JP2867849B2 - 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 for assigning calls using a neural network.
It relates to the learning method of the neural network.

[0002]

2. Description of the Related Art Conventionally, group management control of elevators is mainly based on a call assignment method using an evaluation function, but recently, an optimization problem is approximated by a simple algorithm based on a biological neural network model. The idea of "neural net" that performs analysis and pattern recognition is attracting attention,
By using this neural network, a new elevator group management and control device, which is completely different from the conventional one, is proposed, which can automatically learn and generate a judgment system for determining the optimal assigned car based on actual operation examples. Have been.

A neural network is a network that imitates the human brain. A plurality of neural cell models (neurons) are connected in a complicated manner, and the operation of each neuron and the connection form between the neurons are properly determined to perform pattern recognition. It is possible to embed functions and knowledge processing functions. For example, "Nikkei Electronics" August 1, 1987
P115-P124 of the 0th issue (No. 427) and 1989
"PDP" published by Sangyo Tosho in February, 2002
And the like, and in particular, those having a structure in which neurons are arranged in a hierarchical structure are characterized in that an autonomous learning algorithm called "back propagation" can be used.

[0004] When a 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.
No. 275381, JP-A-2-52875, and JP-A-Hei-3-
No. 59971 and the like.

FIG. 3 shows a typical example of a neural network for call assignment. As shown in FIG. 3, the neural network NN for call assignment is an input signal (system state data).
Neuron NR ₁ corresponding input layer, assignment Unit neuron corresponding output layer NR _3, composed of neuron NR ₂ intermediate layer placed in the middle of the input layer and the output layer.

[0006] The input signals include 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 load in the car, and the number of waiting passengers at the hall. This is a group of system state data 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 car, that is, the same number as the number of elevators, and the car corresponding to the neuron whose output signal (assignment suitability) has the largest value is set as the optimum car, and A call is assigned. The number of neurons in the intermediate layer (the number of neurons in the embodiment is one, 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 weights are initially initialized, but can then be modified using a learning algorithm called "back propagation" to allow more accurate call assignment.

In the back propagation, an output signal of a neural network is compared with a previously generated teacher signal for the same input signal, and a connection weight is corrected so as to minimize the error. In the algorithm, first,
First, all connection weights are initialized (for example, set to random values), and input data for learning (same as input data of a 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 value of each connection weight is corrected in order from the output layer side using the difference (error) so as to reduce the difference. 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. Can be assigned at the same level as the teacher signal for the input data (the new hall call and the system state data at that time).

[0009]

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 assigned 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 not Until you answer the hall call, there will be various unpredictable changes in traffic conditions, such as a new call occurring or a longer stoppage time on the middle floor, so it is truly optimal when a hall call occurs. It is impossible to get an assignment solution.

For this reason, as shown in, for example, Japanese Patent Application Laid-Open No. 2-52875, an optimal operation is analyzed for all hall calls generated during one day, and a neural network Methods for learning have been proposed. This method attempts to capture the problem of finding the optimal solution from among all combinations of assignments for all hall calls during a day as a combination optimization problem well known in the traveling salesman problem and the like. is there.

[0011] However, in order to solve this combinatorial optimization problem, for example, in addition to the simulated annealing method,
Although various ones are known, all of them are to obtain approximate solutions and not to obtain a true optimal solution, and furthermore, the operation is very complicated and the operation takes time. It is very difficult to create a learning sample that serves as a teacher signal.

Further, since such a method requires a deterministic optimal solution based on complete information including the future, it is a probabilistic optimal solution whose future development is unknown at the time of hall call assignment. Are different. Therefore, a neural net made with such learning samples is not always optimal.

[0013]

According to the present invention, in order to solve the above-mentioned problems, each time a new hall call is generated, the system state data of the group is stored and the unanswered call at that time is re-targeted. When an assignment is made and the car responds to the floor where the hall call is generated and stops according to the allocation result, a pair of the answering machine at that time and the system status data of the group at the time of occurrence of the hall call is paired. Is registered as a learning sample, and thereafter, this is repeated to create a large number of learning samples, and the neural network is trained using the large number of learning samples as a teacher signal.

[0014]

According to the present invention, since the learning sample is paired with the system state data of the group at the time of the occurrence of the hall call, not the assigned car at the time but the actual responding car after the reassignment. When the neural network is trained using the training sample of the neural network as a teacher signal, the function of determining the optimal unit, including the function of predicting the situation change from the time when a hall call occurs to the time when the call is actually answered, is included in the neural network. It will be automatically formed.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS One embodiment of the present invention will be described below with reference to FIGS. In the present invention, while actually operating the elevator at the target site (building) or realizing the same operation as the site by simulation, a learning sample is created by the following procedure, and the learning sample is created. Is used as a teacher signal.

FIG. 1 is a flowchart showing a procedure of a learning method according to the present invention, and FIG. 2 is a diagram showing a relationship between a car position and a call occurrence state for explaining the present invention. For example, FIG. Then, Unit A is traveling in the ascending direction near the second floor, Unit B is traveling in the ascending direction near the first floor, and Unit C is 8
The vehicle is traveling in the descending direction near the floor, and the mark ○ indicates a car call,
The △ marks indicate landing calls, respectively. FIG. 2 (b),
The same applies to (c).

First, when a new hall call is generated, the system state data of the group at the time of generation of the call is stored (steps S1 and S2). For example, as shown in FIG. 2 (a), when the hall call H ₁ now 8F upward to the newly generated, the system state data, that A~C Unit cage position and driving direction of the group at that time , Load state, car call occurrence floor, landing call occurrence floor and direction, and the like, and data indicating the state of the group considered necessary for assignment are stored.

Next, in step S3, reallocation is performed for all unanswered hall calls including the new hall call.
Usually, the assignment is often fixed to avoid confusion due to the assignment change, but here, in order to create an optimal learning sample, every time the situation changes due to the occurrence of a hall call, etc.
Reassignment is performed for all unanswered hall calls.

[0019] However, since there is only unanswered hall call one of an H ₁ at the time here in FIG. 2 for simplicity of explanation (a), and allocates this H _1. It is not necessary to use a neural network for the assignment in step S3, and an assignment using a conventional evaluation function or an assignment by a fully automatic group multiplication method is sufficient. That is, the estimated waiting time for the call and the like are calculated for each car, and the car considered to be optimal at this time is selected.

FIG. 2B shows a state in which the hall call H ₁ is assigned to the car A as a result, and after a while, the hall call H _{2 in the} descending direction of the seventh floor is newly generated. Next, in step S4, it is determined whether or not the vehicle has stopped in response to the hall call. However, during the period from (a) to (b) in FIG. 2, the process returns to step S1, and steps S1 and S4 are performed. repeat.

[0021] Then, the hall call again step S1~S3 when H ₂ is generated in shown in FIG. 2 (b) is repeated, stores a group of system state data at this point,
Reassign all unanswered hall calls. FIG.
At the time of (b), since hall call H ₁ and H ₂ are both unanswered reallocate both as a target is performed, resulting hall call H ₂ is assigned to car A, hall calls H ₁
Is assumed to have been reassigned to Unit B. This is shown in FIG. 2 (c).

Thereafter, assuming that a new landing call does not occur and the car A stops in response to the hall call H ₂ and the car B stops in response to the hall call H ₁ , first, the car A responds to the hall call H ₂ . When the call stops, the system status data of the group stored at the time of occurrence of the call, that is, at the time shown in FIG. 2 (b), and the car A which is the responding car are paired and registered as a set of learning samples ( Step S4, S
5).

[0023] Similarly, even when the car B is stopped in response to the hall call H _1, the response No. machine and system status data group stored at the time shown in the call generation time i.e. in FIGS. 2 (a) Is registered as a set of learning samples by pairing with No. B (steps S4 and S5).

Similarly, every time a hall call occurs, reassignment is performed for all unanswered calls, and each time the car stops responding to the hall call, the answering machine and the time of occurrence of the call The learning sample paired with the system state data of the group in is registered. Then, when the number of samples reaches a predetermined number, learning of a target call assignment neural net is performed by the above-described back propagation using the large number of learning samples as teacher signals (step S).
6, S7).

As described above, according to the present invention, the learning sample is registered as a pair with the system state data of the group, not as the assigned unit at the time of the call occurrence, but as a result of the reassignment for each subsequent situation change. because it actually responds No. machine in response to the call, for example, looking at the hall calls H ₁ of FIG. 2, and if the subsequent changes in circumstances (generation of nominal H _2, etc.) allocated to predict at the time point of generation of call The same result will be registered as a learning sample.

Accordingly, when learning of the neural network is completed using a large number of the learning samples and the neural network is applied to the target building, when a new hall call occurs, the system state of the group at that time is input. At that time, the most appropriate car can be assigned taking into account the occurrence of other calls and changes in the situation.

In the above embodiment, the unassigned hall call is reassigned when a new hall call occurs, but the hall call is also reassigned when a car call occurs. Alternatively, the reallocation may be performed at regular intervals.

In the above-described embodiment, the reassignment is performed for all unanswered calls. However, when the present invention is applied to an actual site to create a learning sample, the assignment is changed by changing the assignment. In order to reduce confusion, reallocation may be performed only for unanswered calls that satisfy a predetermined condition (for example, a hall call with a waiting time equal to or longer than a predetermined value).

As described above, the reassignment in step S3 does not necessarily need to use a neural network.
Therefore, a learning sample may be created using the conventional group management device, and then learning of the neural network for call assignment may be performed. However, if the neural network for call assignment is used from the beginning, step S3 is executed. In this case, the neural network is initialized in advance so that the minimum assignment can be performed, a learning sample is created using the neural network, and the neural network is improved by learning thereafter.

[0030]

According to the present invention, the learning samples can be created very simply because they need only be assigned by the conventional evaluation function or the like. Moreover, since this learning sample pairs not the assigned unit at the time of the call occurrence but the actual response unit after reassignment and the system state data of the group at the time of the call occurrence, the learning sample is When learning as a teacher signal, the optimal machine judgment function including the prediction function of the situation change from the time of the occurrence of the hall call until actually responding to the call is automatically formed in the neural network, High-precision assignment can be performed.

[Brief description of the drawings]

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

FIG. 2 is a diagram illustrating a relationship between a car position and a call generation state for explaining the present invention.

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

[Explanation of symbols]

S1~S7 steps H ₁ of the flow chart, H ₂ floor call NN call neuron NR ₃ output layer neuron allocation for the neural network NR ₁ input layer neuron NR ₂ intermediate layer

Claims (2)

(57) [Claims] When a new hall call is generated, a method for learning a neural network for elevator call allocation for inputting system state data of a group at that time and selecting and allocating an optimum car from an output thereof is provided. Each time a hall call occurs, the system status data of the group is stored and reassigned for all unanswered calls at that time, and the car stops responding to the floor where the hall call occurs. A pair of the answering machine at that time and the system state data of the group at the time of occurrence of the hall call are registered as a set of learning samples, and thereafter, this is repeated to create a large number of learning samples, and A learning method of a neural network for elevator call assignment, wherein learning of the neural network is performed using a sample as a teacher signal. 2. A method for learning a neural network for elevator call assignment for inputting system status data of a group at that time and selecting and assigning an optimum car from its output when a new hall call is generated. Each time a hall call occurs, the system status data of the group is stored and, at regular intervals, all unanswered calls at that time are reassigned to the floor where the car is making the hall call. When the vehicle stops responding, the response machine at that time and the system status data of the group at the time of the occurrence of the hall call are registered as a pair and registered as a set of learning samples. A neural network for elevator call assignment, wherein learning of the neural network is performed using the large number of learning samples as teacher signals. Learning method.