JP7645771B2 - Calculation device, work plan creation method, calculation system - Google Patents

Description

The present invention relates to a calculation device, a method for creating a work plan, and a calculation system.

In order to carry out work efficiently, it is common to create work plans in advance. However, when the conditions of the work environment change over time, it is not easy to create an optimal work plan. Patent Document 1 discloses a machine learning device that learns control commands for a machine through machine learning, characterized in that it includes a machine learning device that performs the machine learning and outputs the control commands, a simulator that executes a simulation of the work operation of the machine based on the control commands, and a first determiner that determines the control command based on the results of the simulation executed by the simulator.

JP 2017-185577 A

The invention described in Patent Document 1 leaves room for improvement in terms of calculation accuracy.

A computing device according to a first aspect of the present invention is a computing device that creates a task plan for a predetermined task in an environment in which information about a fluctuating environment can be expressed as environmental parameters, and includes a strict evaluation unit that calculates a required time for the predetermined task by performing a highly accurate evaluation with a high computational cost using a strict model , an evaluation scenario generation unit that generates a plurality of environmental parameters by randomly varying the latest environmental parameters within a predetermined range as evaluation scenarios, and a proxy model that uses learning data including a plurality of combinations of the environmental parameters and the task plan as input and the required time as output to calculate the required time more simply than the strict evaluation unit. a plan generation unit that calculates the required time using the proxy model and calculates a tentative work plan, which is the work plan suitable for any of the plurality of environmental parameters; and a proxy model update unit that, when a difference between an strict required time, which is the time calculated using the strict evaluation unit and is the required time in the tentative work plan, and a proxy required time, which is the time calculated using the proxy model and is the required time in the tentative work plan, is equal to or greater than a predetermined threshold , adds the tentative work plan, the evaluation scenario, and the strict required time to the learning data to generate added learning data, and updates the proxy model using the added learning data .
A method for creating a work plan according to a second aspect of the present invention is a method for creating a work plan by a computing device, for an environment in which information about a fluctuating environment can be expressed as environmental parameters, the method comprising the steps of: a strict evaluation process for calculating a required time for the specified work by performing a highly accurate evaluation with a high computational cost using a strict model ; an evaluation scenario generation process for generating, as an evaluation scenario, a plurality of environmental parameters obtained by randomly varying the latest environmental parameters within a predetermined range; and a process for performing a calculation for the specified work more simply than the strict evaluation process, by using learning data including a plurality of combinations of the environmental parameters and the work plan as input and the required time as output. a plan generation process for performing calculations for the specified work using the proxy model and calculating a tentative work plan, which is the work plan suitable for any of the plurality of environmental parameters; and a proxy model updating process for adding the tentative work plan, the evaluation scenario, and the strict required time to the learning data to generate added learning data, and updating the proxy model using the added learning data, when a difference between an strict required time, which is the time calculated by the strict evaluation process and is the required time in the tentative work plan, and a proxy required time, which is the time calculated using the proxy model and is the required time in the tentative work plan, is equal to or greater than a predetermined threshold.
A calculation system according to a third aspect of the present invention is a calculation system for creating a work plan for a predetermined task in an environment in which information about a fluctuating environment can be expressed as environmental parameters, the calculation system including a strict evaluation unit that calculates a required time for the predetermined task by performing a highly accurate evaluation with a high calculation cost using a strict model, an evaluation scenario generation unit that generates a plurality of environmental parameters by randomly varying the latest environmental parameters within a predetermined range as evaluation scenarios, and a surrogate model that uses learning data including a plurality of combinations of the environmental parameters and the work plan as input and the required time as output to perform calculations for the predetermined task more easily than the strict evaluation unit. a proxy model generation unit that uses the proxy model to perform calculations on the specified work and calculates a tentative work plan, which is the work plan suitable for any of the plurality of environmental parameters; and a proxy model update unit that, when a difference between an strict required time, which is the time calculated using the strict evaluation unit and is the required time in the tentative work plan, and a proxy required time, which is the time calculated using the proxy model and is the required time in the tentative work plan, is equal to or greater than a predetermined threshold , adds the tentative work plan, the evaluation scenario, and the strict required time to the learning data to generate added learning data, and updates the proxy model using the added learning data .

The present invention can improve the accuracy of simple calculations.

A table showing examples of applications to which the arithmetic unit can be applied. 1 is a block diagram of a computing device according to a first embodiment. A diagram showing an example of a work instruction A diagram showing an example of environmental parameters. A diagram showing an example of the expected fluctuation range FIG. 13 is a diagram showing an example of allowable prediction error information. A diagram showing an example of training data FIG. 13 is a diagram showing an example of representative model information. 1 is a flowchart showing an overview of processing by a computing device according to a first embodiment. Flowchart showing surrogate model generation processing Flowchart showing evaluation scenario generation processing Flowchart showing the plan calculation process FIG. 1 is a block diagram of a computing device according to a second embodiment. 11 is a flowchart showing an outline of processing performed by a computing device according to a second embodiment. FIG. 13 is a block diagram of a computing device according to a third embodiment. A diagram showing an example of parameter division A diagram showing an example of training data FIG. 13 is a diagram showing an example of representative model information. 11 is a flowchart showing an outline of processing performed by a computing device according to a third embodiment. 13 is a flowchart showing an overview of processing performed by a computing device according to a modification of the third embodiment.

(Target application)
The computing device according to the present invention creates a work plan in a changing environment. The computing device can be applied to various applications in which various environments change.

Figure 1 is a table showing examples of applications to which the arithmetic device can be applied. For each application, Figure 1 shows the instructions given to the arithmetic device, the calculation results of the arithmetic device, and the variable environmental parameters. An environmental parameter is a parameter that affects an application and can vary in a short time. "Short time" here means a short time compared to the time required for the arithmetic device to perform a calculation. Environmental parameters tend to change more as time passes. Therefore, the longer the calculation takes, the greater the variation in the environmental parameters between the start and end of the calculation. The "calculation results" shown in Figure 1 can be rephrased as a "work plan" calculated by the arithmetic device.

In the "product picking" application shown in the first record in Figure 1, an AGV (Automatic Guided Vehicle) picks up a countless number of products stored in a warehouse. A list of product names and the time when picking of the products should be completed is input to the calculation device. The location of each product in the warehouse is known to the AGV and the calculation device.

The warehouse is large, and it takes time for the AGVs to travel to the location where the products to be picked are stored. Therefore, the calculation device calculates and outputs which products each AGV will pick and in what order. However, the following environmental parameters can change in a short period of time. As shown in Figure 1, the environmental parameters in the "product picking" application are, for example, the picking speed, the number of AGVs, and the speed of the AGVs.

For example, the picking speed, which is the speed at which an AGV picks up goods, fluctuates due to various factors. Also, the number of AGVs that can operate is not constant due to breakdowns and charging. Furthermore, the maximum speed at which an AGV can travel varies depending on factors such as the condition of the floor. These environmental parameters change only slightly over a relatively short period of time, but as the period becomes longer, the changes can accumulate and result in larger changes.

In the "Home Delivery" application shown in the second record in Figure 1, delivery personnel deliver packages to homes. A list of addresses and deadline times is input to the calculation device. The association between addresses and packages to be delivered is managed separately. The calculation device outputs the order in which each worker will visit the delivery destinations. In this application, the environmental parameters are, for example, road conditions such as traffic jams and road closures, and the movement speed of workers, which changes depending on the weather.

In the "multi-product production" application shown in the third record in Figure 1, products are produced using multiple production devices. A list of product names and shipping times is input to the calculation device. The calculation device outputs the order in which products are processed for each production device. In this application, the environmental parameters are, for example, the number and type of production devices and the operating speed of the production devices. The present invention is not limited to the example shown in Figure 1 and can be applied to various applications, but the first embodiment will be described using "product picking" as an example.

-First embodiment-
Hereinafter, the first embodiment of the arithmetic device will be described with reference to FIGS.

Figure 2 is a configuration diagram of the arithmetic device 1 in the first embodiment. The arithmetic device 1 includes a CPU 11, which is a central processing unit, a ROM 12, which is a read-only storage device, a storage unit 13, which is a non-volatile storage device, and a memory 14, which is a storage device capable of high-speed reading and writing. The CPU 11 loads a program stored in the ROM 12 into the memory 14 and executes it to perform the calculations described below.

The computing device 1 may realize the functions described below using a field programmable gate array (FPGA), which is a rewritable logic circuit, or an application specific integrated circuit (ASIC), which is an integrated circuit for a specific application, instead of the combination of the CPU 11, ROM 12, and memory 14. The computing device 1 may also realize the functions described below using a combination of different configurations, for example, a combination of the CPU 11, ROM 12, RAM 13, and an FPGA, instead of the combination of the CPU 11, ROM 12, and memory 14.

The computing device 1 has as its functions a proxy model generation unit 41, a proxy model update unit 42, an evaluation scenario generation unit 43, a plan generation unit 44, a disturbance evaluation unit 45, a strict evaluation unit 51, and a simplified evaluation unit 52. As described above, these functions are realized by, for example, the CPU 11 expanding a program stored in the ROM 12 into the memory 14 and executing it, so in FIG. 2 these functions are conveniently described inside the memory 14.

The memory unit 13 of the computing device 1 stores work instructions 21, environmental parameters 22, expected variation range 23, allowable prediction error 24, learning data 25, proxy model information 26, and proxy model 27. However, the proxy model 27 is illustrated in FIG. 2 for the sake of convenience, and the main component of the proxy model 27 is the proxy model information 26. Details will be described later.

The proxy model generation unit 41 generates one or more proxy models 27. The proxy model update unit 42 updates the proxy models 27 generated by the proxy model generation unit 41. The evaluation scenario generation unit 43 acquires the environmental parameters 22 at the current time, and generates multiple evaluation scenarios that are randomly varied within a predetermined range. The plan proposal generation unit 44 uses the multiple evaluation scenarios generated by the evaluation scenario generation unit 43 and the simplified evaluation unit 52 to calculate a tentative plan proposal, which is a provisional work plan, and an approximate evaluation value of the tentative plan proposal.

The strict evaluation unit 51 performs a highly accurate evaluation with high computational cost using a strict model (not shown). The strict evaluation unit 51 is, for example, a simulator, and simulates the picking operation of each item in a picking task and the movement required for picking to calculate the required work time. The simplified evaluation unit 52 performs a simplified evaluation with low computational cost using a proxy model 27. The proxy model 27 may be realized by machine learning, such as a neural network that approximates a simulation evaluation, or may be a simulation model that is more abstract than the strict evaluation unit 51, or may be a simple mathematical formula.

The evaluation by the strict evaluation unit 51 is more accurate than that by the simple evaluation unit 52. The evaluation by the simple evaluation unit 52 is less accurate than that by the strict evaluation unit 51. The calculation cost means, for example, the amount of calculation, the calculation time, and the amount of calculation resources required for the calculation. For example, if the strict evaluation unit 51 and the simple evaluation unit 52 use the same calculation hardware, the strict evaluation unit 51 takes longer to complete the evaluation than the simple evaluation unit 52. Also, if the strict evaluation unit 51 and the simple evaluation unit 52 complete the evaluation in the same time, the strict evaluation unit 51 requires more calculation resources than the simple evaluation unit 52. Specifically, in order for the strict evaluation unit 51 and the simple evaluation unit 52 to complete the evaluation in the same time, the strict evaluation unit 51 needs to use more calculation cores than the simple evaluation unit 52, or the strict evaluation unit 51 needs to use a calculation core with a larger amount of calculation per unit time than the simple evaluation unit 52.

If there were no restrictions on computational cost, there would be no advantage to using the simplified evaluation unit 52, and it would be desirable to always use the strict evaluation unit 51. However, in reality, there are restrictions on computational cost, so in this embodiment, the strict evaluation unit 51 and the simplified evaluation unit 52 are used in an efficient combination. Furthermore, in this embodiment, the low accuracy of the strict evaluation unit 51 is compensated for by appropriately improving and selecting the simplified evaluation unit 52.

Figure 3 is a diagram showing an example of a work instruction 21. In this example, the work instruction 21 includes multiple combinations of a work instruction ID, a product, a deadline time, and an output quantity. The first record in Figure 3 instructs that six units of product "A" be prepared by 17:00. Note that the work instruction ID is an identifier that identifies an individual work instruction, and does not need to be included in the work instruction 21. Also, a similar instruction may be repeated in place of the output quantity; for example, product "A" and the deadline time "17:00" may be recorded in six consecutive records.

Figure 4 is a diagram showing an example of the environmental parameters 22. The environmental parameters 22 include multiple combinations of detailed parameters and values. Specifically, the environmental parameters 22 store parameters such as the number of AGVs and the picking speed. As described above, the environmental parameters 22 change over a short period of time, so the environmental parameters 22 shown here are merely values at a certain timing. The calculation device 1 may obtain the environmental parameters 22 from another device (not shown), or may obtain the individual values that make up the environmental parameters 22 from a sensor (not shown) connected to the calculation device 1.

Figure 5 is a diagram showing an example of the expected variation range 23. In the expected variation range 23, an expected variation range is described for each detailed parameter of the environmental parameter 22. The fluctuation width of the expected variation range 23 may be set based on statistical information of past detailed parameters, or may be set arbitrarily by the operator. Since the value of the expected variation range 23 is merely a statistical value or an arbitrarily set value, it is possible that the environmental parameter 22 may actually vary beyond the range described in the expected variation range 23. However, in this embodiment, no proactive measures are taken against the environmental parameter 22 varying beyond the expected variation range 23. In the example shown in Figure 5, it is shown that the number of AGVs is ±2, the picking speed is ±20 items/hour, etc.

Figure 6 is a diagram showing an example of the allowable prediction error information G4. The allowable prediction error information G4 defines the allowable range of prediction error between the strict evaluation unit 51 and the simple evaluation unit 52. The value of the allowable prediction error information G4 is set arbitrarily by the operator, for example. The first record in Figure 6 states that the time required to complete all work is allowable if it is within ±30 minutes. This "required time" can also be called "work time" or "processing time." The second record in Figure 6 states that the time until the delivery date is allowable if it is ±15%.

Figure 7 is a diagram showing an example of the training data 25. The training data 25 stores data used to generate the proxy model 27. The training data 25 may store data in advance, or may be empty, i.e., no data is stored, when the proxy model generation process described below is started. Data is added to the training data 25 by the proxy model generation process and the update process described below.

The learning data 25 includes a plurality of samples. Each sample includes an input to the strict evaluation unit 51 and an output from the strict evaluation unit 51. In the example shown in FIG. 7, each record of the learning data 25 includes a sample ID that identifies the sample, a data attribute that indicates the distinction between input and output to the strict evaluation unit 51, a data name, and a data value. Information on records of the learning data 25 that have the same sample ID and whose data attribute is "input" is input to the strict evaluation unit 51 at the same time.

The combination of data name and data value in the learning data 25 corresponds to a record of a work instruction 21. Differences in sample IDs indicate different values of environmental parameters. Although not shown in FIG. 7, the specific values of the environmental parameters corresponding to each sample ID are also recorded separately.

Figure 8 is a diagram showing an example of surrogate model information 26. The surrogate model information 26 stores the names and values of internal parameters, which are information about the surrogate model 27 generated using the training data 25. In the example shown in Figure 8, the surrogate model information 26 stores information about the parameter values for each internal parameter ID that identifies the internal parameter.

The proxy model 27 is configured by combining information about the proxy model 27 (not shown) with the proxy model information 26. The proxy model 27 is, for example, a neural network, and calculations can be performed using the proxy model 27 by combining information about the configuration of the proxy model 27 (not shown) with the parameters stored in the proxy model information 26. Furthermore, for example, the proxy model 27 is a polynomial, and calculations can be performed using the proxy model 27 by combining the mathematical expression of the polynomial (not shown) with the coefficient values in the polynomial stored in the proxy model information 26. This concludes the description of the information stored in the memory unit 13.

(Overview of Processing by the Arithmetic Unit)
9 is a flowchart showing an outline of the processing by the computing device 1. First, in step S31, the proxy model generation unit 41 generates the proxy model 27. This processing is executed in advance before receiving the work instruction 21. Next, in step S32, the computing device 1 acquires the work instruction 21 from outside, for example, a list of product names and deadline times. Next, in step S33, the evaluation scenario generation unit 43 acquires the value of the environmental parameter 22 at the time t, and generates a plurality of virtual environmental parameters (hereinafter referred to as "evaluation scenarios") that are randomly changed based on this value.

In the next step S34, the plan generation unit 44 calculates an optimal work plan (hereinafter referred to as the "provisional work plan") using the proxy model 27, and further calculates the average required time when the provisional work plan is applied to the evaluation scenario using the proxy model 27. "Optimal" here means that the required time is short on average for the multiple environmental parameters generated in step S33. Note that the average required time calculated in this step is also referred to as the "proxy required time" because it is calculated using the proxy model 27.

In the next step S35, the strict evaluation unit 51 calculates the average required time when the tentative work plan calculated in step S34 is applied to the evaluation scenario. That is, the strict evaluation unit 51 performs calculations the same number of times as the number of environmental parameters 22 included in the evaluation scenario to calculate each required time, and calculates the average value of these. Note that the average required time calculated in this step is also called the "strict required time" because it is calculated using the strict evaluation unit 51.

In the next step S36, the plan generation unit 44 determines whether the difference between the average required time of the interim work plan calculated in step S34 and the average required time of the interim work plan calculated in step S35 is within a threshold value. In other words, in this step, the plan generation unit 44 determines whether the difference between the proxy required time and the strict required time is within a threshold value.

If the plan generation unit 44 determines that the difference between the two average required times is within the threshold, the process proceeds to step S37, where the provisional work plan is output, and the process of FIG. 9 ends. If the plan generation unit 44 determines that the difference between the two average required times is equal to or greater than the threshold, the process proceeds to step S38.

In step S38, the surrogate model update unit 42 adds the input to the strict evaluation unit 51 in step S35 and the output of the strict evaluation unit 51 in the same step to the learning data 25, and further updates the surrogate model 27 using this learning data 25, and returns to step S34. Updating the surrogate model 27 means updating the parameter values in the surrogate model information 26. Note that in step S34, which is executed by transitioning from step S38, an updated surrogate model 27 different from that in the initial step S34 is used. Since the learning data 25 of the updated surrogate model 27 contains the same data as the environmental parameters 22 and work instructions 21 to be used in the calculation to be performed, there is a high probability that the accuracy will be higher than that of the learning data 25 before the update. This concludes the explanation of FIG. 9.

Figure 10 is a flowchart showing the proxy model generation process by the proxy model generation unit 41. Figure 10 shows the details of step S31 shown in Figure 9. In step S311, the proxy model generation unit 41 determines the reference value and the variation range of the environmental parameter 22. The proxy model generation unit 41 may adopt the variation range described in the assumed variation range 23, or may randomly determine the value of any of the detailed parameters included in the environmental parameter 22. For example, the proxy model generation unit 41 sets the picking speed to 150 pieces per hour, the number of AGVs to 20, and the AGV speed to 10 km per hour as the reference values of the environmental parameter 22. In addition, the variation ranges of the environmental parameter 22 are set to the picking speed of ±50 pieces/hour, the number of AGVs to ±5, and the AGV speed to ±2 km per hour.

In the next step S312, the proxy model generation unit 41 randomly generates the environmental parameters 22 within the variation range determined in step S301. For example, if 10 picking speeds, 10 number of AGVs, and 10 speeds of AGVs are randomly calculated in step S302, the environmental parameters 22 are calculated in 10 to the power of 3, that is, 1000 different ways.

In the next step S313, the surrogate model generation unit 41 randomly generates multiple execution plans, i.e., work instructions and work sequences. In the next step S314, the surrogate model generation unit 41 combines the multiple environmental parameters 22 generated in step S302 with the multiple execution plans generated in step S303, inputs them to the strict evaluation unit 51, and calculates the required time for each. For example, if there are 1,000 environmental parameters 22 generated in step S302 and 100 execution plans generated in step S303, the strict evaluation unit 51 performs 10,000 calculations in step S304.

In step S315, the surrogate model generation unit 41 compiles the environmental parameters 22 and execution plan input to the strict evaluation unit 51 in step S304, and the required time output by the strict evaluation unit 51 at that time, into one set of learning data. In the above example, 10,000 sets of learning data are obtained. In the following step S316, the surrogate model generation unit 41 generates a surrogate model 27 using the multiple learning data obtained in step S305. If the execution plan is X, the environmental parameters 22 are θ, and the required time is Y, then the surrogate model 27 corresponds to the function f in the following equation 1.

Y≒f(X, θ)...(Formula 1)

The proxy model 27 may be realized by machine learning, such as a neural network that approximates a simulation evaluation, or may be a simulation model that is more abstract than the strict evaluation unit 51, or may be a simple mathematical formula. This concludes the description of the proxy model generation process.

Figure 11 is a flowchart showing the evaluation scenario generation process by the evaluation scenario generation unit 43. Figure 11 shows details of step S33 shown in Figure 9. In step S331, the evaluation scenario generation unit 43 acquires the current value of the environmental parameter 22. In the following step S332, the evaluation scenario generation unit 43 determines the variation range of the environmental parameter 22. For example, the evaluation scenario generation unit 43 adopts a value described in the expected variation range 23.

In the next step S333, the evaluation scenario generation unit 43 randomly generates values for the environmental parameters 22 within the range determined in step S332, with the value acquired in step S331 at the center. For example, if the value of the detailed parameter "picking speed" acquired in step S331 is "150" and the range of "picking speed" determined in step S332 is "±30", multiple values are randomly determined within the range of "±30" with "150" at the center, that is, within the range of "120" to "180".

In the next step S334, the evaluation scenario generation unit 43 compiles the multiple environmental parameters 22 determined in step S333 and stores them in the storage unit 13 as an evaluation scenario. There is no particular limit to the number of environmental parameters 22 included in the evaluation scenario. For example, all possible combinations of the values of each detailed parameter determined in step S333 may be used. This concludes the description of the evaluation scenario generation process.

Figure 12 is a flowchart showing the plan calculation process by the plan generation unit 44. Figure 12 shows details of step S34 shown in Figure 9. In step S341, the plan generation unit 44 randomly generates a work plan and proceeds to step S342. In step S342, the plan generation unit 44 selects one unselected work plan from the work instructions randomly generated in step S341 and proceeds to step S343.

In step S343, the plan generation unit 44 selects one unselected environmental parameter 22 from the evaluation scenario. In the following step S344, the plan generation unit 44 inputs the environmental parameter 22 selected in step S342 and the work plan selected in step S342 to the proxy model 27. Strictly speaking, the plan generation unit 44 inputs the environmental parameter 22 and the work plan to the simplified evaluation unit 52, which performs calculations using the proxy model 27. In the following step S345, the estimated required time calculated by the proxy model 27 in step S344 is temporarily recorded.

In the next step S346, the plan creation unit 44 judges whether or not all of the environmental parameters 22 included in the evaluation scenario have been selected in step S343. If the plan creation unit 44 judges that all of the environmental parameters 22 have been selected, the process proceeds to step S347, and if it judges that at least one unselected environmental parameter 22 remains, the process returns to step S343. In other words, if a positive judgment is made in this step, the estimated required times corresponding to all of the environmental parameters 22 included in the evaluation scenario for a certain work plan have already been calculated.

In step S347, the plan generation unit 44 calculates and records the average value of the estimated required time that has been temporarily recorded in step S345. Once recording of the average value is complete, all values temporarily recorded in step S345 are erased.

In the next step S348, the plan generation unit 44 determines whether or not all of the work plans randomly generated in step S341 have been selected in step S342. If the plan generation unit 44 determines that all of the work plans have been selected, the process proceeds to step S349, and if it determines that at least one unselected work plan remains, the process returns to step S342. If a negative determination is made in this step, the selection status of the environmental parameters 22 is reset, and it is considered that no environmental parameters 22 have been selected so far.

In step S349, the plan generation unit 44 sets a tentative work plan and identifies the average time required by the proxy model of the tentative work plan. Specifically, the plan generation unit 44 first identifies the minimum value from the average values recorded in step S347, and sets the work plan corresponding to the minimum average value as the tentative work plan. This concludes the explanation of Figure 12.

According to the above-described first embodiment, the following advantageous effects can be obtained.
(1) The computing device 1 creates a task plan for a given task in an environment in which information on a changing environment can be expressed as environmental parameters 22. The computing device 1 includes a strict evaluation unit 51 that calculates a required time for the given task, an evaluation scenario generation unit 43 that generates a plurality of environmental parameters 22 as an evaluation scenario based on the latest environmental parameters 22, a proxy model generation unit 41 that uses learning data 25 to create a proxy model 27 for calculating a required time for the given task more easily than the strict evaluation unit 51, a plan proposal generation unit 44 that calculates the required time using the proxy model 27 and calculates a tentative task plan that is a task plan suitable for any of the plurality of environmental parameters 22, and a proxy model update unit 42 that updates the tentative task and the strict required time when a difference between a strict required time calculated using the strict evaluation unit 51 and being a required time for the given task in the tentative task plan and a proxy required time calculated using the proxy model 27 and being a required time for the given task in the tentative task plan is equal to or greater than a predetermined threshold value. Therefore, when the accuracy of the proxy model 27 is low and there is a discrepancy between the calculation results using the proxy model 27 and the calculation results using the strict evaluation unit 51, the proxy model 27 can be updated using the calculation results of the strict evaluation unit 51, thereby improving the accuracy of the proxy model 27 used for the simple calculation.

(2) The environmental parameters 22 include information on the speed at which a specific task is performed, i.e., the picking speed. The evaluation scenario generation unit 43 generates an evaluation scenario based on the expected variation range 23, which describes the range within which the environmental parameters 22 are to be varied.

(Variation 1)
In the first embodiment described above, the arithmetic device 1 has been described as a single hardware device. However, the arithmetic device 1 may be composed of multiple pieces of hardware, and each function may be realized by a different computer. In this case, the functions of the arithmetic device 1 may be realized by multiple pieces of hardware that are connected by a network and located in physically separate locations.

--Second embodiment--
A second embodiment of the arithmetic device will be described with reference to Figures 13 and 14. In the following description, the same components as in the first embodiment are given the same reference numerals, and differences will be mainly described. Points that are not particularly described are the same as in the first embodiment. This embodiment differs from the first embodiment mainly in that the calculation is redone if the environmental parameters fluctuate significantly during the calculation.

Figure 13 is a configuration diagram of the calculation device 1A in the second embodiment. In addition to the configuration in the first embodiment, the calculation device 1A further includes a disturbance evaluation unit 45. The disturbance evaluation unit 45 evaluates the deviation between the environmental parameters 22 at the start of operation of the plan generation unit 44 and the current environmental parameters 22.

Figure 14 is a flowchart showing an outline of the processing of the arithmetic device 1A in the second embodiment. The difference from Figure 9 in the first embodiment is that step S39 is added between steps S34 and S35. In step S39, the disturbance evaluation unit 45 compares the value of the environmental parameter 22 acquired at the beginning of step S33, i.e., step S331 in Figure 11, with the latest value of the environmental parameter 22, and determines whether the value of the environmental parameter 22 has changed significantly since the execution of step S33. If the disturbance evaluation unit 45 determines that the value has changed significantly, it returns to step S33, and if it determines that the value has not changed significantly, it proceeds to step S35. The criterion for determining whether the change in the environmental parameter 22 is large may be determined separately and individually for each detailed parameter, or the value of the expected variation range 23 may be used.

According to the above-described second embodiment, the following advantageous effects can be obtained.
(3) The evaluation scenario generating unit 43 generates an evaluation scenario based on the start-time environmental parameters, which are the latest environmental parameters 22. When the difference between the completion-time environmental parameters, which are the environmental parameters 22 when the calculation of the tentative work plan using the evaluation scenario, and the start-time environmental parameters is equal to or greater than a predetermined threshold, the plan generating unit 44 causes the evaluation scenario generating unit 43 to create an evaluation scenario using the latest environmental parameters 22 and calculates the tentative work plan again. Therefore, since the environmental parameters 22 fluctuate in a short time, calculation using perfectly matching values of the environmental parameters 22 is impossible, but if the environmental parameters 22 fluctuate significantly while the plan generating unit 44 is generating the plan, a new plan is generated, so that a plan can be generated using environmental parameters 22 that are less different from the actual values.

(Modification of the second embodiment)
The calculation of the tentative work plan may be redone when the elapsed time from the start of calculation exceeds a predetermined threshold. For example, step S39 in Fig. 14 may be further modified so that if the elapsed time from the start of the processing of step S33 exceeds a predetermined time, a positive determination is made and the process returns to step S33. Also, immediately after step S35, a branch may be added to return to step S33 if the elapsed time from the start of the processing of step S33 exceeds a predetermined time.

--Third embodiment--
A third embodiment of the arithmetic device will be described with reference to Figures 15 to 19. In the following description, the same components as in the first embodiment are given the same reference numerals, and differences will be mainly described. Points that are not specifically described are the same as in the first embodiment. This embodiment differs from the first embodiment mainly in that a plurality of proxy models are used selectively.

Figure 15 is a configuration diagram of a calculation device 1B in the third embodiment. In addition to the configuration in the first embodiment, the calculation device 1A further stores parameter classification 28 in the memory unit 13. Furthermore, learning data 25A and surrogate model information 26A are stored in the memory unit 13 instead of learning data 25 and surrogate model information 26.

The surrogate model 27 has the advantage of requiring less calculations than the strict evaluation unit 51, but the accuracy of the calculations tends to decrease when the value of the environmental parameter 22 changes. For this reason, multiple surrogate models 27 are generated in advance and the surrogate models 27 are used according to the value of the environmental parameter 22. In this embodiment, the value of the environmental parameter 22 is divided into multiple categories in advance, and a data set is formed for each category. Then, a surrogate model is generated for each data set. Information indicating the category of the environmental parameter 22 is stored as a parameter category 28 in the storage unit 13. In this embodiment, multiple surrogate models 27 are generated, so the learning data 25A and surrogate model information 26A for generating these multiple surrogate models 27 also differ from those in the first embodiment.

Figure 16 is a diagram showing an example of parameter category 28. Parameter category 28 shows the correspondence between the dataset ID and the value range of environmental parameter 22. For example, the value range of environmental parameter 22 corresponding to dataset ID "DS001" is the number of AGVs of "8 to 12" and the picking speed of "80 to 120 items/hour". The value range of environmental parameter 22 corresponding to the next dataset ID "DS002" has the same picking speed, but differs from the first dataset in that the number of AGVs is "13 to 20". Note that the values of environmental parameter 22 of the datasets do not have to be completely identical, and there may be overlap in the value range.

Figure 17 is a diagram showing an example of learning data 25A in this embodiment. The difference from learning data 25 in the first embodiment is that data set ID information has been added. In the example shown in Figure 17, data for two sample IDs whose environmental parameters 22 fall into the same category is shown.

Figure 18 is a diagram showing an example of the proxy model information 26A in this embodiment. The difference from the proxy model information 26 in the first embodiment is that a proxy model ID, which is an identifier for the proxy model, and a dataset ID, which indicates the dataset used to generate the proxy model, are added.

Figure 19 is a flowchart showing an overview of the processing of the calculation device 1B in the third embodiment. The difference from Figure 9 in the first embodiment is that step S391 is added between steps S33 and S34. In step S391, the plan generation unit 44 selects the optimal proxy model 27, and uses that proxy model 27 in steps S34 and after.

The specific processing of the plan generation unit 44 in this step is as follows. First, the plan generation unit 44 reads the value of the current environmental parameter 22, and identifies the dataset ID corresponding to the range in the parameter category 28 that includes the value of the current environmental parameter 22. The plan generation unit 44 then reads the internal parameter value of the proxy model 27 that corresponds to the identified dataset ID by referring to the proxy model information 26A, and uses the proxy model 27.

According to the above-described third embodiment, the following advantageous effects can be obtained.
(CL4) The evaluation scenario generation unit 43 generates a plurality of evaluation scenarios with different value ranges for the environmental parameters. The proxy model generation unit 41 generates proxy models 27 corresponding to each of the evaluation scenarios, and the plan generation unit 44 selects one of the proxy models 27 according to the value of the environmental parameter 22. Therefore, by selecting the proxy model 27 according to the situation, it is possible to improve the accuracy of the calculation.

(Modification of the third embodiment)
The configuration of the second embodiment may be combined with the configuration of the third embodiment. In this case, there are advantages that the plan generation unit 44 performs calculations using the optimal proxy model 27, and that recalculation can be performed if the environmental parameters 22 change significantly during processing by the plan generation unit 44. In this case, the parameter classification 28 can be used for the judgment in step S39.

Figure 20 is a flowchart showing an overview of the processing of the calculation device 1B in this modified example. However, in Figure 20, the description from the start to step S31 is omitted for convenience of drawing. Comparing this flowchart with Figure 9 in the first embodiment, step S391 has been added between step S33 and step S34, and step S39 has been added between step S34 and step S35.

In each of the above-mentioned embodiments and variations, the functional block configurations are merely examples. Some functional configurations shown as separate functional blocks may be configured together, or a configuration shown in a single functional block diagram may be divided into two or more functions. In addition, some of the functions of each functional block may be provided by other functional blocks.

In each of the above-described embodiments and variations, the program is stored in ROM 12, but the program may be stored in storage unit 13. Furthermore, the arithmetic device 1 may be provided with an input/output interface (not shown), and the program may be read from another device when necessary via the input/output interface and a medium available to the arithmetic device 1. Here, the medium refers to, for example, a storage medium that is detachable from the input/output interface, or a communication medium, i.e., a network such as a wired, wireless, or optical network, or a carrier wave or digital signal that propagates through the network. Furthermore, some or all of the functions realized by the program may be realized by a hardware circuit or FPGA.

The above-mentioned embodiments and modifications may be combined with each other. Although various embodiments and modifications have been described above, the present invention is not limited to these. Other aspects that are conceivable within the scope of the technical concept of the present invention are also included within the scope of the present invention.

1, 1A, 1B... Calculation device 13... Storage unit 14... Memory 21... Work instructions 22... Environmental parameters 23... Expected variation range 24... Allowable prediction error 25, 25A... Learning data 26, 26A... Proxy model information 27... Proxy model 28... Parameter classification 41... Proxy model generation unit 42... Proxy model update unit 43... Evaluation scenario generation unit 44... Plan generation unit 45... Disturbance evaluation unit 51... Strict evaluation unit 52... Simplified evaluation unit

Claims (6)

A computing device that creates a task plan for a predetermined task in an environment in which information about a changing environment can be expressed as an environmental parameter, comprising:
a strict evaluation unit that calculates a required time for the predetermined task by performing a highly accurate evaluation with a high computational cost using a strict model;
an evaluation scenario generating unit that generates, as an evaluation scenario, a plurality of environmental parameters obtained by randomly varying the latest environmental parameters within a predetermined range;
a surrogate model generation unit that uses learning data including a plurality of combinations of the environmental parameters and the work plan as inputs, and the required time as output, to create a surrogate model for calculating the required time more simply than the strict evaluation unit;
a plan generation unit that calculates the required time using the representative model and calculates a tentative work plan, which is the work plan suitable for all of the plurality of environmental parameters;
and a proxy model updating unit that, when a difference between an strict required time, which is the time calculated using the strict evaluation unit and is the required time in the tentative work plan, and a proxy required time, which is the time calculated using the proxy model and is the required time in the tentative work plan, is equal to or greater than a predetermined threshold , adds the tentative work plan, the evaluation scenario, and the strict required time to the learning data as added learning data, and updates the proxy model using the added learning data . 2. The computing device according to claim 1,
The environmental parameters include information on the speed at which the predetermined task is performed;
The evaluation scenario generation unit is a computing device that generates the evaluation scenario based on environmental parameter variation information in which the predetermined range is described. 2. The computing device according to claim 1,
the evaluation scenario generation unit generates the evaluation scenario based on start-time environmental parameters which are the latest environmental parameters;
The plan proposal generation unit, when a difference between the completion environmental parameters, which are the environmental parameters at the time when calculation of the interim work plan using the evaluation scenario is completed, and the start environmental parameters, is equal to or greater than a predetermined threshold, causes the evaluation scenario generation unit to create the evaluation scenario using the latest environmental parameters and calculates the interim work plan again. 2. The computing device according to claim 1,
the evaluation scenario generation unit generates a plurality of evaluation scenarios each having a different value range for the environmental parameter;
the surrogate model generation unit generates the surrogate models corresponding to each of the evaluation scenarios;
The plan generation unit selects one of the representative models depending on the value of the environmental parameter. A method for creating a work plan, the method comprising: creating a work plan for a predetermined work in an environment in which information about a changing environment can be expressed as an environmental parameter by a computing device, the method comprising:
A rigorous evaluation process that uses a rigorous model to perform a highly accurate evaluation with a high computational cost to calculate a required time for the predetermined task;
an evaluation scenario generation process for generating a plurality of environmental parameters by randomly varying the latest environmental parameters within a predetermined range as evaluation scenarios;
a surrogate model generation process for generating a surrogate model for performing calculations related to the predetermined work more easily than the strict evaluation process, using learning data including a plurality of combinations of the environmental parameters and the work plan as inputs, and the required time as output;
a plan generation process for performing a calculation related to the predetermined work using the proxy model and calculating a tentative work plan, which is the work plan suitable for all of the plurality of environmental parameters;
a proxy model updating process for adding the tentative work plan, the evaluation scenario, and the strict required time to the learning data to generate added learning data, and updating the proxy model using the added learning data, when a difference between an strict required time, which is the time calculated by the strict evaluation process and is the required time in the tentative work plan, and a proxy required time, which is the time calculated using the proxy model and is the required time in the tentative work plan, is equal to or greater than a predetermined threshold. A computing system for creating a task plan for a predetermined task in an environment in which information about a changing environment can be expressed as an environmental parameter, comprising:
a strict evaluation unit that calculates a required time for the predetermined task by performing a highly accurate evaluation with a high computational cost using a strict model;
an evaluation scenario generating unit that generates, as an evaluation scenario, a plurality of environmental parameters obtained by randomly varying the latest environmental parameters within a predetermined range;
a surrogate model generation unit that uses learning data including a plurality of combinations of the environmental parameters and the task plan as inputs, and the required time as output, to generate a surrogate model for performing calculations related to the predetermined task more easily than the strict evaluation unit;
a plan generation unit that performs a calculation related to the predetermined work using the proxy model and calculates a tentative work plan, which is the work plan suitable for all of the plurality of environmental parameters;
a proxy model updating unit that, when a difference between an strict required time, which is the time calculated using the strict evaluation unit and is the required time in the tentative work plan, and a proxy required time, which is the time calculated using the proxy model and is the required time in the tentative work plan, is equal to or greater than a predetermined threshold , adds the tentative work plan, the evaluation scenario, and the strict required time to the learning data to generate added learning data, and updates the proxy model using the added learning data .

Images