JP2964902B2 - Learning method of neural network for elevator call assignment - 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 system for allocating calls using a neural network, and more particularly to a method of learning a neural network for the allocation.

[0002]

2. Description of the Related Art Conventionally, elevator group management control has mainly been a call allocation method using an evaluation function or a call allocation control by an expert system using fuzzy logic. A new method of assigning calls using a neural network has been proposed.

A neural network is a network imitating the human brain. A plurality of neural cell models (neurons) are connected in a complex manner, and the operation of each neuron and the form of connection between the neurons are properly determined to perform pattern recognition. The function and the knowledge processing function can be embedded. For example, "Nikkei Electronics" 1987
P115-P124 and 1 of August 10, issue (No. 427)
It is disclosed in the book "PDP model" published by Sangyo Tosho Co., Ltd. in February 989, etc. In particular, those in which neurons are arranged in a hierarchical structure can use an autonomous learning algorithm called "back propagation". There are features.

When this neural network is used, there is no need for a human to consider an assignment algorithm at all, and a decision system for deciding an optimum assigned car in response to various traffic conditions can be automatically generated. For example, Japanese Patent Application Laid-Open No. 1-275381 "Elevator group management control device" and Japanese Patent Application Laid-Open No. 3-31173 "Elevator group management control device" Japanese Patent Application No. 5-2438
No. 17, "Method of Learning Neural Network for Elevator Call Assignment" and the like.

FIG. 3 shows an example of a neural network for call assignment. As shown in FIG. 3, the neural network NN for call assignment includes an input layer NR1 corresponding to an input pattern (elevator system state data), an output layer NR3 corresponding to an output pattern (assignment suitability),
It is composed of neurons of an intermediate layer NR2 located between the input layer and the output layer.

The input pattern uses various data representing the state of the elevator system (floor and direction of landing call, position and driving direction of each car, car call, load state, etc.) as parameters necessary for call assignment. , 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 parameters.

When input data is given to each neuron in the input layer, a signal is sequentially transmitted to the output layer, and as a result, a certain value is output from each neuron in the output layer. In the output layer, there are neurons for the number of elevators, and for various input patterns, the neuron corresponding to the car that is optimal for allocation outputs “1” (maximum value), and the other neurons output “0”. Is learned in advance.

Accordingly, among the values of each neuron in the output pattern, a neuron that outputs a value closest to “1” (maximum value) indicates that it is optimal for allocation.
The unit corresponding to this neuron is selected as the assigned unit. Note that the number of neurons in the intermediate layer (one layer in the embodiment, but may be plural) is appropriately determined according to the number of elevators, the properties of the building, and the like.

Although not shown, a connection weight (synapse weight) indicating the strength of the connection between the neurons is set between the neurons. The connection weight is initially set to an appropriate value, but can be modified so that a more accurate call assignment can be performed using a learning algorithm called “back propagation”.

Since this back propagation is well known and will not be described in detail, a learning sample (an input pattern and a desired output pattern corresponding to the input pattern, that is, a teacher signal are paired). An algorithm that compares the output pattern for the same input pattern with the teacher signal and corrects the connection weights to minimize the error. Initially, all weights are initialized (for example, random The input pattern of the learning sample is given to each neuron of the input layer. The output pattern at this time and the output pattern of the learning sample (teacher signal)
Then, using the difference (error), the values of the connection weights are sequentially corrected from the output layer side so that the difference becomes smaller.

If this is repeated using a large number of learning samples until the error converges, the call assignment function at the same level as the teacher signal is automatically embedded in the neural network, and the learning input The same level of call assignment as the teacher signal can be performed not only for patterns but also for unknown input patterns.

[0012]

As described above, a teacher signal is always required for learning using the back propagation method. However, in the case of elevator call assignment, an optimal teacher signal is obtained. Is very difficult. This is because when a new hall call occurs, it is relatively easy to find the optimal car at that time based on the car position of each car and the status of other calls, but in practice the assigned car is assigned to that car. Until the call is answered, there will be various unpredictable changes in traffic conditions, such as another new call or longer downtime on the middle floor, so that when the hall call occurs, It is very difficult to obtain a truly optimal assignment solution. In addition, the method of performing learning by back propagation has a problem in that call assignment with higher accuracy than a teacher signal cannot be performed.

For this reason, it is conceivable that self-organization of the neural network is attempted by using a so-called reinforcement learning method that does not require a teacher signal. Since this reinforcement learning method is already well known, a detailed description thereof will be omitted. First, all connection weights are initialized, and then one or a plurality of arbitrary connection weights are perturbed. That is, the value of the connection weight is slightly changed. 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 if the result is improved, the perturbation is accepted. To the state before giving. When this procedure is repeatedly performed in this manner, each connection weight gradually converges toward the optimum value. This is the reinforcement learning method.

Therefore, when this reinforcement learning method is applied to a neural network for assigning calls to elevators, the procedure is generally as shown in FIG.
1 gives (+) perturbation to the neural network for assignment. That is, the value of one or a plurality of arbitrary connection weights is slightly increased. Then, the elevator is operated for a certain period of time while performing the assignment by the neural network after the (+) perturbation, and the waiting time of each call is totaled from the assignment result during that time (step S2).

Next, in step S3, a (-) perturbation is applied to the initial neural network, that is, the value of the connection weight is slightly reduced this time, and the neural network after the (-) perturbation is performed in the same manner as described above. The operation is performed for a fixed time while the assignment is performed, and the waiting time of each call is totaled from the assignment result during the operation (step S4).

In the case of (+) perturbation and in the case of (-) perturbation, for example, an average waiting time is calculated from the above totaled result, and the allocation performance is compared (step S5). The connection weight is selected and updated to that value (step S6). Thereafter, when the updated neural network is used as a new neural network for allocation and the above procedure is repeated, the values of the connection weights eventually converge to the optimum values.

In the above procedure, in order to accurately compare the allocation performance in the case of (+) perturbation and (-) perturbation in step S5, steps S2 and S4 are required.
Must be performed under the same conditions, that is, under the same call occurrence conditions. For this purpose, learning can be performed using a simulation device.However, it is extremely difficult to accurately grasp in advance the properties and traffic conditions of the building in which the elevator is actually installed. Since the traffic situation fluctuates even after installation, it is necessary to continue reinforcement learning while actually driving in the building after installing the elevator.

However, when reinforcement learning is performed while actually driving at the site, it is very difficult to make the operating conditions in steps S2 and S4 identical as described above. This means that it is necessary to perform the operation for about 30 minutes each in order to obtain statistics such as the waiting time and the like. For example, even during the operation at the step S2 during the UP peak time, at the time of the step S4, This is because traffic conditions may change during normal times, and even in the same UP peak time zone, the occurrence of calls may differ significantly between the first half hour and the second half hour. .

For this reason, in order to perform reinforcement learning while actually driving at the site according to the procedure of FIG. 4, for example, if the tally in the case of (+) perturbation is obtained in step S2, (-) in step S4 Aggregation by perturbation requires the same conditions as possible, such as the same time period of the next day, or, for more accuracy, the same time period of the same day of the next week. Executing a cycle will take one day or one week,
Not only does it take a very long time to complete learning, but also in buildings where traffic conditions frequently fluctuate, there is a problem that neural network learning cannot follow such fluctuations.

The present invention has been made in view of such problems, and has as its object to provide a method capable of performing reinforcement learning in a short period of time while actually driving at a site.

[0021]

In order to achieve the above object, according to the present invention, a plurality of nets having different perturbation amounts are created by perturbing the connection weights of the neural network for assignment. For example, a (+) net perturbed to the (+) side and a (−) net perturbed to the (−) side are respectively created, and the short-time operation by the assignment in the (+) net and the (−) net are performed. After repeating the operation of the elevator for a predetermined cycle with the short-time operation based on the assignment by the neural network as one cycle, the assignment performance by the (+) net and the assignment performance by the (-) net during that period are compared. And giving the perturbation with the better performance to the connection weight of the neural network for assignment to update
Thereafter, the above procedure is repeated to correct the connection weight of the neural network for assignment.

[0022]

According to the present invention, the allocation performance is compared and the connection weight is not corrected for each cycle of the short-time operation by the (+) net and the short-time operation by the (-) net, but the link weight is corrected. After the operation is repeated, the allocation performance of each neural network is compared with the statistical result of the predetermined cycle, and the connection weight is updated according to the result.

[0023]

An embodiment of the present invention will be described below.
FIG. 1 is a flowchart showing a learning procedure of the present invention, in which reinforcement learning is performed while an elevator is actually operated by an assignment neural network.

First, in step S1, the initial net is copied and used as a neural net for assignment. As this initial net, a neural net that can be actually assigned to some extent as it is, such as one already learned by simulation or the like, or one already used for assignment at another site, is used. .

In step S2, the perturbation amount of the i-th connection weight of the output layer of the neural network for allocation is set. In steps S3 and S4, the (+) net given the (+) side perturbation and ( Create (-) nets with perturbation on the-) side. Here, the amount of perturbation of the weight can be set to an appropriate value. For example, an actual assignment is performed by using the neural network for assignment, and an appropriate amount of perturbation (the assignment result is affected, and the Is within the allowable range) by calculation.

Next, in step S5, the elevator is assigned on the (+) net until the fixed number of hall calls are assigned and serviced (step S7). In step S6, the actual waiting time of each hall call is recorded. And (+)
When the number of hall calls assigned and serviced by the net reaches a certain number, hall calls are allocated in the same way until a certain number is reached by the (-) net, and the waiting time of each hall call is recorded during that time ( Steps S8 to S10).

Here, the above-mentioned fixed number of hall calls is, for example, about five, and one period in which the assignment is made by the (+) net or the (-) net is short enough that the traffic situation does not change significantly during that period. The time is set to be, for example, 1 to 2 minutes (at most, about 5 minutes). Then, one cycle includes the short-time operation for one period by the (+) net and the short-time operation for one period by the (-) net. FIG. 2 shows this state.

In FIG. 2, each arrow has a starting point indicating the point of occurrence and floor at which the hall call occurred, an end point indicating the point at which the car actually served the call, and the length of the arrow indicates the actual waiting time of the call. Each represents time. As shown in FIG. 2, when the number of hall calls serviced by the (+) net allocation becomes a fixed number (four in this example, 1 to 4),
Next, the assignment is made by the (-) net, and this is repeated as one cycle. At this time, a call that extends over both the period of assignment by the (+) net and the period of assignment by the (−) net, such as 5, is not subject to evaluation because both are affected.

Returning to FIG. 1, steps S5 to S10 are repeated via step S11, that is, the short-time operation in which the (+) net allocation period and the (-) net allocation period are one cycle is performed for a predetermined number of cycles (10 (About 20 times), and when the predetermined number of cycles is reached, the process proceeds from step S11 to S12. In step S12, the waiting time of each hall call in the operation period of the predetermined cycle is divided into the hall call assigned by the (+) net and the hall call assigned by the (-) net, and totaled. Calculate the average waiting time at. Then, in step S13, the average waiting time, that is, the allocation performance is compared.

Thus, instead of comparing each cycle,
Comparing by counting the number of predetermined cycles, even if there is a slight change in the traffic situation in each cycle, the period of (+) net and the period of (-) net are almost the same time zone / traffic as a whole. In this situation, the allocation performance of the two can be accurately compared under almost the same conditions.

As a result of comparison, if the performance of the (+) net is better, the i-th connection weight of the output layer of the neural network for assignment is perturbed by (+), and vice versa. And update (step S14). In this way, when the above procedure is further repeated, the connection weight gradually converges, and a neural network most suitable for the traffic condition of the building is constructed.

In the above embodiment, only two (+) nets and (-) nets are created. However, the number is not limited to two, and a plurality of nets having different perturbation amounts are created. Similarly, each of them is operated for a short time, and the operation is performed for a predetermined number of cycles as one cycle. Then, the assignment performance of each net is evaluated, and the best perturbation of the net is assigned to the assignment neural net. It can also be provided and updated.

Further, in the above embodiment, the evaluation of the allocation performance is performed by the average value using the actual waiting time of the hall call, but the present invention is not limited to this. It may be performed by using other indices such as a certain long wait probability or power consumption, or may be performed by comprehensive evaluation thereof.

In the above embodiment, the correction of the weight is performed only on the output layer. However, the present invention can be similarly applied to the correction of the weight of the input layer and the intermediate layer.

Further, in the above embodiment, the period of time of the allocation by the (+) or (-) net is a period until a certain number of hall calls are serviced, but may be a predetermined time. The predetermined time is set to be longer than the maximum expected waiting time (to take a sample with a long waiting time), but it is preferable to set the maximum time to about 5 minutes as described above.

[0036]

According to the present invention, the learning by the reinforcement learning method can be performed in a short time while actually driving at the site. In addition, even when the traffic condition changes after the elevator is installed, it is possible to quickly follow the change and constantly maintain highly accurate assignment performance.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a learning procedure according to the present invention.

FIG. 2 is a diagram for explaining a relationship between allocation periods of each net according to the present invention.

FIG. 3 is a diagram showing an example of a neural network for call assignment.

FIG. 4 is a flowchart showing a conventional learning procedure.

[Explanation of symbols]

 NN Neural network for call assignment NR1 Input layer neurons NR2 Hidden layer neurons NR3 Output layer neurons 1 to 5

Claims (4)

(57) [Claims] When a new hall call is generated, a system state data of the group at that time is input, and an elevator call allocation neural network learning method used for selecting and allocating an optimum car from its output is assigned. A perturbation is applied to the connection weights of the neural network for use to create a plurality of nets having different amounts of perturbation, and the short-time operation by the assignment in each of the plurality of nets is combined into one cycle, and the elevator having a predetermined number of cycles is used. After the operation is repeated, the allocation performance of the nets during the operation is compared, and the perturbation of the net having the best performance is given to the connection weight of the neural network for allocation to be updated, and thereafter, the above procedure is repeated. The connection weight of the neural network for assignment is modified by Learning method of neural network for call assignment. 2. A method according to claim 1, wherein when a new hall call is generated, the system state data of the group at that time is input, and an elevator car allocation neural network learning method is used for selecting and allocating an optimum car from the output. A (+) net in which the connection weight of the neural network for use is perturbed to the (+) side and a (-) net in which the connection weight is perturbed to the (−) side are respectively created, and the short time by the assignment in the (+) net is created. After repeating the operation of the elevator for a predetermined cycle with the operation and the short-time operation by the assignment in the (-) net as one cycle, the assignment performance by the (+) net and the assignment by the (-) net in the meantime. Compared with the performance, the perturbation with the better performance is given to the connection weight of the neural network for assignment and updated, and thereafter, A method of learning a neural network for allocating an elevator call, wherein a connection weight of the neural network for allocating is corrected by repeating a procedure. 3. The learning method for an elevator call assignment neural network according to claim 1, wherein the assignment performance is evaluated based on an actual waiting time of the hall call. 4. The method according to claim 1, wherein the assignment performance is evaluated by comprehensive evaluation of a plurality of indices.
The learning method of the neural network for elevator call assignment described in (1).