JPH05165839A - Scheduling system and condition adjustment neural network - Google Patents

Abstract

PURPOSE:To attain a high speed manufacture scheduling system equipped with a knowledge obtaining function by using the optimizing function and learning function of a neural network. CONSTITUTION:After the setting of a condition necessary for a scheduling is operated by a constraint condition setting device 1, the resetting of a variable constraint condition is operated by a variable constraint condition adjusting neural network 2. The variable constraint condition adjusting neural network 2 obtains the knowledge of the adjustment of the variable constraint condition by the learning function, and it is not necessary for an operator to operate the adjustment. The weight of the neural network used by a scheduling neural network 4 is set by the set constraint condition by a scheduling weight setting device 3. The scheduling neural network 4 efficiently operates the scheduling by using the optimizing function of the neural network. A display 5 operates the display and change of the result of the scheduling.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a manufacturing scheduling for automatically obtaining an optimum manufacturing schedule in a manufacturing line and a neural network used therefor.

[0002]

2. Description of the Related Art An example of the flow of a conventional manufacturing scheduling system is shown in FIG. In the system shown,
In the constraint condition setting device 1, a plurality of production conditions with different values of the variable constraint condition are prepared. Then, the scheduling device 6 schedules each of the prepared production conditions. An electronic computer is used as the scheduling device. In general, if the variable constraints are adjusted to reduce the production cost too much, the production capacity will be reduced, the production will not be in time, and the scheduling will fail. If the production cost is raised too much, the production capacity will be excessive. The scheduling staff evaluates the scheduling 8 using all the obtained scheduling results 7, and selects the one with the minimum production cost and the successful scheduling.

[0003]

However, in the conventional scheduling system as described above, the execution time of the electronic computer greatly increases as the number of products or processes to be scheduled increases. Furthermore, it takes a very long time to obtain a solution because scheduling is performed for all prepared cases. In addition, the operator must consider the adjustment of the variable constraint condition for assigning the variable constraint condition corresponding to the situation of excess or deficiency of the production capacity, and the knowledge of the expert is required.

The present invention has been made to solve the above problems, and an object of the present invention is to obtain a scheduling system capable of high-speed scheduling without operator intervention and a neural network used in this system or the like. ..

[0005]

In order to solve the above-mentioned problems, the scheduling system of the first invention is a neural network of variable constraint condition adjustment having a learning function,
It has a scheduling neural network with an optimization function and has the following elements. (A) A constraint condition setting device that sets a constraint condition for scheduling, and (b) a variable constraint condition whose constraint is changeable among the constraint conditions set by the constraint condition setting device described above is input and adjusted. Variable constraint adjustment neural network, (c) Scheduling in which the constraint conditions set by the constraint condition setting device and the variable constraint conditions adjusted by the variable constraint neural network are input and weights are set for the respective conditions. Weight setting device, (d)
A scheduling neural network that performs scheduling based on the weight set by the scheduling weight setting device.

The condition adjusting neural network of the second invention is used, for example, in a variable constraint condition adjusting neural network for adjusting variable constraint conditions of a scheduling system. Instead, it is equipped with a neural network that divides the whole into windows of a prescribed range and adjusts it, and gains knowledge about variable constraint adjustment for each window, automatically adjusts variable constraint conditions, and speeds up manufacturing scheduling. And has the following elements. (A) condition input means for inputting the condition to be adjusted as time series data, (b) index input means for inputting the index for adjusting the condition to be adjusted as time series data,
(C) Using time-series data within a predetermined range of the time-series data input by the condition input means and the index input means,
A neural network that finds and outputs the optimum condition at a predetermined time point within the range, and moves and moves the predetermined range in the time series direction to find and output the optimum condition at each time point.

[0007]

In the scheduling system according to the first aspect of the present invention, the variable constraint adjusting neural network is provided with a learning function so that a desired variable constraint can be learned from the schedule result that changes according to the variable constraint and the current variable constraint. It is possible to adjust variable constraints and acquire knowledge. Further, the scheduling neural network can perform scheduling at high speed by having a neural network having an optimization function.

In the learning of the condition-adjusting neural network according to the present invention, the variable constraint condition can be adjusted without being affected by the deviation on the time axis by learning while the input and output of the neural network are translated in parallel on the time axis. .. For example, in a manufacturing scheduling system, the correlation between the excess and deficiency of production capacity and the required production capacity does not change much regardless of the pattern in the time series. However, it is extremely difficult to make a neural network learn the invariance of such a time relation. Therefore, by cutting out a period in a predetermined range from the schedule target period with a window and sliding this window on the time axis within the schedule target period, it is possible to repeatedly execute learning and condition adjustment in a limited range. It was done.

[0009]

【Example】

Example 1. FIG. 1 shows an embodiment of a scheduling system according to the present invention. The system shown in FIG. 1 includes a constraint condition setting device 1, a variable constraint condition adjusting neural network 2, a scheduling weight setting device 3, a scheduling neural network 4, and a display 5.

First, the general operation of the system shown in the figure will be described. The person in charge of scheduling sets conditions necessary for manufacturing scheduling via the constraint condition setting device 1. Variable constraint conditions that need to be adjusted (for example, the amount of personnel input, the amount of material input, etc.) are sent to the variable constraint adjustment neural network 2, and other conditions are fixed conditions (for example, holidays and product manufacturing). Available lines, delivery dates, production quantities, etc.) are sent to the scheduling weight setting device 3. The variable constraint condition adjustment neural network 2 sets a new variable constraint condition from knowledge learned in advance, with the current variable constraint condition and the excess / deficiency of the production capacity as inputs. The newly set variable constraint condition is sent to the scheduling weight setting device 3. The scheduling weight setting device 3 receives the fixed condition and the variable constraint condition as inputs, and the synapse weight (weight) of the neural network used in the scheduling neural network 4.
To set. The set weight is sent to the scheduling neural network 4. The scheduling neural network 4 schedules the manufacturing process using the set weight. The excess / deficiency situation of the variable constraint condition in the scheduling result is fed back to the input of the variable constraint condition adjustment neural network 2. The scheduling result is displayed on the display 5. The person in charge of scheduling changes the schedule from the display 5 as necessary, and the internal state of the scheduling neural network is changed to reflect the result.

Next, FIG. 2 shows an example of the manufacturing process targeted by this embodiment. The product manufacturing process consists of A, B, C, D and E. In the figure, 9, 10, 11, 12, and 13 are equipments in charge of processes A, B, D, D, and E, respectively. 14,
Reference numerals 15 and 16 are flows each including all processes and are called lines. The products are manufactured in one line. For each product, the process is carried out in the order A, B, C, D, E so that the product that started later in the line does not overtake the product that started earlier. The scheduling system determines the start date of each process of each product by the method described below.

Input / output of the variable constraint condition adjusting neural network 2 will be described with reference to FIG. In the figure, 2 is a variable constraint condition adjusting neural network, 21 is an input, and 22 is an output. 23, 24 and 25 are graphs of time series data given to the input, and 23 is a time axis, 2
4 is the time-series data of the current variable constraint conditions to be input, 2
5 is time series data of the excess and deficiency of the current production capacity, which is an index for adjusting the variable constraint condition. 26 is time 2
7 shows a range of time (hereinafter referred to as a window) given in the variable constraint condition adjustment neural network 2 to obtain a desirable variable constraint condition, and 27 is the time (day) at the center of the window. The figure schematically shows that the time-series data included in the window 26 is input to the variable constraint condition adjustment neural network 2 in order to obtain a desired variable constraint condition at a certain time (day) 27. That is, only the data contained in the window 26 of the time series 24 and 25 is selected and given to the variable constraint condition adjustment neural network 2, and based on the input, the variable constraint condition adjustment neural network 2 determines the time (day). The desired variable constraint in 27 is output to 22.

Next, the variable constraint condition shown by 24 in FIG. 3 will be described. The daily production capacity of each process in the manufacturing line is not fixed and can be enhanced for a certain period by means such as manpower input. Therefore,
During the period when the quantity of orders is large and the production is not in time by the delivery date,
Increase production capacity so that deadlines can be met. The production capacity is determined for each process and index as shown in FIG. 4. For example, the production capacity of the process A in the production capacity index 1 period is 100 [t / day], and the production capacity index 2 is 12 periods.
It is set as 0 [t / day]. When this system is movable, the operator must set the constraint condition setting device 1
Initialize using. Regarding the variable constraint condition 24, for example, time-series data such as the variable constraint condition 24 shown in FIG. 3 is given for the process A. Specifically, this time-series data is scheduled on a date and on that date. It is given by the production capacity index of the completed process A.
In this way, the index is selected as the variable constraint condition 24, and the production capacity before adjustment is set. The variable constraint adjustment neural network 2 is shown in FIG. 5 for each process of each line.
The variable constraint adjustment neural network 2 as shown in FIG. The variable constraint condition adjustment network 2 for each process of each line obtains the optimum production capacity for each process of each line according to the production capacity before adjustment and the excess / deficiency production capacity 25, and the scheduling system is the closest production capacity to this. Select the index and output. For example,
If the index 1 is input as the variable constraint condition 24 before adjustment of a certain date for the process A on the line 14,
The variable constraint adjustment neural network 2 for the process A on the line 14 outputs, for example, the index 2 as the optimum production capacity on that date. The production capacity at the input / output of the variable constraint adjustment neural network is normalized so as to be 1 at the maximum production capacity of the production line.

Next, the excess / deficiency production capacity 25 shown by 25 in FIG. 3 will be described. Each product has a set production start date and delivery date, and all processes must be completed within the period from the production start date to the delivery date. Undercapacity is calculated as follows. First, when a plurality of products are assigned to the same process on the same day as a result of scheduling, the production capacity shortage on that day is (the number of overlapping products-1) x the production capacity on that day. FIG. 6 shows how to determine the production capacity shortage when a certain product cannot be scheduled within the delivery date from the production start date. Reference numeral 71 denotes the scheduling result 28 output by the scheduling system. The left end of the rectangle 77 corresponding to each process shows the start date of the process, and the right end shows the end date of the process. This scheduling result 71 shows the case where all the processes cannot be completed, and the production start date 74 to the delivery date 76
If the compression is performed without changing the production start date 75 so as to fall within the range of, the scheduling result of 72 is obtained. For example, in process A in the figure, the compression rule is a1 / a2, the compression ratio of each process is calculated, and (compression ratio-1) x production capacity is insufficient for each day of the period to which each process is allocated on the compressed schedule. Production capacity. The time series of the insufficient production capacity due to this product in the process A is as 73. On the other hand, if there is no shortage of production capacity on a certain process, production is not allocated, and there is a day that is not a line cleaning period between productions, the production capacity on that day is defined as overproduction capacity. Overcapacity is negative and undercapacity is positive, normalized on the same scale as unadjusted capacity.

Next, returning to FIG. 3, the role of the window 26 will be described. The scheduling system slides the window 26 from the production start date to the production end date, and the time (day) 27 in the middle of the window 26.
To determine the optimum production capacity of. For example, if there are 30 days from the production start date to the production end date and the size of the window 26 is 9 days, the window 26 is shifted by 30 days (30 times). Then, using the variable constraint condition 24 and the excess / deficient production capacity 25 for the preceding and following 9 days, the time (day) 2
7 to find the optimum production capacity. The variable constraint condition adjusting neural network 2 is prepared for each process of each line, and this work is performed for each process of each line as described above. Thus, using the window 26,
Variable constraint condition adjustment neural network 2 has 9
The optimum production capacity is output using the variable constraint condition 24 and the excess / deficiency production capacity 25 for each day, and the optimum production capacity is obtained using the variable constraint condition 24 and the excess / deficiency production capacity 25 for 30 days of all processes. The result can be output at higher speed than that. Further, the variable constraint condition adjusting neural network 2 is configured by a neural network as shown in FIG. 5 and has a learning function. Therefore, the learning function works more efficiently as the same pattern or similar patterns occur. It will be. Therefore, rather than learning the patterns of the variable constraint condition 24 and excess / deficient production capacity 25 for 30 days and outputting the optimum answer to the new variable constraint condition 24 and excess / deficient production capacity 25 for 30 days, It is more efficient for the learning function to learn the patterns of variable constraint conditions 24 and excess / deficient production capacity 25 for a day and output the optimum answer to the new variable constraint conditions 24 and excess / deficient production capacity 25 for 9 days. Will work.

As described above, the window 26 reduces the amount of data to increase the execution speed of the variable constraint condition adjustment neural network 2, and increases the appearance rate of the same or similar patterns to increase the variable constraint condition adjustment neural network. It is provided to enhance the learning function of the network 2. Therefore, the window 26
A smaller size makes it possible to achieve higher speed and higher learning capability, but on the other hand, there is a risk that scheduling will be missed in the entire process. This is because the variable constraint adjustment neural network 2 uses the window 26
The optimum answer is output within the size (for example, 9 days), and when the size of the window 26 is equal to the entire process (for example, 30 days), the variable constraint adjustment neural network 2 is the entire process (30 days). This is because the optimum production capacity for the day in the center of the window 26 is output in the case of passing through.

Next, the input / output and learning method of the variable constraint condition adjusting neural network 2 will be described. The variable constraint adjustment neural network 2 can learn the capability adjustment as follows after the scheduling is completed. First, during the adjustment, the input of the constraint condition adjustment neural network 2 is stored at some time points. After the scheduling is completed, a scheduling expert adjusts the scheduling result 28 which seems to be better. The result is used as a desired output, and the stored input is used for learning as described below, whereby the capability of the variable constraint condition adjusting neural network 2 is improved.

The steps of inputting / outputting and learning the production capacity of the variable constraint condition adjusting neural network shown in FIGS. 3 and 5 will be specifically described. Learning has been scheduled,
Alternatively, after being interrupted by the operator, the expert presents the desired production capacity, which is used as teacher data. Prepare the time series data of production capacity before adjustment (the production capacity for each day is lined up) in array p [i] (i = 1, 2, ... N, N is the number of days to be scheduled) The time-series data of production capacity is arrayed as q [i] (i = 1, 2, ... N, N
Prepared for the number of days to be scheduled. The variable constraint adjustment neural network 2 shifts the window 26 of FIG. 3 by one day, and the input neuron sequence r [i] (i
= 1, 2, ... 2L, L is the number of days included in the window and is an odd number. ), And learns the element s [i] of the time series s [i] (i = 1, 2, ... N) of the production capacity adjusted by the expert as a teacher signal. The learning procedure is described below.

1. Let i = 1. 2. The production capacity 24 before adjustment is set in the input neuron. r [j] = p [i- (L-1) / 2 + j-1] (j = 1, 2, ... L) where i- (L-1) / 2 + j-1 is less than 1 or N If so, set the default (usually lowest) capacity to r [j]. 3. Insufficient production capacity 25 is set in the input neuron. r [j + L] = q [i- (L-1) / 2 + j-1] (j = 1, 2, ... L) where i- (L-1) / 2 + j-1 is less than 1 or N If it is larger, 0 is set to r [j + L]. 4. r [i] (i = 1, 2, ... 2L) is input,
A variable capacity adjustment neural network is trained using s [i] as a teacher signal. Learning is performed by back propagation and the error obtained in this process is recorded. 5. If i is smaller than N, add 1 to i and perform step 2
Return to. 6. If the average value of all the errors obtained in the process of step 4 is sufficiently small, the process ends. Otherwise, go back to step 1.

Next, the operation of the scheduling neural network 4 will be described with reference to FIG. This network determines the allocation of products to the system, the order, and the start date of each process of each product so as to satisfy the scheduling constraint. There are many constraint conditions, but here we use the constraint condition "The manufacturing period of a product on a certain process must precede the manufacturing period of the next product to be produced on the same process" as an example. The operation will be described.

FIG. 8 is a diagram functionally classifying some of the neurons of the scheduling neural network 4. In the figure, each symbol has the following meaning.・ M: Total number of products. Α _i ^ps : the start date when the process p of the product i is broken by the line s. Expressed as the number of days from the first day of the schedule period. B _ij ^s : 1 if the product i is to be manufactured jth in the line s, 0 otherwise.

The scheduling neural network 4 comprises a neuron having an output value of a and a neuron having an output value of b. the output value of a is from internal state Sa of the neuron, a = calculated by N (1 / (1 + exp (-S a))), the output value of b is the internal state S of neurons
It is calculated from b by b = 1 / (1 + exp (−S _b )).

FIG. 9 is a diagram showing an example of the case where the product i and the product j are scheduled in the same process on the same line. FIG. 9A shows the case where the scheduling is successful.
(B) shows the case of failure.

As described above, the success or failure of the scheduling can be represented by the penalty function φ (x) as shown in FIG. 10, for example. This function φ (x) is used for the product i in the process p of the line S and the product j in the same process of the same line next
When manufacturing and, it is a function showing the degree of overlapping of schedules, and φ (x) = 0 when there is no overlapping, and the larger the overlapping, the larger the positive value shown in FIG.

The constraint condition "the manufacturing period of the product on a certain process must be earlier than the manufacturing period of the next product to be produced on the same process" is shown in FIG.
It is satisfied by minimizing F in Equation 1 of (a).

Next, the operation of the scheduling neural network 4 will be described. First, the internal state Sa of the neuron in the scheduling neural network 4 is initialized by a separately determined algorithm, and Sb
Is initialized to 0 and then changed so that the value of each expression corresponding to each constraint condition decreases. For example, with respect to Equation 1, a,
b is changed as in Expressions 2 and 3 of FIG.

Example 2. In the above embodiment, the case of the manufacturing scheduling system has been described, but other scheduling systems such as a travel schedule, a sales schedule, and a test schedule may be used.

[0028]

Since the scheduling system according to the first aspect of the present invention is configured as described above, by using a neural network having a learning function for adjusting the variable constraint conditions, a variable value matching the scheduling result can be obtained. Constraints can be learned and the degree of adjustment of variable constraints can be held as knowledge. This allows the operator to perform scheduling without being aware of the variable constraint condition adjustment. Further, by using the above variable constraint condition adjusting neural network, it is not necessary to prepare a plurality of variable constraint conditions and perform scheduling for each case, and a neural network having an optimization function for scheduling is provided. By using it, the scheduling can be speeded up.

Since the condition-adjusting neural network according to the second aspect of the invention is configured as described above, the input / output is learned while moving in parallel on the time axis, so that the deviation on the time axis is eliminated. Variable constraints can be adjusted without being affected.

[Brief description of drawings]

FIG. 1 is a flowchart of a process according to a first embodiment of the present invention.

FIG. 2 is a diagram showing an example of a manufacturing process targeted for scheduling in the first embodiment.

FIG. 3 is a conceptual diagram illustrating input / output data of a variable constraint condition adjustment neural network.

FIG. 4 is a diagram showing an example of production capacity.

FIG. 5 is a schematic diagram of a variable constraint adjustment neural network.

FIG. 6 is a diagram showing a method of calculating a production capacity shortage.

FIG. 7 is a schematic diagram of a scheduling neural network.

FIG. 8 is a diagram showing neurons of the scheduling neural network 4.

9 is a diagram showing an example of a scheduling result of the scheduling neural network 4. FIG.

FIG. 10: Scheduling neural network 4
FIG. 6 is a diagram showing a function φ (x) used for the scheduling of FIG.

FIG. 11: Scheduling neural network 4
FIG. 7 is a diagram showing an equation for explaining the operation of FIG.

FIG. 12 is a flow chart of processing showing a conventional example.

[Explanation of symbols]

 1 Constraint Condition Setting Device 2 Variable Constraint Condition Adjustment Neural Network 3 Scheduling Weight Setting Device 4 Scheduling Neural Network 5 Display 24 Variable Constraint Condition (Example of Condition to be Adjusted) 25 Over / Under Production Capacity (Example of Indicator) 26 Window (Predetermined Example of range) 27 Time (day) 28 Scheduling result

[Procedure amendment]

[Submission date] March 11, 1992

[Procedure Amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0007

[Correction method] Change

[Correction content]

[0007]

In the scheduling system according to the first aspect of the present invention, the variable constraint adjusting neural network is provided with a learning function so that a desired variable constraint can be learned from the schedule result that changes according to the variable constraint and the current variable constraint. Therefore, the variable constraint condition and the knowledge can be acquired. Further, the scheduling neural network can perform scheduling at high speed by having a neural network having an optimization function.

[Procedure Amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0013

[Correction method] Change

[Correction content]

Next, the variable constraint condition shown by 24 in FIG. 3 will be described. The daily production capacity of each process in the manufacturing line is not fixed and can be enhanced for a certain period by means such as manpower input. Therefore,
During the period when the quantity of orders is large and the production is not in time by the delivery date,
Increase production capacity so that deadlines can be met. The production capacity is determined for each process and index as shown in FIG. 4. For example, the production capacity during the production capacity index 1 of the process A is 100 [Kg / day] , and the production capacity index 2 is 12 periods.
0 [Kg / day] . When this system operates , the operator must set the constraint condition setting device 1
Initialize using. Regarding the variable constraint condition 24, for example, time-series data such as the variable constraint condition 24 shown in FIG. 3 is given for the process A. Specifically, this time-series data is scheduled on a date and on that date. It is given by the production capacity index of the completed process A.
In this way, the index is selected as the variable constraint condition 24, and the production capacity before adjustment is set. The variable constraint adjustment neural network 2 is shown in FIG. 5 for each process of each line.
The variable constraint adjustment neural network 2 as shown in FIG. The variable constraint condition adjustment network 2 for each process of each line obtains the optimum production capacity for each process of each line according to the production capacity before adjustment and the excess / deficiency production capacity 25, and the scheduling system is the closest production capacity to this. Select the index and output. For example,
If the index 1 is input as the variable constraint condition 24 before adjustment of a certain date for the process A on the line 14,
The variable constraint adjustment neural network 2 for the process A on the line 14 outputs, for example, the index 2 as the optimum production capacity on that date. The production capacity at the input / output of the variable constraint adjustment neural network is normalized so as to be 1 at the maximum production capacity of the production line.

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0017

[Correction method] Change

[Correction content]

Next, the input / output and learning method of the variable constraint condition adjusting neural network 2 will be described. The variable constraint adjustment neural network 2 can learn the capability adjustment as follows after the scheduling is completed. First, during the adjustment, the input of the constraint condition adjustment neural network 2 is stored at some time points. After scheduling the end, the scheduling of the skilled person Sukeji
Make the scheduling adjustments that seem to be better for Turing Result 6 . The result is used as a desired output, and the stored input is used for learning as described below, whereby the capability of the variable constraint condition adjusting neural network 2 is improved.

[Procedure amendment 4]

[Document name to be amended] Statement

[Correction target item name] 0028

[Correction method] Change

[Correction content]

[0028]

Since the scheduling system according to the first aspect of the present invention is configured as described above, by using a neural network having a learning function for adjusting the variable constraint conditions, a variable value matching the scheduling result can be obtained. Constraints can be learned and the degree of adjustment of variable constraints can be held as knowledge. This allows the operator to perform scheduling without being aware of the variable constraint condition adjustment. Further, by using a variable constraint adjustment neural network described above, when necessary to the scheduling for each case a variable constraint to multiple cases prepared eliminates both Sukeji
The scheduling can be speeded up by using the neural network with the optimization function for the tasking .

[Procedure Amendment 5]

[Document name to be amended] Statement

[Correction target item name] Explanation of code

[Correction method] Change

[Correction content]

[Explanation of Codes] 1 constraint condition setting device 2 variable constraint condition adjusting neural network 3 scheduling weight setting device 4 scheduling neural network 5 display 6 scheduling result 24 variable constraint condition (an example of condition to be adjusted) 25 excessive and insufficient production capacity (index 26 window (an example of a predetermined range) 27 time (day)

[Procedure correction 6]

[Document name to be corrected] Drawing

[Name of item to be corrected] Figure 1

[Correction method] Change

[Correction content]

[Figure 1]

[Procedure Amendment 7]

[Document name to be corrected] Drawing

[Correction target item name] Figure 8

[Correction method] Change

[Correction content]

[Figure 8]

[Procedure Amendment 8]

[Document name to be corrected] Drawing

[Name of item to be corrected] Fig. 10

[Correction method] Change

[Correction content]

[Figure 10]

[Procedure Amendment 9]

[Document name to be corrected] Drawing

[Name of item to be corrected] Fig. 11

[Correction method] Change

[Correction content]

FIG. 11

Claims (2)

[Claims] 1. A scheduling system having the following elements: (a) a constraint condition setting device for setting a constraint condition for scheduling; (b) the constraint condition can be changed among the constraint conditions set by the constraint condition setting device. A variable constraint condition adjustment neural network for inputting and adjusting a variable constraint condition, (c) a constraint condition set by the constraint condition setting device, and a variable constraint condition adjusted by the variable constraint condition neural network. A scheduling weight setting device that inputs and sets weights for each condition, (d) A scheduling neural network that performs scheduling based on the weights set by the scheduling weight setting device. 2. A condition adjustment neural network having the following elements: (a) condition input means for inputting conditions to be adjusted as time series data, and (b) time series data as an index for adjusting the conditions to be adjusted. (C) using time series data in a predetermined range of the time series data input by the condition input means and the index input means,
A neural network that finds and outputs the optimum condition at a predetermined time point within the range, and moves and moves the predetermined range in the time series direction to find and output the optimum condition at each time point.