JPH08310746A - Learning method for waiting time predicting neural net - Google Patents

Abstract

PURPOSE: To efficiently conduct learning and predict the waiting time with high accuracy by using the system state data of a group when a hall call occurs as the input pattern, and using only the waiting time of the allocated elevator for the hall call as the teacher signal to generate the learning sample. CONSTITUTION: This waiting time predicting neural net is constituted of a three-layer structure: input layer, intermediate layer, and output layer to cope with the number of elevators, e.g. three. The system state data of a group (data on all elevators) are used as the input patter, only the waiting time of the allocated elevator is used as the teacher signal to generate the learning sample, and the learning sample is used for learning. Only the connection Weight connected to the neuron applied with the teacher signal is corrected based on the error between the teacher signal and the output in the output layer. The corrected value is reflected on the connection weight connected to the neuron of the other output layer located at the position symmetric with the connection weight.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method of a waiting time prediction neural network for predicting the waiting time of a hall call in elevator group management control.

[0002]

2. Description of the Related Art Hitherto, the group management control of elevators has been mainly performed by a call assignment method using an evaluation function and a call assignment control by an expert system using a fuzzy theory. A new method has been proposed in which call allocation is performed using the above-mentioned neural network, and it is also proposed to use the neural network to predict the waiting time of a hall call.

A neural network is a network that imitates the human brain. A plurality of nerve cell models (neurons) are connected in a complicated manner, and pattern recognition is performed by deciding the operation of each neuron and the connection form between neurons. Functions and knowledge processing functions can be embedded, for example, "Nikkei Electronics" 1987.
It is disclosed in P115-P124 of the August 10, issue (No. 427) and the book "PDP model" published by Sangyo Tosho Co., Ltd. in February 1989, in which neurons are arranged in a hierarchical structure. Is characterized by the availability of an autonomous learning algorithm called "back propagation".

When this neural network is used, it is not necessary for a human to consider an allocation algorithm or a waiting time prediction algorithm, and in addition, a decision system for determining an optimal allocation car in response to various traffic situations. There is an excellent effect that a system for predicting the waiting time and the waiting time can be automatically generated. For example, as an example used for elevator call assignment, Japanese Patent Laid-Open No. 1-275381, "Elevator Group Management Control Device", 3-311
No. 73 “Elevator group management control device”, Japanese Patent Application No. 5-2
No. 43817 "Learning method for neural net for elevator call assignment". Further, examples used for the waiting time prediction include JP-A-2-286581 "Elevator group management control device" and JP-A-5-246633 "Elevator group management control device construction method".

An example of the call assignment neural network is shown in FIG. As shown in FIG. 5, the call assignment neural net NN has an input layer NR1 corresponding to an input pattern (system state data of a group of elevators).
And the output layer N corresponding to the output pattern (appropriation of allocation)
R3 and an intermediate layer NR2 placed between the input layer and the output layer
And neurons.

The input patterns are various data representing the system status of the elevator group (call allocation such as the distance from the current position of each unit to the floor where the new hall call is generated, the planned number of stops on the way, the congestion situation in the car, etc.). The data that seems to be necessary) are converted into a form that can be input to the neural network, and the number of neurons in the input layer corresponds to the total number of the data.

When input data is given to each neuron of the input layer, a signal is sequentially transmitted to the output layer, and as a result, some value is output from each neuron of the output layer. In the output layer, there are as many neurons as the number of elevators, and for various input patterns, the neuron corresponding to the number that is optimal for allocation is "1" (or maximum value), and the other neurons are "0". Pre-learned to output.

Therefore, after the learning, it is shown that among the values of each neuron in the output pattern, the neuron that outputs the value closest to "1" (or the maximum value) is the most suitable for allocation, The number corresponding to this neuron is selected as the assigned number. The number of neurons in the intermediate layer (the number of which is one in the embodiment, but may be plural) is appropriately determined according to the number of elevators, the property of the building, and the like.

Although not shown, connection weights (synaptic weights) representing the connection strength of neurons are set between the neurons. This connection weight is initially set to an appropriate value, but it can be modified thereafter by using a learning algorithm called "backpropagation" so that more accurate call allocation can be performed.

Since this back propagation is well known, a detailed description thereof will be omitted, but a learning sample prepared in advance (an input pattern and a desired output pattern corresponding to the input pattern, that is, a teacher signal is paired. , Which compares the output pattern for the same input pattern with the teacher signal, and modifies the connection weights to minimize the error. Initially, all weights are initialized (for example, at random). Set to a certain value) and give the learning pattern input pattern to each neuron in the input layer. Then, the output pattern at this time and the output pattern of the learning sample (teacher signal)
And the difference (error) is compared, and the value of each connection weight is corrected in order from the output layer side so that the difference becomes smaller.

When this is repeated until the error converges using a large number of learning samples, the call assignment function of the same level as the teacher signal is automatically embedded in the neural network, and the learning input is used. Not only the pattern but also the unknown input pattern can be assigned the call at the same level as the teacher signal.

In the same way as above, the neural network for waiting time prediction also uses various data representing the state of the elevator system when the hall call occurs as an input pattern, and the actual waiting time for it as a teacher signal. By similarly learning, it becomes possible to predict the waiting time at the same level as the teacher signal for an unknown input pattern.

[0013]

By the way, when a neural network for predicting the waiting time of a hall call is created,
The simplest method is shown in FIG. 6, in which only the information regarding the number (or first-arrival unit) to which the hall call is assigned is input to the neural network, and the waiting time of the hall call for the assigned number (first-arrival unit) is output. It is conceivable to have such a configuration.

However, the states of machines other than the allocated machine (first-come-first-served machine) are not affected at all. For example, if there is another elevator in front of the assigned car (tentatively designated as No. 1), it will be difficult for the calls that occur thereafter to be allocated to No. 1, and the waiting time for the hall call will be when there is no other elevator in front. It is more likely to be shorter than Therefore, in order to improve the accuracy of waiting time prediction, it is necessary to input information on other machines than the allocated machine (first-come-first-served machine) to the neural network, and the configuration shown in FIG. 7 can be considered. .

In this case, it can be easily imagined that it is better to fix the position for inputting the information of the assigned vehicle (first-arrival vehicle) in view of learning efficiency. However, the neural network as shown in FIG. 7 has no symmetry in its structure and is not universal. For example, as shown in the above-mentioned Japanese Patent Laid-Open No. 5-246633, when a call assignment neural network is created based on a neural network in which the waiting time is learned,
The waiting time prediction neural network having the structure shown in FIG. 7 has the following problems.

When learning is performed using the same learning sample (which requires less memory), the layout of the input patterns must be changed in the waiting time prediction neural network and the call allocation neural network. -When switching the waiting time prediction neural network to the call allocation neural network, the structure between the output layer and the intermediate layer must be changed. Therefore, as shown in FIG. 8, it is conceivable to output the predicted waiting time of all the cars so as to have a structure having symmetry, but in that case, how to make the teacher signal becomes a problem.

In other words, it is difficult to define the waiting time (time required for arrival) other than the first-arriving car in response to the hall call. Although it is conceivable to use the time of actual arrival or the time of passage, the operation of the elevator after that will change depending on whether the hall call is assigned or not. In particular, if you turn over in the middle floor or stop in an empty basket, the value will be significantly different from the expected value. That is,
The waiting time (time required for arrival) for a car that cannot be assigned a hall call cannot be correctly known.

Therefore, when the neural network for waiting time prediction is constructed as shown in FIG. 8 and the learning sample is used for learning, the waiting time of all machines is given as a teacher signal in one learning sample. Then, there is a risk of learning meaningless information, which not only deteriorates learning efficiency but also decreases prediction accuracy.

[0019]

In order to solve the above-mentioned problems, in the present invention, the system status data of a group (data relating to all units) is used as an input pattern, while the teacher signal waits for an assigned unit (first unit). A learning sample with only time is created, and learning is performed using this learning sample.
At this time, in the output layer, only the connection weight connected to the neuron to which the teacher signal is given is corrected based on the error between the teacher signal and the output, and the corrected value is set at a position symmetrical to the connection weight. It is characterized in that it is reflected in the connection weight connected to the neuron of the other output layer in.

[0020]

In the present invention, at the time of learning, in the output layer, only the connection weight connected to the neuron to which the teacher signal is given is corrected based on the error, and the connection weight is connected to the neuron of another output layer. For the weight, the value after the above correction is used as it is.

[0021]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram (partially omitted) of the structure of a neural network for waiting time prediction according to the present invention. Here, as in FIG. 8, an input layer (first layer) and an intermediate layer (second layer). The output layer (third layer) has a three-layer structure, and the number of elevators is three. FIG. 2 is a diagram showing the structure of each neuron.

1 and 2, U _mi ^(K) is the output from the i-th neuron in the m-th layer of ^K , and Q _mi ^(K) is the output from the i-th neuron in the m-th layer of K. , P _i ^(K) is the i-th input data of Unit K (P _i ⁽⁰⁾ is the i-th input data for the whole group), and W _mij ^{(K, L)} is the i-th unit of the m-th layer of Unit K. And the j of the m-1th layer of L machine
The connection weights with the th neuron are shown.

FIG. 3 is a flow chart showing the procedure for creating a learning sample according to the present invention. First, in step S1, it is determined whether or not a new hall call is generated. When a call is generated, the call is confirmed in step S2. The time of occurrence is stored, and the system state data of the group at the time of occurrence of the call is stored in step S3.

When it is confirmed in step S4 that the car has stopped in response to the call, in step S5 the difference between the current time and the call generation time is calculated to calculate the actual waiting time for the hall call. . Then, in step S6, the system state data of the group at the time the call is generated and the actual waiting time are registered as a learning sample. When the above procedure is repeated each time a hall call is generated, a large number of learning samples in which only the waiting time of the assigned machine (or the first-arrival machine) is used as the teacher signal is registered.

The learning sample preparation procedure described above may be performed while the elevator is actually operating in the field or by simulation. At that time, the call may be assigned by any of a method using a neural network, a method using a conventional evaluation function, a method using a fuzzy rule, and the like.

FIG. 4 is a flow chart showing the learning procedure of the neural network according to the present invention, and the learning is performed by the above learning sample. First, in step S11, a variable N representing a sample number is initialized to 0, and step S12
Increments the value of N to N + 1. Step S1
In step 3, the input pattern and output target of the Nth learning sample are set, and in step S14, the operation of each layer is sequentially performed on the Nth sample from the input layer of the neural network.

Then, in step S15, the error between the output of the output layer and the output target is calculated, and in step S16, the connection weight between the intermediate layer and the output layer is corrected based on the error. Now, for example, in step S14, if the N-th learning sample is that the second machine is the assigned machine (first-come-first-served machine), the teacher signal is given only to the neurons in the output layer of the second machine. Only the error between the output of the neuron in the output layer (third layer) Q ₃₁ ⁽²⁾ and the teacher signal is calculated, and in step S16, based on this error, U ₃₁ ⁽ the output layer neuron of Unit 2 ^). Modify the connection weights connected to ²⁾ in sequence.

At this time, the neuron U in the output layer of Unit 1
The modification of the connection weight connected to ₃₁ ⁽¹⁾ and the connection weight connected to neuron U ₃₁ ⁽³⁾ in the output layer of Unit ³ is because the elevator generally has symmetry with respect to each unit. The correction result for Unit 2 can be used as it is. For example, W ₃₁₁ ^{(1,1) at} symmetrical positions
And W ₃₁₁ ^(3,3) can be the same value as W ₃₁₁ ^(2,2), and W ₃₁₁ ^(1,2) and W ₃₁₁ ^(1,3) , W ₃₁₁ ^(2,1) , W
₃₁₁ ^(2,3) , W ₃₁₁ ^(3,1) and W ₃₁₁ ^(3,2) can have the same value.

In order to always make the connection weights at the symmetrical positions have the same value, the result may be copied each time the correction is made based on the error. If the connection weights are configured to point to the same address in the memory, the correction for one connection weight will be immediately reflected in the other equivalent connection weights, which can be performed more easily and efficiently.

Then, in step S17, step S1
Based on the correction result of the weight of 6, the error is calculated for each neuron of the intermediate layer, and each connection weight between the input layer and the intermediate layer is sequentially corrected in step S18 based on the error. Thus, the above procedure is repeated, and step S1
When it is confirmed in step 9 that all the samples have been completed, the process goes through step S20 and is repeated until the error converges.
When the steps 11 to S20 are repeated and the error is sufficiently converged, the learning is completed, and thereafter, this neural network can be used for waiting time prediction.

[0031]

According to the present invention, only one teacher signal is given to neurons in a plurality of output layers, but in reality, learning is equivalent to giving a teacher signal to all output layer neurons. Can be done, and learning can be done efficiently. Also, the assigned number (first-come-first-served number) as a teacher signal
Since only the waiting time is given, the waiting time can be predicted with high accuracy.

[Brief description of drawings]

FIG. 1 is a diagram showing a configuration of a waiting time prediction neural network according to the present invention.

FIG. 2 is a diagram showing a structure of a neuron.

FIG. 3 is a flowchart showing an example of a procedure for creating a learning sample according to the present invention.

FIG. 4 is a flowchart showing a learning procedure of a neural network according to the present invention.

FIG. 5 is a diagram showing an example of a neural network for call assignment.

FIG. 6 is a diagram showing a configuration of a waiting time prediction neural network.

FIG. 7 is a diagram showing a configuration of a waiting time prediction neural network.

FIG. 8 is a diagram showing a configuration of a waiting time prediction neural network.

[Explanation of symbols]

U ₁₁ ^{(0) to} U ₁₁ ⁽³⁾ , U ₁₂ ^{(0) to} U ₁₂ ⁽³⁾ Each neuron in the input layer U ₂₁ ^{(1) to} U ₂₁ ⁽³⁾ , U ₂₂ ^{(1) to} U ₂₂ ^{(3 )} Each neuron in the intermediate layer U ₃₁ ^{(1) to} U ₃₁ ⁽³⁾ Each neuron in the output layer P ₁ ^{(0) to} P ₁ ⁽³⁾ , P ₂ ^{(0) to} P ₂ ⁽³⁾ Each data of the input pattern Q ₃₁ ^{(1) to} Q ₃₁ ⁽³⁾ Output of Units 1 to 3 (predicted waiting time) W _mij ^{(K, L)} connection weight

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Ichiro Nagashima 1-28-10 Sho, Ibaraki, Osaka Prefecture Fujitec Co., Ltd. (72) Inventor Takeshi Midoriya 1-28-10 Sho, Ibaraki, Osaka Fujitec (72) Inventor Tomoaki Tanabe 1-28-10 Sho, Ibaraki, Osaka Prefecture Fujitec Co., Ltd. (72) Inventor Naoki Ota 1-28-10 Sho, Ibaraki, Osaka Fujitec Co., Ltd. Within

Claims (3)

[Claims] 1. When a hall call occurs, the system state data of the elevator group at that time is converted into a form that can be used as an input pattern of a neural network and input, and as an output pattern, prediction of each unit for the hall call. In a learning method of a waiting time prediction neural network that outputs a waiting time, the system state data of the group at the time when a hall call occurs is used as an input pattern, and only the waiting time of the assigned machine (or first arrival machine) for that hall call is used as a teacher signal. A learning method for a waiting time prediction neural network, wherein the learning sample is created and learning is performed using the learning sample. 2. When a hall call occurs, the system state data of the elevator group at that time is converted into a form that can be used as an input pattern of a neural network and input, and the output pattern of each unit's prediction for the hall call. In a learning method of a waiting time prediction neural network that outputs a waiting time, the system state data of the group at the time when a hall call occurs is used as an input pattern, and only the waiting time of the assigned machine (or first arrival machine) for that hall call is used as a teacher signal. When a learning sample is created and learning is performed using the learning sample, in the output layer, only the connection weight connected to the neuron to which the teacher signal is given is corrected based on the error between the teacher signal and the output. Then, the corrected value is connected to a neuron in another output layer that is symmetric to the connection weight. A learning method for a waiting time prediction neural network, which is characterized in that it is reflected in the existing connection weight. 3. The waiting time prediction neural network according to claim 2, wherein the reflecting means has a structure in which a value of a connection weight at a symmetrical position indicates a same address on a memory in advance on a program. Learning method.