JPH08208136A - Learning of neural-net for elevator call assignment - Google Patents

Abstract

PURPOSE: To shorten learning time by selecting the neural-net of a building with a specification most similar to a new building from a data base if a controller for group control in the new building requires to learn a new neural-net, and learning it as the initial net. CONSTITUTION: If a controller for group control in a new building requires to learn a new neural-net, the neural-net of the building with a specification most similar to that of the new building is selected from a data base, and is incorporated into the controller for group control. Identification of whether or not the specification of the building is most similar to that of the new building is conducted, for example, in the following order: 1) the number of cars is the same, 2) the number of floors is the same or similar, 3) the nature of a building is the same, 4) the characteristic of a floor to be served is similar, and 5) the specification of an elevator is the same. The selected neural-net is learned as the initial net.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a learning method for newly creating a neural network in elevator group management control for assigning calls using a neural network.

[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 a call is assigned by using the neural network described above.

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.
August 115 (No 427) P115-P124 and 1
It is disclosed in the book "PDP model" published by Sangyo Tosho Co., Ltd. in February 989, and in particular, those in which neurons are arranged in a hierarchical structure can use an autonomous learning algorithm called "back propagation". There are features.

By using this neural network, it is not necessary for a human to consider the allocation algorithm at all, and in addition, a decision system for deciding the optimum allocation car can be automatically generated in response to various traffic situations. For example, as an example used for call assignment of elevators, JP-A-1-275381 "Elevator group management control device" and JP-A-3-31173 "Elevator group management control device", Japanese Patent Application No. 5-2438
No. 17 "learning method of elevator call assignment neural network" and the like.

An example of the call assignment neural network is shown in FIG. As shown in FIG. 5, the call assignment neural net NN includes an input layer NR1 corresponding to an input pattern (elevator system state data) and an output layer NR3 corresponding to an output pattern (allocation suitability).
It is composed of neurons of the intermediate layer NR2 placed between the input layer and the output layer.

The input pattern uses various data representing the state of the elevator system (floor call generation floor and direction, position and operating direction of each unit, car call, load state, etc.) as parameters necessary for call assignment. , Is converted into a form that can be input to a neural network, and the number of neurons in the input layer corresponds to the total number of its parameters.

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 machine that is optimal for allocation outputs "1" (maximum value), and the other neurons output "0". Have been learned in advance to do.

Therefore, among the values of each neuron of the output pattern, it is shown that the Nyuron which outputs the value closest to "1" (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 neurons in the embodiment is one, but may be more than one) 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.

[0012]

Conventionally, the learning of this assignment neural net is performed by setting a blank condition, that is, the connection weight of the neural net to a random value, or by fixing a preset net as an initial net. I set it to the value of and started learning from that state.

For this reason, there is a problem that it takes a long time for learning to converge, and it takes a long time especially for group management with a large number of cars and floors. The present invention has been made in view of such problems, and an object thereof is to reduce the learning time when newly creating a neural network.

[0014]

In order to achieve the above object, in the first invention of the present application, a learning sample created by a simulation operation of an elevator or a learning sample created while an elevator is operating in an actual building. A large number of neural nets that have been trained using are stored as a database, and when learning a new neural network for a new building, select the neural network of the building with the most similar specifications from the building from the database. Then, the learning is performed using this as an initial net.

Further, in the second invention of the present application, a plurality of neural nets are selected from the database and installed in the group management device, and the learning sample collected after the elevator is operated is used to determine the degree of agreement with the teacher signal. Are compared for each neural net, and the one with the highest degree of matching among them is used as the initial net for learning.

[0016]

According to the first invention of the present application, when the learning of a new neural network for a new building becomes necessary, the neural network of the building having the most similar specifications to the building is selected from the database, Learning is performed by using it as the initial net.

According to the second invention of the present application, a plurality of neural nets having the most similar specifications to the building are selected from the database, and the learning sample collected in the building is used to determine the degree of agreement with the teacher signal. Are compared,
The neural net with the highest degree of matching is selected as the initial net.

[0018]

Embodiments of the present invention will be described below. First, a procedure for creating a learning sample after the operation of the elevator, learning the neural network, and then registering the learned neural network in the database will be described with reference to the flowchart of FIG.

First, in step S1, it is determined whether or not a new hall call is generated, and when the call is generated, step S2 is performed.
Then, the elevator system state data at the time the call is generated is converted into an input pattern of the neural network. When it is confirmed in step S3 that the call has been serviced, the serviced machine is converted into an output pattern of the neural network as an allocated machine in step S4. At this time, if the number of units initially assigned is different from the number of units actually serviced due to allocation change or reassignment, that is, if the initial assignment was not the best as a result (however, the service was switched to dedicated operation Except when there was no),
Convert the actual serviced number to the output pattern as the assigned number. Note that the allocation during this period is performed by the conventional method using an evaluation function, fuzzy group management, or the like.

Then, in step S5, the input pattern and the output pattern (teacher signal) are paired and stored as a learning sample. Further, the above procedure is repeatedly executed through step S6, and when the number of learning samples reaches a predetermined number, learning of the neural network is performed using this. When the completion of learning is confirmed through step S8, the learned neural net is stored in the memory in the file format in step S9. In this way, the maintenance staff downloads the neural network stored in the memory for each building using a maintenance tool or the like, takes it home and registers it in the database.

The contents registered in this database are as follows:
The connection structure of the neural network, each connection weight, and the specifications of the building. The specifications of the building include the number of floors, the number of floors, the character of the building (whether it is a condominium, a multi-tenant building, a company-owned building, etc.), the characteristics of the service floor (such as express zones or floors with incomplete service), and elevator specifications ( Include all data that affects the allocation of group management controls, such as rated speed and loading). In this way, many learned neural nets are registered in the database.

Next, when shipment of the group management control device to a new building is decided and learning of a new neural network becomes necessary, the neural network of the building having the most similar specifications to the building is selected from the above database. And incorporate it into the group management control device. In that case, when determining whether the specifications are most similar to each other, priority is given to the item that is considered to have the largest influence on the allocation performance, and the determination is performed in the following order, for example.

1. The same number of cars. 2. The number of floors is the same or close. 3. The character of the building is the same. 4. Those with similar service floor characteristics. 5. Elevator specifications are the same.

Then, learning is started by using the selected neural net as an initial net. This learning may be before the installation of the elevator or after the operation of the elevator, but before the installation of the elevator, a learning sample is created by simulating the operation of the elevator in the new building, After the operation of the elevator, as described above, it is carried out by creating learning samples while allocating in the conventional method.

By the way, even if a neural network with the most similar building specifications is selected as the initial net from the database as described above, the selected neural network is not always the initial net depending on the traffic situation of the new building. It may not be optimal.

Therefore, the following method can be taken as a solution. That is, in the same manner as above, a plurality of neural nets that are considered to have the most similar building specifications are selected and incorporated into the group management device of the new building. Then, the degree of coincidence with the teacher signal is compared for each neural net by using the learning sample collected after the elevator is operated, and the neural net with the highest degree of coincidence is selected as the initial net. The procedure of this method is shown in the flowcharts of FIGS.

First, after a plurality of selected neural nets are incorporated into the group management device, a predetermined number of learning samples are collected after the operation of the elevator in steps S21 to S26 as in the case described above. Then, for the first neural network (steps S27 to S29), the first learning sample is applied to the neural network to calculate the output value for the input pattern (steps S30 to S29).
32), an error from the teacher signal is calculated and stored (step S
33). Then, the same procedure is repeated for all the samples (step S34), and the average value of the error is calculated for the first neural network (step S35).

This is repeated for all the neural networks selected in the same manner (step S36), and the average value of the error is calculated for each neural network.
Then, the net having the smallest average error value is determined to have the highest degree of coincidence with the teacher signal, and this is set as an initial net, and learning is started for this net (step S37).

In this embodiment, the average value of the errors is calculated as a criterion for judging the degree of coincidence with the teacher signal, but the present invention is not limited to this. For example, the hit rate of the assigned machine for all samples is calculated. May be used.

FIG. 4 is a diagram comparing the convergence time in the case of learning by the conventional method with the convergence time in the case of performing the learning according to the present invention. As is clear from the figure, in the present invention, the learning time becomes T2 by appropriately selecting the initial net in contrast to the conventional learning time T1, and the convergence time of learning is significantly larger than that in the conventional method. You can see that it will be shortened.

[0031]

According to the present invention, the neural network having the most similar specifications is selected from among the neural nets which have already been trained, and the learning is performed by using the neural net as the initial net. It can be shortened.

Further, a plurality of neural nets are incorporated in advance, the degree of coincidence with the teacher signal is compared, and the neural net having the highest degree of coincidence is selected from among them, so that the specifications are similar, but the traffic is similar. Even for buildings with different situations, the optimal initial net can be selected to significantly reduce the learning time.

[Brief description of drawings]

FIG. 1 is a flowchart showing a procedure for registering a neural network in a database according to the present invention.

FIG. 2 is a flowchart showing a procedure for selecting an initial net in the present invention.

FIG. 3 is a flowchart showing a procedure of selecting an initial net in the present invention.

FIG. 4 is a diagram showing a comparison of convergence times between the present invention and a conventional learning method.

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

[Explanation of symbols]

 NN Call assignment neural network NR1 Input layer neuron NR2 Intermediate layer neuron NR3 Output layer neuron

 ─────────────────────────────────────────────────── ─── Continuation of the front page (72) Inventor Ichiro Nagashima 1-28-10 Sho, Ibaraki-shi, Osaka Inside Fujitec Technical Research Institute (72) Inventor Takeshi Midoriya 1-28-10 Sho, Ibaraki, Osaka Fujitec Technical Laboratory Co., Ltd. (72) Inventor Tomoaki Tanabe 1-28-10 Sho, Ibaraki City, Osaka Prefecture Fujitec Technical Laboratory Co., Ltd. (72) Naoki Ota 1-28-10 Sho, Ibaraki City, Osaka Stocks Company Fujitec Technical Research Institute

Claims (3)

[Claims] 1. When a hall call occurs, various data representing the state of the elevator system at that time are converted into a form that can be used as an input pattern of a neural network and input, and an optimum car is selected from the output pattern. In the learning method of the neural network for call assignment of the group management elevator, which is assigned to the hall call, the learning sample created by the simulation operation of the elevator or the learning sample created while the elevator is operating in the actual building. A large number of stored neural networks are stored as a database, and when learning a new neural network for a new building, select the neural network of the building with the most similar specifications from the building from the database, and initialize it. Characterized by learning as a net Learning method of neural network for elevator call assignment. 2. When a hall call occurs, various data representing the state of the elevator system at that time are converted into a form that can be used as an input pattern of a neural network and input, and an optimum car is selected from the output pattern. In the learning method of the neural network for call assignment of the group management elevator, which is assigned to the hall call, the learning sample created by the simulation operation of the elevator or the learning sample created while the elevator is operating in the actual building. A large number of neural networks are stored as a database, and when learning a new neural network for a new building, multiple neural networks of the building with the most similar specifications to the building are selected from the database and group management is performed. Learning collected in the equipment and collected after the elevator is in operation Neural network for elevator call assignment, characterized in that the degree of coincidence with the teacher signal is compared for each neural net using the training sample, and the one with the highest degree of coincidence is used as the initial net for learning. How to learn the net. 3. The elevator call allocation according to claim 1 or 2, wherein whether or not the specifications of buildings are similar is determined in the order of the number of cars, the number of floors, and the character of the building. Method for learning neural network.