JP2500407B2 - Elevator group management control device construction method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an elevator group management control device for assigning calls using a neural network, and more particularly to a group management control device capable of efficiently adapting the neural network to each site. Regarding the construction method.

[0002]

2. Description of the Related Art Conventionally, a call assignment method using an evaluation function has been the mainstream of elevator group management control, but recently, an approximation problem has been approximated by a simple algorithm based on a model of a biological neural circuit. The concept of "neural network" that performs analysis and pattern recognition is drawing attention.
Using this neural network, we propose a new elevator group supervisory control device that is completely different from the conventional one, which can automatically learn and generate a decision system that determines the optimal assigned car based on actual driving examples. Has been done.

A neural network is a network that imitates the human brain. A plurality of nerve cell models (neurons) are connected intricately, 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” August 1, 1987.
0th issue (No. 427) P115-P124 and 1989
"PDP" published by Sangyo Tosho Co., Ltd.
A model having a structure in which neurons are arranged in a hierarchical structure is characterized in that an autonomous learning algorithm called "back propagation" can be used.

When the neural network is used, it is not necessary for a human to consider the allocation algorithm at all, and moreover, it is possible to automatically generate a judgment system for deciding the optimum allocation car in response to various traffic situations. For example, Japanese Patent Laid-Open No. 1-
275381, JP-A-2-52875, 3-Kaikai 3-
59971 and the like.

FIG. 9 shows a typical example of a call assignment neural network. As shown in FIG. 9, a neural network NN for call assignment includes an input layer neuron NR ₁ corresponding to an input signal (system state data), an output layer neuron NR ₃ corresponding to an assignment machine, an input layer and an output. It is composed of neurons NR _{2 of the} intermediate layer placed in the middle of the layer.

[0006] Input signals are used for call allocation such as a floor where a new hall call is generated, a position of each elevator, a driving direction, another hall call registration floor, a car call registration floor, a car load, and the number of waiting passengers at a hall. This is system state data of a group considered necessary, and the number of neurons in the input layer is determined according to the information amount of the input signal. The neurons in the output layer correspond to each machine, that is, the same number as the number of elevators, and the machine corresponding to the neuron whose output signal (appropriateness of allocation) has the largest value is the optimum machine, and Calls are assigned. 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 in the initialized state, but can be modified thereafter by using a learning algorithm called "back propagation" so that more accurate call allocation can be performed.

In the back propagation, for the same input signal, the output signal of the neural network is compared with the teacher signal created in advance, and the connection weight is corrected so as to minimize the error. In the algorithm, first
Initially, all connection weights are initialized (set to random values, for example), and input data for learning (the same as the input data of the teacher signal) is given to each neuron in the input layer.
Then, the output signal of each neuron in the output layer at this time is compared with the output target of the teacher signal, and the difference (error) is used to correct the value of each connection weight in order from the output layer side so that the difference becomes smaller. I will do it. When this is repeated using a large number of learning data until the error converges, the call assignment function at the same level as the teacher signal is automatically embedded in the neural network, and not only the learning data but also the unknown The input data (new hall call and system state data at that time) can be assigned the same level of call assignment as the teacher signal.

[0009]

In order to learn by using the backpropagation method as described above, a teacher signal is always required, but in the case of elevator call assignment, an optimum teacher signal is obtained. Is very difficult. When a new hall call occurs, it is relatively easy to find the optimal assigned car at that time from the car position of each car and the situation of other calls, but in reality the allocated car While responding to, various unpredictable changes will occur in traffic conditions, such as new calls or longer stoppages on the intermediate floors. It is impossible to get a quota solution.

Moreover, the method of performing learning by back propagation has a problem that it is not possible to perform call assignment control with higher accuracy than the teacher signal. Therefore, the so-called reinforcement learning method (reinforcem
ent learning) can be used for self-organization of neural networks. This reinforcement learning method is also well known, so a detailed description thereof will be omitted. However, first, all the connection weights are initialized, and then any one or more connection weights are perturbed. That is, the value of the connection weight is slightly changed. Then, the output of the neural network or the evaluation value of the system controlled by the neural network is compared and evaluated before and after the perturbation is given, and the perturbation is accepted if it is improved, and the perturbation is given if it is not improved. Returns to the previous state. When this procedure is repeated in this way, each connection weight gradually converges toward the optimum value, but the biggest drawback of this method is that the learning efficiency is very poor. Therefore, it is very difficult to apply this reinforcement learning method to the neural network for call assignment as it is.

Further, as disclosed in, for example, Japanese Unexamined Patent Publication No. 2-52875, optimal operation is analyzed for all hall calls generated during the day, and learning of the neural network is performed based on this analysis. A method of doing so has also been proposed. This method attempts to find the problem of finding an optimal solution from all combinations of allocations for all hall calls for one day as a combinatorial optimization problem that is well known in the traveling salesman problem. However, the solution obtained by this method is only an approximate solution, and a method for obtaining an optimal solution has not been established yet.In addition, this method allows self-organization of the neural network during the actual operation of the elevator. There is a problem that you cannot do it.

[0012]

In order to solve the above problems, the present invention constructs a call assignment neural network in three stages. In the first stage, a neural network having the traffic condition data of the group as an input signal and having a highly accurate teacher signal, which is closely related to the call assignment and has a predicted waiting time of a hall call as its output signal, is created. The learning is performed by using the predicted waiting time as a teacher signal.

In the second stage, the output layer is reconfigured for call assignment while leaving the input layer and the intermediate layer of the neural network for which learning has been completed, and the output layer is learned as a teacher signal for call assignment suitability. In the third stage, the neural net that has completed the second stage is applied to the site for call assignment, and its output layer is adapted to the site by the reinforcement learning method during operation of the elevator.

[0014]

In the first stage, a neural network that outputs information closely related to call assignment is constructed so that it can be used later for call assignment. Moreover, since learning is performed using accurate teacher signals, a neural network with high prediction accuracy can be constructed.

Since the neural network in the first stage cannot be used for call assignment as it is, only the output layer is reconfigured for call assignment in the second stage. At the end of the first step, the input layer and the intermediate layer are configured as a neural network with high prediction accuracy, and at the end of the second step, the output layer also completes learning so that call allocation is possible. Therefore, in the third stage, reinforcement learning is performed only on the output layer.

[0016]

An embodiment of the present invention will be described below with reference to the drawings. FIG. 1 is a flowchart showing the procedure of the first stage of constructing a neural network according to the present invention. First,
It is closely related to call assignment, that is, the input signal can be shared with the call assignment, and the structure of the neural network that uses the information that can obtain a highly accurate teacher signal as the output signal and the determination of the input signal (or Fix).
(Procedure 11).

Here, the predicted waiting time for a new hall call is considered as the above output signal. As mentioned above, it is difficult to obtain the optimal allocation solution for hall calls even if all the daily operations are analyzed. However, regarding the estimated waiting time for hall calls, how will traffic conditions change during that time? And since the actual waiting time, that is, the correct solution can be obtained when the car answers the call, and the input signal required to obtain the expected waiting time is considered to be almost the same as in the case of call assignment, A neural network having the predicted waiting time as an output signal is constructed.

An example of the structure of this neural network is shown in FIG.
Shown in Since it is used for call allocation later, the configuration is similar to that of FIG. 9, and each neuron in the output layer corresponds to the predicted waiting time of each machine for the hall call. That is, the neural network is constructed so that the new hall call and the system state data at that time are input as the input signal and the predicted waiting time of each unit is obtained as the output signal.

Returning to FIG. 1, next, learning is performed using the back propagation method (procedure 12). In that case, a teacher signal is required, but the estimated waiting time for hall calls is
Since the actual waiting time data is obtained when the car responds to the hall call, this is used as the teacher signal. A large number of this teacher signal may be created in advance by simulation, or actual measurement data may be registered at the installation site of the elevator.

When the learning is completed, it is judged whether or not the prediction accuracy of the predicted waiting time is sufficient by actually using this neural network (procedure 13). If the prediction accuracy is not sufficient, return to step 11 to modify the neural network by changing or adding an input signal, and then to steps 12 and 13.
repeat. Then, when the prediction accuracy becomes sufficiently high, the first stage is ended.

At the time when the first stage is completed, this neural network uses the output as the predicted waiting time for the hall call, and since the learning is performed by using the accurate teacher signal, various traffic after the hall call is generated. This means that the predictive function for changes in the situation is embedded with high accuracy.

In the above example, only the predicted waiting time is considered as the output of the neural network. However, the number of neurons in the output layer is further increased to closely relate to the call assignment such as the load in the car at the time of answering the hall call. However, it is also possible to add, as an output signal, information by which an accurate teacher signal can be obtained. Also, the predicted waiting time is not limited to the predicted waiting time for a newly generated hall call, and when the new hall call is assigned, the predicted waiting time that is the maximum among the hall calls handled by the car is used as the output signal. You may do it.

FIG. 3 is a flowchart showing the procedure of the second stage of constructing a neural network according to the present invention. The output of the neural network constructed in the first stage is the estimated waiting time and cannot be used for call assignment as it is, so only the output layer is reconfigured for call assignment (procedure 2).
1). That is, as described in the first step, the input layer and the intermediate layer are the same as in the case of call assignment, and the prediction function is already embedded with high accuracy by learning using the accurate teacher signal. Therefore, it is used for call assignment as it is, and only the output layer is reconfigured for call assignment.
An example of the reconstructed neural net is shown in FIG.

In FIG. 4, the number of neurons in the output layer is the same as that of the neural net in FIG. 2 since each neuron corresponds to each machine. Therefore, in this example, reconstruction means the number of neurons and the connection configuration. 2 means that only the connection weights of the output layer (weights between the intermediate layer and the output layer) are initialized with the state of FIG. If there are multiple types of output signals and the number of neurons in the output layer is larger than the number of cages in the neural network of FIG. 2, the number of neurons in the output layer is reduced and the output layer is changed including the change of connection. Will be reconfigured.

Here, after the reconstruction of the output layer, that is, the initialization of the connection weights, only the connection weights of the output layer are modified by the back propagation method using the teacher signal of the call assignment as in the conventional method. (Procedure 22). The call assignment teacher signal consists of a pair of an input signal and an output target for it. For example, it corresponds to a new hall call and the input signal such as the car position of each unit at that time, which is considered to be optimal. The output target of the neuron is set to 1, and the output targets of other units are set to 0. The criteria for determining the optimum allocation when creating this teacher signal may be any such as the call allocation based on the same evaluation function as the conventional one, or the call allocation based on the fully automatic group-coupling method. .

When a large number of these teacher signals have been created in advance and learning has been completed (procedure 23), the second stage ends. At this time, as described in the above-mentioned Japanese Patent Laid-Open No. 1-275381, the learning can be efficiently performed if the learning is performed in consideration of the symmetry between the machines. At the time when the second stage is completed, the neural network shown in FIG. 4 has the function of accurately predicting the change in traffic condition after the occurrence of a new hall call in the input layer and the middle layer as described above. Moreover, since the learning for the call assignment of the output layer is completed in the second stage, the call assignment can be performed using this neural network. Therefore, in this state, the neural network shown in FIG. 4 is incorporated into the group supervisory control device and the process proceeds to the next third stage.

FIG. 5 is a flow chart showing the procedure of the third step of constructing a neural network according to the present invention. In the third stage, call assignment is performed using the neural network shown in FIG. 4, and unnecessary reinforcement learning of the teacher signal is performed for only the output layer as follows.

First, the output layer of the neural network is initialized (procedure 31). Next, using this neural network, call allocation control is performed for a predetermined time (or for a predetermined number of hall calls), and a control evaluation index value within that time (or a predetermined number of hall calls) is calculated and stored. (Procedure 3
2). As the control evaluation index value (hereinafter, simply referred to as an evaluation value), for example, an average value of the actual waiting time of hall calls assigned within that time or a predetermined number of hall calls can be considered.

Next, a perturbation is given to the connection weight of the output layer of the neural network. That is, the value of any one or a plurality of connection weights is slightly changed (procedure 33). Then, call allocation control is performed again for a predetermined time or a predetermined number of hall calls, and the evaluation value is calculated in the same manner as in step 32 (step 34).

The calculation result is compared with the evaluation value stored in the procedure 32 (procedure 35), and if not improved, the connection weight is returned to the state before perturbation (procedure 3).
6) Then, the procedure 33 and subsequent steps are repeated again. If the evaluation value is improved, the perturbation given in step 33 is accepted, and the evaluation value calculated in step 34 is stored. Repeat steps 33 to 37 until learning is completed in this way.
The function of the neural network is gradually improved by accepting the perturbation only when the evaluation value is improved as a result of perturbing the connection weight. Then, when the evaluation value becomes a sufficiently satisfactory value, or when the evaluation value does not show any improvement even if the perturbation is given, the learning is completed (procedure 38), and the third stage is ended.

The procedures of the third step can be performed by simulation or under actual group control while performing group management control. Since the calculation of the evaluation value in that case can limit the occurrence of a large number of hall calls in the case of simulation, the comparison of the evaluation value before and after perturbing the connection weight can be easily performed. However, it is impossible to reproduce exactly the same situation of hall calls when it is done on-site, so the evaluation values are compared using a statistical method (for example,
The method of determining whether there is a significant difference by student test etc. is well known).

By using the neural network which has completed the third stage in this way, it is possible to perform highly accurate call allocation, and if the third stage procedure is performed regularly or continuously thereafter while performing group management control on site. , It is possible to constantly adapt to changes in traffic conditions in buildings.

By the way, in the above-mentioned embodiment, since data relating to all the machines are input as the input signals of the neural network, and the neurons in the output layer correspond to the respective machines, the number of elevators for group management increases. It is necessary to increase the number of neurons accordingly, and it is difficult to standardize the connection structure of the neural network. However, as shown in FIG. 6, which of the two machines has the highest allocation suitability is output. By using a neural network that can be represented as follows, call assignment can be performed using the same neural network regardless of the number of elevators, and the present invention can be applied to the construction of this neural network.

The neural network of FIG. 6 has the first allocation
A neural network corresponding to a candidate car and a neural network corresponding to a car to be compared are combined to have a common output layer neuron. For example, in this neural network, the input signal I ₁ associated with the first candidate car
(Car position, driving direction, car load, distance to new hall call, call in charge) and input signal I ₂ related to the car to be compared are input to each neuron of each input layer N ₁ and N _2. When input, the output of the neuron in the output layer is learned so as to be equal to or more than the threshold when the first candidate car has a higher allocation suitability.

FIG. 7 shows the procedure for assigning a call using this neural network. First, the input signal of each car is calculated and stored (procedure 41). Next, for example, the first car is temporarily set as the first candidate car, and its input signal I
₁ is input to each neuron of the input layer N ₁ (procedure 42),
The second car is set as a car for comparison and its input signal I ₂ is input to each neuron of the input layer N ₂ (procedure 43). Then, the output of the neural network is calculated (procedure 44), and it is judged whether or not the value is larger than the threshold value (procedure 4).
5).

If it is larger than the threshold value, a predetermined score is added to the first candidate car, that is, the first car (procedure 46),
If it is smaller than the threshold value, the procedure directly returns to step 43 through step 47, and this time the output of the neural net is calculated using the third car as the comparison target car. Thus the first
When all the comparisons between the candidate car, that is, the first car and the other candidate cars are completed (procedure 47), the procedure 4 is performed again through the procedure 48.
Returning to step 2, this time, the second car is set as the first candidate car, and the steps 43 to 47 are repeated with the other cars sequentially as the cars to be compared.

When all the cars are finished as the first candidate car and all the combinations in which the other cars are compared are completed (procedure 48), it is determined that the car with the maximum score has the highest allocation suitability. It is assigned (procedure 49). As described above, when the neural network shown in FIG. 6 is used, the call can be assigned only by repeatedly using the neural network shown in FIG. 6 regardless of the number of elevators, and the neural network can be easily standardized. (A call assignment method using such a neural network is disclosed in Japanese Patent Application No. 2-134902). However, even when such a neural network is used, the present invention can be applied. In addition, the high-performance neural network can be efficiently constructed as follows by the procedures of the first to third steps.

First, in the first stage, as shown in FIG.
A neural network is constructed in which system state data when a new hall call is generated for a car is used as an input signal, and a predicted waiting time for the hall call or a call already assigned is used as an output signal. Since it is possible to easily obtain actual measurement data for the estimated waiting time, this is used as a teacher signal and learning is performed using the back propagation method.

In the second stage, two sets of neural nets shown in FIG. 8 are used, only the output layer is made common, and only the output layer is learned by using the teacher signal for call assignment. This teacher signal can be easily created using the conventional call assignment determination criteria, as in the case described above.

In the third stage, reinforcement learning is performed only on the output layer as in the case described above, and the construction of the neural network for call assignment is completed.

[0041]

As described above, according to the present invention, learning is performed by using the predicted waiting time or the like that can obtain a teacher signal more accurate than the teacher signal of call assignment. Since it is embedded and learning is performed by using the reinforcement learning method for call assignment, it is possible to perform call assignment control with higher accuracy than before.

Further, according to the present invention, it is possible to automatically adapt even during operation of the elevator according to the situation of each site and its change, and in that case, the reinforcement learning is limited to only the output layer. Since it is carried out in a very efficient manner, it has an excellent effect of being able to adapt to changing circumstances very efficiently and quickly.

[Brief description of drawings]

FIG. 1 is a flowchart showing a procedure of a first stage of constructing a neural network according to the present invention.

FIG. 2 is a diagram showing an example of a configuration of a neural network in the first stage of the present invention.

FIG. 3 is a flowchart showing a procedure of a second stage of constructing a neural network according to the present invention.

FIG. 4 is a diagram showing an example of a configuration of a neural network in a second stage of the present invention.

FIG. 5 is a flowchart showing a procedure of a third stage of constructing a neural network according to the present invention.

FIG. 6 is a diagram showing another embodiment of the configuration of the neural network for call assignment.

7 is a flowchart showing the procedure of call assignment control using the neural network of FIG.

8 is a diagram showing an example of a configuration of a neural network used for constructing the neural network shown in FIG.

FIG. 9 is a diagram showing an example of a general configuration of a call assignment neural network.

[Explanation of symbols]

NN Neural network NR ₁ Input layer neuron NR ₂ Intermediate layer neuron NR ₃ Output layer neuron

Claims (2)

(57) [Claims] 1. A neural network for call assignment, which activates a plurality of elevators for a plurality of floors, inputs system status data of a group, and outputs assignment suitability of each unit for newly generated hall calls. In the method of constructing an elevator group supervisory control device that selects and allocates the most suitable car by the output signal of, the system status data of the group is used as an input signal, and the expected waiting time of each unit for hall calls is output as its output signal. Of a neural network which has a first waiting step for learning the neural network using the predicted waiting time as a teacher signal, and a neural network for which learning in the first step has been completed.
A second stage in which only the output layer is reconfigured for call assignment while the input layer and the intermediate layer are left unchanged, and the output layer is learned as a call assignment suitability as a teacher signal; and the neural that has completed the learning in the second stage Applying the net to the site for call assignment, and adapting the output layer to the site by the reinforcement learning method while the elevator is in operation The third step is to construct a neural network for call assignment. A method for constructing a characteristic elevator group management control device. 2. The elevator group supervisory control system according to claim 1, wherein an actual waiting time when the car responds to the hall call is used as the predicted waiting time which is the teacher signal in the first stage. How to build.