shown in FIG. 4 allows each input data to be set according to the weight set in the learning process without considering what is important for assignment and what is not important for input data for the first machine. In the state in which the importance of is taken into consideration, data will be output from units ▲ U ⁽¹⁾ ₅ ▼ and ▲ U ⁽¹⁾ ₆ ▼ as data related to Unit 1, so this can be output in the second layer in the middle. By comparing with the corresponding signals of the units, the relative evaluation between the units can be performed more accurately with the minimum connection configuration and therefore in the shortest time. Also, there is symmetry in the relationship between the status of each machine and allocation (for example, if allocation to machine 1 is optimal in a certain state, similar conditions exist in machine 2 and machine 3). By making the units connected symmetrically with respect to each machine, the processing for each machine can be performed in the same manner, and the weights can be set symmetrically to improve the efficiency of learning.

Next, network learning will be described. The characteristics of the neural net are determined by the configuration and weight of the connection between each unit, but since the configuration of the connection is fixed here, it is necessary to set the weight. This weight setting uses learning by back propagation known as a general setting method. Below, the fourth
The neural net having the configuration shown in the figure will be described.

First, the weight values between the units are randomly set to appropriate values. Then, many learning samples (combinations of input patterns and output targets) are created, and these are read into the learning computer. This learning sample assumes, for example, one combination of the car states of Nos. 1 to 3 and a new hall call, and in this state the number of cars to be assigned is best judged by an expert or using a simulation or the like. If the No. 1 unit is optimal, set the output target of the unit in the output layer ▲ U ⁽¹⁾ ₉ ▼ of the No. 1 unit to 1 for the input pattern at that time, and set the output targets of the other units in the output layer. It was created as 0.
Learning is performed using the large number of learning samples. An example of the processing procedure is shown in the flowchart of FIG.

First, the index N indicating the sample number is set to 0 (procedure
S11), next, N = N + 1 is set (procedure S12), and the processes from step S13 onward are sequentially performed from sample number 1. In step S13, the input pattern and output target of the first sample are first set, and in step S14 the output of the neural net for this input pattern is calculated. Since the weights are set randomly at the beginning, the output value of each unit in the output layer at this time is different from the output target, so the difference between this output target and the actual output is taken as the error. Calculate (Procedure S1
Five). Error in each unit of output layer ▲ E ^(α) ₉ ▼
Is the output target of each unit is ▲ T ^(α) ₉ ▼ ▲ E ^(α) ₉ ▼ = ▲ T ^(α) ₉ ▼-▲ Q ^(α) ₉ ▼ (α
Is the machine number).

In the operation of the neural network, the input data is sequentially processed from the input layer to the output layer, but the correction of the weight is performed from the output layer to the input layer by using the error.

When the output error is obtained, it is used to correct the weight between each unit of the intermediate second layer and the output layer (step S16).
If the weight between the unit of the middle second layer ▲ U ^(α) _k ▼ and the unit of the output layer ▲ U ^(α) ₉ ▼ is ▲ W ^(α) _k9 ▼,
The correction is calculated as follows.

Δ ▲ W ^(α) _k9 ▼ = λ ・ ▲ E ^(α) ₉ ▼ ・ ▲ Q ^(α) ₉
▼ + (1-λ) ・ Δ '▲ W ^(α) _k9 ▼ Δ ▲ W ^(α) _k9 ▼: ▲ W ^(α) _k9 ▼ current correction amount Δ '▲ W ^(α) _k9 ▼: ▲ W ^(α) _k9 ▼ previous correction amount (initially 0) λ: Learning rate (constant, for example 0.5) ▲ Q ^(α) _k ▼: Output of the unit ▲ U ^(α) _k ▼.

For example, the corrected value of the weight ▲ W ⁽¹⁾ ₇₉ ▼ between the unit ▲ U ⁽¹⁾ ₇ ▼ and the unit ▲ U ⁽¹⁾ ₉ ▼. Is Δ ▲ W ⁽¹⁾ ₇₉ ▼ = λ · ▲ E ⁽¹⁾ ₉ ▼ · ▲ Q ⁽¹⁾ ₇ ▼ + (1-λ) · Δ '▲ W ⁽¹⁾ ₇₉ ▼.

Similarly, when the correction of the weight between each unit of the intermediate second layer and the output layer is completed, the error of each unit of the intermediate second layer is calculated (procedure S17), and the error is used to calculate the intermediate first layer. The weight between the layer and the intermediate second layer is modified (step S18). This correction is calculated as follows.

The unit ▲ U ^(α) _k ▼ of the intermediate second layer is calculated by using the error ▲ E ^(α) ₉ ▼ of the output layer, Becomes

At this time, the weight ▲ W is placed between the unit ▲ U ^(α) _j ▼ of the intermediate first layer and the unit ▲ U ^(α) _j ▼ of the intermediate second layer.
^(Α) _jk ▼ corrected value Is Δ ▲ W ^(α) _jk ▼ ＝ λ ・ ▲ E ^(α) _k ▼ ・ ▲ Q ^(α) _j
▼ + (1-λ) · Δ '▲ W ^(α) _jk ▼, and between the unit of the first intermediate layer ▲ U ^(β) _j ▼ and the unit of the second intermediate layer ▲ U ^(α) _k ▼. Weight ▲ W
The corrected value of ^{(β, α)} _jk ▼ Is Δ ▲ W ^{(β, α)} _jk ▼ = λ ・ ▲ E ^(α) _k ▼ ・ ▲ Q
^(Β) _j ▼ + (1-λ) · Δ ′ ▲ W ^{(β, α)} _jk ▼. However, β represents the machine number like α, and α ≠
β. Other symbols are the same as those described above.

For example Unit ▲ U ⁽¹⁾ ₇ ▼ error ▲ E ⁽¹⁾ ₇ ▼ is From this, the weight value between the units ▲ U ⁽¹⁾ ₅ ▼ and ▲ U ⁽¹⁾ ₇ ▼ is the corrected value of ▲ W ⁽¹⁾ ₅₇ ▼. Is Δ ▲ W ⁽¹⁾ ₅₇ ▼ = λ ・ ▲ E ⁽¹⁾ ₇ ▼ ・ ▲ Q ⁽¹⁾ ₅ ▼ + (1-λ) ・ Δ '▲ W ⁽¹⁾ ₅₇ ▼ Also, the unit ▲ U ⁽²⁾ Weight between ₅ ▼ and ▲ U ⁽¹⁾ ₇ ▼ ▲
W ^(2,1) ₅₇ ▼ corrected value Is Calculate as Δ ▲ W ^(2,1) ₅₇ ▼ = λ ・ ▲ E ⁽¹⁾ ₇ ▼ ・ ▲ Q ⁽²⁾ ₅ ▼ + (1-λ) ・ Δ '▲ W ^(2,1) ₅₇ ▼ respectively. it can.

Similarly, if all the weights between the intermediate first layer and the intermediate second layer are corrected, then the error of each unit of the intermediate first layer is calculated (procedure S19), and that value is used to calculate the error between the input layer and the intermediate layer. The weight between layers is corrected (procedure S20). The calculation of this correction is performed as follows.

The error ▲ E ^(α) of the unit ▲ U ^(α) _j ▼ of the intermediate first layer
_j ▼ is Next, unit ▲ U of the input layer ^(alpha) _i ▼ and intermediate first layer unit ▲ U ^(alpha) weight between the _{^{j ▼ ▲ W (α) ij}} ▼
The modified value of Is Δ ▲ W ^(α) _ij ▼ = λ · ▲ E ^(α) _j ▼ · ▲ Q ^(α)
_i ▼ + (1−λ) · Δ ′ ▲ W ^(α) _ij ▼.

For example Unit ▲ U ⁽¹⁾ ₅ ▼ error ▲ E ⁽¹⁾ ₅ ▼ is From this, the corrected value of the weight ▲ W ⁽¹⁾ ₁₅ ▼ between the units ▲ U ⁽¹⁾ ₁ ▼ and ▲ U ⁽¹⁾ ₅ ▼ Is It can be calculated as Δ ▲ W ⁽¹⁾ ₁₅ ▼ = λ ・ ▲ E ⁽¹⁾ ₅ ▼ ・ ▲ Q ⁽¹⁾ ₁ ▼ + (1-λ) ・ Δ '▲ W ⁽¹⁾ ₁₅ ▼.

Similarly, if all the weights between the input layer and the intermediate first layer are modified, it means that the modification of the weights by the first learning sample is completed, and the procedure returns from step S21 to step S12, and then 2
The above procedure is repeated for the second learning sample using the corrected weights. When all the N learning samples prepared in this way are completed, it is confirmed in step S22 whether the error in the output layer has become sufficiently small. If the error has not converged, the procedure returns to step S11 to repeat learning with N learning samples again, and when the error has converged, the learning is completed.

Then, when the weight set by the completion of learning is used, as described above, the input patterns other than the learning sample are assigned with the same judgment criteria as the learning sample.

Next, a method for simplifying the establishment of the weight by this learning by utilizing the property peculiar to the elevator will be described.

In an elevator, since each car generally has symmetry, for example, in the case of three elevators, considering an input pattern that is best assigned to the first car, the input data for the second car of this input pattern is considered. Even if the input data for Unit 3 and 3 are exchanged, allocation to Unit 1 is also optimal. In addition, if the input data for the No. 1 machine of this input pattern is replaced with the input data for the No. 2 machine, the optimum input pattern will be assigned, and if it is replaced with the input data of the No. 3 machine, it will be optimally allocated to the No. 3 machine. It becomes the input pattern. That is, in the case of three elevators, six kinds of learning samples can be considered for one input pattern. Therefore, originally, a large number of learning samples should be used with the six kinds of learning samples as one set. It is necessary to perform learning, but if learning is repeated until the error converges in this way, the weight corresponding to the same connection position of each machine will eventually become the same value among the setting values of each weight. Is expected. For example, in the example of FIG. 4, ▲ W ⁽¹⁾ ₅₇ ▼ and ▲ W ^(2,1) ₅₇ ▼ are different values, but ▲ W ⁽¹⁾ ₅₇ ▼ and ▲ W ⁽²⁾ ₅₇ ▼, ▲ ^{Three of} W ⁽³⁾ ₅₇ ▼ have the same value, and ▲ U ⁽¹⁾ ₁₅ ▼ and ▲ U ⁽¹⁾ ₂₅ ▼ have different values.
It is expected that three of ▲ U ⁽¹⁾ ₁₅ ▼, ▲ U ⁽²⁾ ₁₅ ▼, and ▲ U ⁽³⁾ ₁₅ ▼ will have the same value. Also, ▲ U ^(1,2) ₅₇ ▼ and ▲ U ^(2,1) ₅₇ ▼, or ▲ U ^(2,3) ₆₈ ▼ and ▲ U
^(3,2) ₆₈ ▼ are expected to have the same value. Therefore, for the weights that are expected to eventually become the same value, set the same value from the beginning,
Further, the number of times of learning can be reduced by treating the same value in the learning process.

For example, regarding the weight ▲ W ^(α) ₅₇ ▼ between the units ▲ U ^(α) ₅ ▼ and ▲ U ^(α) ₇ ▼, the correction amount Δ first calculated by the above-described procedure for a certain learning sample. ▲ W
^{If (1)} ₅₇ ▼, Δ ▲ W ⁽²⁾ ₅₇ ▼, Δ ▲ W ⁽³⁾ ₅₇ ▼ are used as temporary values, the average of the following three values is calculated and calculated as the actual correction amount. After correction Is always corrected to the same value for each unit. Similarly, ▲ W
^{(1, 2)} ₅₇ ▼ and ▲ W ^(2,1) in the case of ₅₇ ▼ modifications, each modification amount Δ ▲ W ^(1,2) ₅₇ ▼ and ▲ W ^(2,1) ₅₇ ▼ mean value Is the actual correction amount, ▲ W ^(1,2) ₅₇ ▼ and ▲ W
^(2,1) ₅₇ ▼ will always be corrected to the same value. When learning using the symmetry of the elevator in this way,
In this case, one learning sample has the same effect as that obtained by learning about six types of learning samples, and the number of learning sample sets to be prepared, that is, the number of times of learning can be greatly reduced.

Next, a case will be described where learning is performed online instead of offline, that is, while the elevator system is operating.

Since the traffic conditions in each building differ depending on the nature and use of the building, when learning is done offline, it is necessary to prepare a learning sample specific to each building and repeat learning for each building, which is a very complicated work. Will be forced. In addition, if the weight settings are fixed due to the results of learning conducted offline, and if the traffic conditions of the building change significantly from the middle due to tenant changes etc., appropriate allocation may not be performed. There is.

In order to solve the problem caused by the offline learning, the weight setting is first learned offline by using a standard learning sample common to each building, and then (after the elevator is installed), the elevator setting is performed. It is necessary to create a learning sample peculiar to the building during the operation of, and perform learning automatically and correct the weights at any time. This online learning procedure will be described with reference to the flowchart of FIG.

First of all, the neural net is calculated by the weight set in the off-line learning, and the new hall call is assigned. When the allocation is done, the pair of the input pattern at that time and the allocated machine number is registered as one sample in the temporary sample collection (composed of RAM etc.) (procedure R11, R1).
2). Information about the call and the assigned number is detected and stored until the assigned number is cleared by the assigned number responding to the call or when the call is fully passed by the assigned number (procedures R13 to R15). The information on the call and the assigned number is, for example, the elapsed time from the occurrence of the call, the fact that another number has arrived first or stopped on the floor of the call, and the assigned number has passed through the floor.

When the assigned number is erased, for example, when the assigned number responds to the call, it is determined in step R16 whether the assigned number was appropriate. This is based on predetermined conditions, such as the fact that the waiting time was longer than a predetermined time, that another unit arrived first, or that there was an unloaded item on that floor. However, if it is true, the assignment is inappropriate, that is, the sample registered in step R12 is deleted as inappropriate for learning (step R1
7) Otherwise, leave the sample as is and proceed to step R18. And procedure R11 ~ until a certain time passes
Repeat R18, while registering a large number of samples in the same procedure as above. After a certain period of time, the sample collection created in the above procedure is added to a predetermined standard sample collection registered in advance, and this is used as a learning sample, and relearning is performed (procedure R19). The learning procedure is exactly the same as the offline learning procedure shown in FIG. When learning is completed, the temporary sample collection is cleared (procedure R20), samples are created again while assigning the modified weights, and the above procedure is repeated to automatically correct the weights.

By incorporating online learning in this way, offline learning is common to each building, so it only needs to be done once, and after that, with a sample unique to each building, it is always automatically corrected to the optimum weight for that building. Therefore, even if traffic conditions change due to changes in tenants, etc., it is possible to automatically respond.

〔The invention's effect〕

According to the present invention, it is possible to automatically create a judgment system that determines an optimal allocation car according to various traffic situations by simply creating a large number of learning samples, and to perform extremely advanced allocation control. It becomes possible to do.

Further, by devising the connection structure of the neural net and the value of the weight, the time required for the calculation and the number of learning can be reduced.

In addition, using online learning,
Standardization can be achieved regardless of the application, and even if the property of the building changes in the middle, it can be automatically handled.

[Brief description of drawings]

FIG. 1 is a diagram showing an embodiment of the overall configuration of the present invention, and FIG.
FIG. 4 is a flow chart showing an allocation processing procedure according to the present invention, FIG. 3 is a diagram showing an example of an input pattern according to the present invention, FIG. 4 is a diagram showing an example of a structure of a neural net according to the present invention, and FIG. FIG. 6 is a flowchart showing a processing procedure of weight learning according to the present invention, FIG. 6 is a flowchart showing a processing procedure when performing weight learning according to the present invention online, and FIG. 7 is a diagram for explaining the present invention. FIG. 8 is a diagram showing the structure of the unit, FIGS. 8 (a) and 8 (b) are diagrams showing an example of the input / output characteristics of the unit, and FIG. 9 is a diagram showing the structure of a general neural net. 1 …… Hall call button 2 …… Hall call signal 3A to 3C …… Operation control device for each unit 4 …… Car information signal 5 …… Microcomputer 6 …… Input / output interface 7 …… Input pattern calculation means 8 …… Neural network computing means 9 ... Assignment determining means 10 ... Assignment signal

Claims (4)

[Claims] 1. A group management of elevators in which a plurality of elevators are activated for a plurality of floors, and an optimal elevator is selected and responded to by responding to a hall call common to the plurality of elevators. In the control device, means for calculating the input pattern to the neural net from the hall call and the state of each car, means for calculating the neural net based on the input pattern, and determination of allocation based on the calculation result of the neural net An elevator group supervisory control device, comprising: 2. A group management of elevators in which a plurality of elevators are activated for a plurality of floors and an optimal elevator is selected and responded to by responding to a hall call common to the plurality of elevators. In the control device, means for calculating the input pattern to the neural net from the hall call and the state of each car, means for calculating the neural net based on the input pattern, and determination of allocation based on the calculation result of the neural net And means for creating training samples while the elevator is running,
A group management control device for an elevator, comprising means for performing learning using the learning sample and correcting the weight of the neural net. 3. A neural net comprises at least an input layer, an intermediate first layer, an intermediate second layer and an output layer, and each unit of the input layer and each unit of the intermediate first layer relates to its own machine for each machine. Connect only unit comrades to each other, the first intermediate
The group management of elevators according to claim 1 or 2, wherein each unit of the layer and each unit of the second intermediate layer are configured to mutually connect a unit related to its own unit and a unit related to another unit. Control device. 4. A means for calculating an input pattern to a neural net from a hall call and the state of each car, a means for calculating a neural net based on the input pattern, and an allocation based on the calculation result of the neural net. The neural network of an elevator equipped with a means for making a decision
In the process of learning the net, each weight corresponding to the same connection position of each machine is set to the same value, and the correction is performed by using the average value of the correction amount of each weight for one learning sample as the actual correction amount. A method of learning a neural net of an elevator, characterized in that