JP2018181184A - Evaluation device, evaluation method, and program therefor - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an evaluation device, an evaluation method, and a program therefor that can deal with the case where control performance of a learning module is degraded after additional learning.SOLUTION: An evaluation device 60 evaluating a second learning module additionally learning a first learning module comprises: a learning target receiving part 203 that receives a learning target to be achieved by the second learning module; an evaluation part 204 that evaluates the second learning module in terms of at least evaluation items included in the learning target to create evaluation data; a determination part 205 that determines whether the second learning module achieves the learning target by using the learning target and evaluation data; and a learning module selection part 206 that, when the second learning module is determined not to achieve the learning target, acquires a third learning module different from the second learning module on the basis of at least the learning target.SELECTED DRAWING: Figure 6

Description

  The present invention relates to an evaluation device, an evaluation method and a program thereof.

  It is known in the prior art to control a system using neural network technology. For example, Patent Document 1 discloses an elevator system that uses a neural network to select an optimal car to be moved from a plurality of cars to a landing.

  Further, Patent Document 2 discloses a method of performing additional learning on a neural network in order to reduce erroneous recognition in a pattern recognition apparatus using the neural network.

JP, 2005-255289, A JP-A-9-62648

  In control systems using neural network technology, additional learning of neural networks may be performed to improve control performance. For example, at the time of the first learning, learning data A randomly selected from data obtained in the period of time 1 to time 2 in FIG. 11 is used, and then data obtained in the period of time 2 to time 3 in FIG. There may be a case where additional learning is performed using learning data B randomly selected from.

  In such a case, the learning data B may contain inappropriate learning data due to random selection of the learning data. As described above, when additional learning is performed using additional learning data B including inappropriate learning data, the control performance of the neural network after the additional learning is degraded compared to the neural network before the additional learning. There is.

  In addition, when a special event occurs in the period from time 2 to time 3, there is a possibility that the neural network after the additional learning becomes specialized for the special event. As a result, the control performance of the neural network after the additional learning may be degraded compared to the neural network before the additional learning.

  Then, this invention aims at providing the evaluation apparatus which can cope with the case where the control performance of a learning module deteriorates after additional learning, an evaluation method, and its program.

  An evaluation apparatus for evaluating a second learning module which additionally learns a first learning module according to an aspect of the present invention includes: a learning target receiving unit that receives a learning target to be achieved by the second learning module; Whether the second learning module has achieved the learning goal by using the evaluation unit that evaluates the second learning module about the evaluation items included in the goal and generates evaluation data, and the learning goal and the evaluation data A determination unit is provided, and a learning module selection unit that acquires a third learning module different from the second learning module based on at least the learning target when it is determined that the learning target is not achieved.

  According to this aspect, if the control performance of the learning module does not achieve the learning goal after the additional learning, it is possible to avoid continuing to use the learning module that did not achieve the learning goal. The reliability of the system used can be increased. For example, by avoiding the use of a learning module that did not achieve the learning goal, it is possible to prevent a decrease in the processing accuracy of the entire system. In addition, when not achieving the learning goal, a control device that performs control using the learning module is evaluated, by using a configuration that acquires a learning module different from the learning module that did not achieve the learning goal based on at least the learning goal. Since none of the evaluation devices need to hold multiple learning modules, the hardware resources for recording learning modules can be minimized.

  In the evaluation device, the third learning module may be a first learning module. According to this aspect, since control of the system can be continued using the learning module prior to performance degradation, it is possible to prevent deterioration in the stability of the system using the learning module.

  The evaluation device may further include a factor estimation unit configured to estimate factor that the learning module could not achieve the learning target, and to generate factor improvement data related to learning data used for additional learning of the learning module. According to this aspect, useful information can be obtained from the degraded learning module. By utilizing such information, it is possible to reduce learning data and processing required to improve the accuracy of the learning module, and to reduce the load on the CPU.

  In the above-described evaluation device, the learning module selection unit may generate a learning instruction based on the factor improvement data. According to this aspect, it is possible to cause the learning module to perform efficient additional learning, which can overcome the problem of the degraded learning module. By using the learning module that has been subjected to this efficient additional learning in the system, it is possible to improve the performance deterioration due to the updating of the learning module in the shortest time.

  In the evaluation device, the factor estimation unit may output the estimated factor, receive a user input regarding the estimated factor, and generate factor improvement data based on the user input. According to this aspect, it is possible to generate the information on the performance deterioration factor based on the more reliable information.

  According to another aspect of the present invention, an evaluation method for evaluating a second learning module obtained by additionally learning a first learning module includes the steps of: receiving a learning target to be achieved by the second learning module; The second learning module is evaluated with respect to the evaluation items included in, and it is determined whether the second learning module has achieved the learning target, using the step of generating the evaluation data and the learning target and the evaluation data. And a step of acquiring a third learning module different from the second learning module based on at least the learning target if it is determined that the learning target is not achieved.

  A program according to another aspect of the present invention is a program for causing a computer to evaluate a second learning module to which the first learning module is additionally learned, and a learning goal to be achieved by the second learning module. The second learning module uses the learning goal and the evaluation data to evaluate the second learning module for the process of accepting the second learning module for at least the evaluation item included in the learning goal, and the learning goal for the second learning module. And a process of acquiring a third learning module different from the second learning module when it is determined that the learning goal is not achieved.

  According to the present invention, it is possible to provide an evaluation apparatus, an evaluation method, and a program thereof that can cope with the case where the control performance of the learning module degrades after the additional learning.

It is a figure showing the network composition of the learning system concerning the embodiment of the present invention. It is a figure showing physical composition of a learning device concerning an embodiment of the present invention. It is a figure showing physical constitution of an additional learning control device concerning an embodiment of the present invention. It is a functional block diagram of a learning device concerning an embodiment of the present invention. It is a functional block diagram of an additional learning control device concerning an embodiment of the present invention. It is a flowchart which shows the flow of the additional learning process performed by the additional learning control apparatus which concerns on embodiment of this invention. It is a flowchart which shows the flow of the additional learning process performed by the additional learning control apparatus which concerns on embodiment of this invention. It is an example of the screen output by the additional learning control apparatus which concerns on embodiment of this invention. It is a figure which shows the network configuration of the learning system which concerns on another embodiment of this invention. It is a figure which shows the physical structure of the evaluation apparatus which concerns on embodiment of this invention. It is a figure which shows the time series at the time of selecting learning data.

  Embodiments of the present invention will be described with reference to the accompanying drawings. The following embodiments are for the purpose of facilitating the understanding of the present invention, and are not for the purpose of limiting and interpreting the present invention. Moreover, the present invention can be variously modified without departing from the scope of the invention. Further, those skilled in the art can adopt an embodiment in which each of the elements described below is replaced with an equivalent one, and such embodiments are also included in the scope of the present invention.

(Network configuration)
A network configuration of a learning system 1 according to a predetermined embodiment of the present invention will be described with reference to FIG. The learning system 1 includes a learning device 10, an additional learning control device 20, one or more sensors 30, and a storage device 40. The learning device 10 is connected to the additional learning control device 20, one or more sensors 30 and a storage device 40 via the communication network N. The communication network N may be any of a wired communication network and a wireless communication network configured by a wired or wireless channel, and may be the Internet or a Local Area Network (LAN).

  The learning device 10 learns the learning module based on the learning data stored in the storage device 40, and stores the learned module in the storage device 40. The learning device 10 according to the present embodiment includes a learning module, but the learning module may be provided in a device separate from the learning device 10.

  The learning module includes one unit of dedicated or general-purpose hardware or software having a learning ability, or a combination of one unit of the hardware or software. Some learning modules that perform the learning include those that have already performed some learning using learning data, and those that have not yet learned. Here, the learning ability refers to the ability to improve the processing ability of a certain task based on the experience obtained from the learning data.

  The additional learning control device 20 outputs the output data according to the feature of the input data using the learned module. The additional learning control device 20 according to the present embodiment acquires a learned module or a copy of the learned module from the learning device 10, and sets it as a learning module. The additional learning control device 20 can evaluate the output data output using the set learning module, and can output evaluation data. The additional learning control device 20 can use the learning target to be achieved by the set learning module and the evaluation data to determine whether the currently set learning module has achieved the learning target. When it is determined that the learning goal is not achieved, the additional learning control device 20 may acquire, for example, a learning module set previously, from the learning device 10 and set it as a learning module. The additional learning control device 20 has a functional configuration of an evaluation device 60 described later, and substantially includes the evaluation device 60.

  The copy of the learned module includes one unit of dedicated or general-purpose hardware or software capable of reproducing the function of the learned module, or a combination of one unit of the hardware or software.

  The duplicate of the learned module may not necessarily have the learning ability. Also, the configuration of the learned module and the configuration of the duplicate of the learned module do not necessarily have to match. Also, a copy of the learned module includes a learning module obtained by so-called distillation. That is, the copy of the learned module includes the learned other learning module obtained by learning the learning module and another learning module having a different structure so as to maintain the function of the learned module.

  Here, the other learning module may be simpler in structure than the learned module and may be more suitable for deployment, and the output of the learned module may be used for learning of the other learning module. Data may be used. The copy of the learned module is obtained by changing the regularization method to prevent overfitting, changing the learning rate of back propagation, changing the algorithm of updating the weighting coefficient, or the like in the learning process of the learning module. Containing learned modules.

  Further, acquiring a learned module or a duplicate of the learned module means acquiring information necessary for reproducing the function of the learned module in the additional learning control device 20. For example, when the learning module includes a neural network, acquiring the learned module or a copy of the learned module means at least the number of layers of the neural network, the number of nodes related to each layer, and the weight parameter of the link connecting the nodes. It means obtaining information on the bias parameter for each node and the function form of the activation function for each node.

  The sensor 30 may be any of a physical quantity sensor that detects a physical quantity, a chemical quantity sensor that detects a chemical quantity, and an information sensor that detects information, but is not limited to these and may include any sensor. The physical quantity sensor is, for example, a camera that detects light and outputs image data or moving image data, a heart rate sensor that detects heart rate of a person and outputs heart rate data, and a blood pressure sensor that detects blood pressure of a person and outputs blood pressure data And a vital sensor such as a body temperature sensor that detects body temperature of a person and outputs body temperature data, and includes a sensor that detects any other physical quantity and outputs an electrical signal. The stoichiometry sensor includes, for example, a gas sensor, a humidity sensor, an ion sensor, and a sensor that detects an arbitrary stoichiometry and outputs an electrical signal. The information sensor includes, for example, a sensor that detects a specific pattern from statistical data, and also includes a sensor that detects any other information.

  The storage device 40 stores sensing data output by the sensor 30. The storage device 40 also stores the learned module output by the learning device 10. Although FIG. 1 illustrates the storage device 40 as a single storage unit, the storage device 40 may be configured by one or more file servers.

  In addition, in FIG. 1, although the learning apparatus 10, the additional learning control apparatus 20, and the memory | storage device 40 are each comprised separately, you may comprise these integrally. That is, all of the learning device 10, the additional learning control device 20, and the storage device 40 may be integrally configured, and two of the learning device 10, the additional learning control device 20, and the storage device 40 may be selectively integrated. It may be configured. At this time, the respective elements of the learning device 10, the additional learning control device 20, and the storage device 40 which are integrally configured are connected via an internal bus.

(Physical configuration: learning device)
The physical configuration of the learning device 10 according to a predetermined embodiment of the present invention will be described with reference to FIG. The learning device 10 includes a control unit 10a, a storage unit 10b, a communication unit 10c, an input unit 10d, and a display unit 10e. These components are mutually connected so as to be able to transmit and receive data via a bus.

  The control unit 10a includes a central processing unit (CPU) corresponding to a hardware processor, and a random access memory (RAM) corresponding to a memory. The CPU develops the program stored in the storage unit 10 b in the RAM, interprets and executes the program developed in the RAM, and the control unit 10 a functions as each unit in FIG. 4 described later.

  The type of hardware processor is not limited to the CPU. For example, using a CPU, a graphics processing unit (GPU), a field-programmable gate array (FPGA), a digital signal processor (DSP), an application specific integrated circuit (ASIC) alone or in combination as a hardware processor Can. The RAM is a storage unit in which data can be rewritten, and is formed of, for example, a semiconductor storage element. The RAM temporarily stores programs and data such as applications executed by the CPU.

  The storage unit 10 b is a non-volatile storage medium such as, for example, a hard disk drive (HDD) or a solid state drive (SDD). The storage unit 10 b stores programs and data executed by the CPU.

  The communication unit 10 c is a hardware interface that connects the learning device 10 to the communication network N.

  The input unit 10 d receives an input from the user, and is configured of, for example, a keyboard, a mouse, and a touch panel.

  The display unit 10 e visually displays the processing result of the CPU, and is configured of, for example, a Liquid Crystal Display (LCD).

  The learning device 10 may be configured, for example, by executing a learning program according to the present embodiment by a CPU of a general personal computer. The learning program may be stored and provided in a computer-readable storage medium such as a RAM or the storage unit 10b, or may be provided via a communication network N connected by the communication unit 10c. These physical configurations are examples and may not necessarily be independent configurations.

(Physical configuration: additional learning control device)
The physical configuration of the additional learning control device 20 according to the predetermined embodiment of the present invention will be described with reference to FIG. Similar to the learning device 10, the additional learning control device 20 also includes a control unit 20a including a CPU and a RAM, a storage unit 20b for storing data and the like, a communication unit 20c for connecting to the network N, and an input for receiving input from the user. It has a unit 20d, a display unit 20e and the like. These components are mutually connected so as to be able to transmit and receive data via a bus. The CPU develops the program stored in the storage unit 20b in the RAM, interprets and executes the program developed in the RAM, and the control unit 20a functions as each unit in FIG. 5 described later.

  The additional learning control device 20 may be configured, for example, by executing the additional learning control program by the CPU of a general personal computer. The additional learning control program may be provided by being stored in a computer-readable storage medium such as a RAM or the storage unit 20b, or may be provided via the communication network N connected by the communication unit 20c.

(Functional configuration: learning device)
The functional configuration of the learning device 10 according to the predetermined embodiment of the present invention will be described with reference to FIG. The learning device 10 includes a learning instruction receiving unit 101, a learning data acquisition unit 102, a learning control unit 103, a learning module 104, a learned module output unit 105, and a learned module extraction unit 106.

  The learning instruction receiving unit 101 receives a learning instruction from the user via the input unit 10 d or a learning instruction from the additional learning control device 20 via the communication unit 10 c, and the information included in the learning instruction is learning data described later. Delivered to the acquisition unit 102. In the present embodiment, the learning instruction includes learning data acquisition conditions, designation of input parameters, and the like. Among the data that can be used as learning data for causing the learning module 104 to learn, the learning data acquisition condition is a condition that is necessary to satisfy a learning instruction from the user. For example, among the data acquired by the sensor 30, an acquisition date and time may be designated. The input parameter is a factor that affects the control performance of the learned module among the information included in the learning instruction.

  The learning data acquisition unit 102 receives the learning data acquisition condition, and acquires learning data from the storage device 40 based on the received learning data acquisition condition.

  The learning control unit 103 causes the learning module 104 to learn using the learning data acquired by the learning data acquisition unit 102. The learning control unit 103 completes the learning based on the learning instruction received by the learning instruction receiving unit 101. The criteria for determining that learning has been completed may be, for example, when learning is performed using a predetermined number of learning data. Alternatively, the learning may be completed when the control performance of the learned module satisfies the learning target described later. When the learning is completed, the learning control unit 103 stores the learned module in the storage device 40. At this time, in the present embodiment, the learning control unit 103 stores the learned module in association with the learning module identifier that can uniquely identify the learned module and the learning data acquisition condition.

  Here, it is desirable that the learned modules stored in the storage device 40 be version-controlled so that the update history can be known. Version management may be performed by the learning module identifier itself or by separately provided version information.

  The learning module 104 is a module for realizing machine learning. Here, an embodiment to which a neural network is applied as an example of the learning module 104 will be described. However, the neural network is only an example of the learning module 104, and the learning device 10 may apply another configuration as the learning module 104.

  The learned module output unit 105 outputs the learned module and the learning module identifier to the outside such as the additional learning control device 20, for example.

  The learned module extraction unit 106 receives the learning module extraction condition, and acquires the learned module from the storage device 40 based on the received learning module extraction condition. In the present embodiment, the learning module extraction condition includes the learning module identifier of the learning module currently set and the extraction point. The extraction point includes information that can specify a learning module to be extracted, such as a date before performance degradation or specification of a version retroactively set from a currently set learning module. For example, the extraction point can be “before December 31, 2017”, “previous version of currently set learning module”.

(Functional configuration: additional learning control device)
The functional configuration of the additional learning control device 20 according to the predetermined embodiment of the present invention will be described with reference to FIG. The additional learning control device 20 includes a learned module reception unit 201, a learning module 202, a learning target reception unit 203, an evaluation unit 204, a determination unit 205, a learning module selection unit 206, a factor estimation unit 207, a control unit 208, and a database (DB ) 209.

  The learned module receiving unit 201 receives the learned module and the learning module identifier, and sets the received learned module as the learning module 202. In the present embodiment, the learned module receiving unit 201 receives a learned module and a learning module identifier from the learned module output unit 105 of the learning device 10, and sets the received learned module as the learning module 202. Note that the learned module receiving unit 201 may receive the learned module from the storage device 40 and set it as the learning module 202. Here, an embodiment to which a neural network is applied as an example of the learning module 202 will be described. However, the neural network is only an example of the learning module 202, and the additional learning control device 20 may apply another configuration as the learning module 202.

  The learning target receiving unit 203 receives the learning target to be achieved by the learning module 202 via the input unit 20 d, and stores the received learning target in the DB 209. In the present embodiment, the learning goal includes evaluation items and conditions. The evaluation item is an item for evaluating the learning module 202, and is an item used to determine the accuracy of output data output from the learning module, for example.

  As the condition for the evaluation item, for example, “the evaluation item is less than or equal to the reference value x”, “the evaluation item is immediately before” with respect to the evaluation item “the number of external operations for the device controlled by the control unit 208 / day”. "The evaluation item is less than or equal to the reference value y", "the evaluation item is set immediately before" for the evaluation item "energy consumption of the device controlled by the control unit 208 / month", which is smaller than the learning module set. “The evaluation item is less than or equal to the reference value z”, “for the evaluation item“ the number of times the value calculated by the learning module 202 exceeds the allowable change rate / month ”within the predetermined time”, The evaluation item may be smaller than that of the learning module set immediately before.

  The evaluation unit 204 evaluates the learning module 202 using sensing data stored in the storage device 40 to generate evaluation data, and associates the generated evaluation data with the learning module identifier received by the learned module reception unit 201. The data is stored in DB 209. In the present embodiment, the evaluation unit 204 can perform evaluation on at least an evaluation item included in the learning target, and generate evaluation data. The evaluation unit 204 can automatically perform evaluation after a certain period of time has elapsed since the learning module 202 was set, or may separately perform evaluation by accepting an evaluation instruction from a user. For example, the evaluation unit 204 may perform the evaluation in response to receiving the learning target. In addition to the evaluation items included in the learning target, the evaluation unit 204 may evaluate the evaluation items set in advance.

  The determination unit 205 determines whether the currently set learning module has achieved the learning target, using the learning target received by the learning target receiving unit 203 and the evaluation data stored in the DB 209. If it is determined that the learning goal has been achieved, the determination unit 205 ends the process.

  When it is determined that the learning goal is not achieved, the learning module selection unit 206 transmits the learning module extraction condition to the learned module extraction unit 106 of the learning device 10 as described later, and learning obtained as a response thereof The completed module is set as the learning module 202. The learned module obtained by the learning module selection unit 206 transmitting the learning module extraction condition to the learned module extraction unit 106 is evaluated whether it achieves the learning goal again before being set as the learning module 202. It is also good.

  The learning module selection unit 206 may also obtain factor improvement data by instructing a factor estimation unit 207 described later to estimate a factor that could not achieve the learning target. Thereafter, using the factor improvement data obtained from the factor estimation unit 207, the learning module selection unit 206 may instruct the learning device 10 to perform readditional learning.

  The factor estimating unit 207 estimates factors that can not achieve the learning goal when it is determined that the currently set learning module does not achieve the learning goal. In the present embodiment, in response to the instruction from the learning module selection unit 206, the factor estimation unit 207 uses the learning target and the sensing data of the storage device 40, and causes the factor estimation unit 207 to fail to achieve the learning target. presume. The sensing data of the storage device 40 is statistically processed to estimate the factor that could not achieve the learning goal.

  The control unit 208 performs control using the value calculated by the learning module 202. Although a control unit that performs control is illustrated as an example of performing control of a system using a learning module in the present embodiment, an embodiment of the present invention applies to various systems that execute processing using a learning module as described later. It can apply.

  A learning target DB 2091 and an evaluation data DB 2092 are stored in the DB 209. In the present embodiment, the learning goal DB 2091 stores evaluation items and conditions included in the learning goal accepted by the learning goal accepting unit 203. Further, the evaluation data DB 2092 stores the evaluation data generated by the evaluation unit 204 and the learning module identifier of the learning module to be evaluated in association with each other.

(Additional learning process)
First Embodiment
A first embodiment of the additional learning process performed by the additional learning control device 20 will be described along the flowchart of FIG. Embodiments of the present invention can be applied to various systems that execute processing using a learning module, and the field is not particularly limited, but in the following description, an air conditioning control system will be described as an example.

  In the first embodiment, the control unit 208 is, for example, an air conditioning control unit. The learning module 202 calculates the room temperature setting value using the current room temperature, the outside air temperature, the value of humidity, etc., the known date and time, the volume of the room, etc. output by the sensor 30 as input parameters. The control unit 208 of the additional learning control device 20 performs air conditioning control using the room temperature setting value calculated by the learning module 202. Here, an embodiment will be described in which the control unit 208 actually performs air conditioning control using the learning module 202, and the evaluation unit 204 evaluates the control result by the control unit 208. However, alternatively, the additional learning control device 20 may include a simulation unit (not shown). In an alternative embodiment, the simulation unit may perform air conditioning control using the learning module 202, and the evaluation unit 204 may evaluate the simulation results by the simulation unit.

  Further, although an embodiment in which the additional learning control device 20 includes the learning module 202 and the control unit 208 will be described here, alternatively, a control device separate from the additional learning control device 20 includes the learning module and the control unit. You may have.

  As a result of learning performed in advance by the learning control unit 103 of the learning device 10, a neural network learned using the learning data of the period from January 1, 2010 to December 31, 2012 in the storage device 40 0 and neural network 1 learned using learning data from January 1, 2010 to December 31, 2014, and learning from January 1, 2010 to December 31, 2016 It is assumed that a neural network 2 learned using data is stored. The neural network 0 is stored in the storage device 40 in association with the learning module identifier 0. The neural network 1 is stored in the storage device 40 in association with the learning module identifier 1. The neural network 2 is stored in the storage device 40 in association with the learning module identifier 2.

  It is assumed that the additional learning control device 20 sets the neural network 1 as the learning module 202 until January 31, 2017. The storage device 40 stores sensing data output by the sensor 30 as needed. That is, in the storage device 40, sensing data regarding input parameters such as the current room temperature and the outside air temperature, and the number of times of external operation with respect to the room temperature setting value, power consumption etc. during the air conditioning control using the neural network 1. The sensing data on the evaluation items of are stored. Further, as a result of the evaluation performed previously by the evaluation unit 204, evaluation data 0 of the neural network 0 and evaluation data 1 of the neural network 1 are stored in the evaluation data DB 2092 of the additional learning control device 20.

  In S601, the learned module receiving unit 201 of the additional learning control device 20 receives the neural network and the learning module identifier, and sets the received neural network as the learning module 202. In this embodiment, on February 1, 2017, the learned module reception unit 201 receives the neural network 2 and the learning module identifier 2 from the learned module output unit 105 of the learning device 10, and receives the received neural network 2 Are set as the learning module 202. From February 1, 2017, the control unit 208 performs air conditioning control using the room temperature set value calculated by the set neural network 2. Meanwhile, as described above, the storage device 40 stores the sensing data output by the sensor 30 as needed during the air conditioning control using the neural network 2.

  In S602, the learning target receiving unit 203 of the additional learning control device 20 receives the learning target to be achieved by the learning module 202 via the input unit 20d, and stores the received learning target in the learning target DB 2091 of the DB 209. In the present embodiment, on April 1, 2017, a learning goal is received from the administrator of the additional learning control apparatus 20 via the input unit 20d, and the learning goal receiving unit 203 transfers the received learning goal to the learning goal DB 2091. Remember. The learning target includes the evaluation items “the number of external operations per day for the room temperature setting value / day” and the condition “less than the learning module set immediately before”.

  In step S603, the evaluation unit 204 of the additional learning control device 20 evaluates the learning module 202 to generate evaluation data, and associates the generated evaluation data with the learning module identifier received by the learned module receiving unit 201. Are stored in the evaluation data DB 2092 of In the present embodiment, in response to the acceptance of the learning target in S602, the evaluation unit 204 uses sensing data stored in the storage device 40 during the period in which air conditioning control is performed using the neural network 2. The neural network 2 is evaluated.

  Here, the evaluation unit 204 may evaluate not only the evaluation items included in the learning target received in S602 but also other evaluation items set in advance. With such a configuration, it is possible to cope with the case where a comparison condition with a learning module that has been set previously is designated for different evaluation items in future learning goals. In the present embodiment, the evaluation unit 204 generates evaluation data 2 by evaluating not only the evaluation item “the number of external operations with respect to the room temperature setting value / day” included in the learning target but also other evaluation items set in advance. The evaluation data 2 is stored in the evaluation data DB 2092 in association with the learning module identifier 2.

  Next, in step S604, the determination unit 205 of the additional learning control device 20 uses the learning target received by the learning target receiving unit 203 in step S602 and the evaluation data stored in the evaluation data DB 2092 to perform learning that is currently set. The module determines whether the learning goal has been achieved. If it is determined that the learning goal has been achieved (S604: Yes), the additional learning control device 20 ends the process.

  In the present embodiment, the determination unit 205 includes a learning target including the evaluation item “the number of external operations per day for the room temperature setting value / day” and the condition “less than the learning module set immediately before”, evaluation data 1 and evaluation. Using the data 2, it is determined whether the neural network 2 has achieved the learning goal. As a result, it is determined that the learning target is not achieved (S604: No), and the process proceeds to S605.

  In S605, the learning module selection unit 206 of the additional learning control device 20 transmits the learning module extraction condition to the learned module extraction unit 106 of the learning device 10, and sets the learning module obtained as the response as the learning module 202. , End the process. In the present embodiment, the learning module selection unit 206 causes the learning module extraction unit 106 to set the learning module extraction condition “learning module identifier of learning module currently set: learning module identifier 2; extraction point: learning currently set The “previous version of the module” is transmitted, and the neural network 1 obtained as a response is set as the learning module 202.

Second Embodiment
In the first embodiment, an example has been described in which the neural network 2 determined as not achieving the learning target is returned to the neural network 1 that is the version before update. In the second embodiment, an example will be described in which the neural network is additionally trained when it is determined that the learning target is not achieved.

  FIG. 7 is a flowchart showing the flow of the additional learning process performed by the additional learning control device 20. In the second embodiment, the description of matters common to the first embodiment will be omitted, and only different points will be described. Since S701 to S705 in FIG. 7 are the same processes as S601 to S605 in FIG. 6, detailed descriptions of these processes will be omitted.

  In S706, the learning module selection unit 206 instructs the factor estimation unit 207 of the additional learning control device 20 to estimate the factor that could not achieve the learning goal, and the factor estimation unit 207 achieves the learning goal. Estimate the factors that could not be done. In the present embodiment, the factor estimating unit 207 estimates factors using the sensing data of the storage device 40 for the evaluation item “the number of external operations with respect to the room temperature setting value / day” included in the learning target. For example, the factor estimating unit 207 estimates factors using sensing data of a period from January 1, 2010 to December 31, 2016, which is a period of learning data used for the neural network 2.

  In the present embodiment, the factor estimating unit 207 uses, for example, a known algorithm, the evaluation item “the number of external operations to the room temperature setting value / day” included in the learning target is largely deviated from the values of other days. Find possible patterns for the day. For example, it is assumed that a pattern is found that “the number of external operations per day with respect to the room temperature setting value / day” is high in a specific period “from July 1, 2016 to July 31, 2016”. In addition, the pattern is not limited to the period, but when the values such as the outside air temperature and humidity included in the input parameter continue to differ from the predetermined values, the pattern in which the “input parameter value” is large or small You may find out. In addition, when the volume of the room included in the input parameter is changed, a pattern may be found that the "room volume" has changed.

  In S 707, the factor estimating unit 207 determines whether a pattern of outliers has been found. When the outlier pattern is not found (S 707: No), the additional learning control device 20 ends the process. On the other hand, when a pattern of outliers is found (S 707: Yes), the process proceeds to S 708, and the factor estimating unit 207 may output the found out value pattern. In the present embodiment, as shown in FIG. 8, a text 801 indicating the found outlier pattern “the number of external operations per day with respect to the room temperature setting value is large from July 1, 2016 to July 31, 2016. The factor determination screen including the performance deterioration factor radio button 802 for setting whether the pattern is the performance deterioration factor, the detail setting button 803, and the determination button 804 is output. The detail setting button will be described later.

  In step S709, the factor estimating unit 207 determines whether a user input has been received. If the user input is received, the process proceeds to S 710, the factor estimating unit 207 generates factor improvement data based on the user input, and returns the generated factor improvement data to the learning module selection unit 206. In the factor improvement data, it is not desirable to include in the learning data by being selected as the performance deterioration factor by the performance deterioration factor radio button 802 among the patterns of the outliers outputted by the factor estimation unit 207 in S708. It contains the indicated pattern.

  In the present embodiment, the administrator of the additional learning control device 20 confirms that indoor work has been performed in the period from July 1, 2016 to July 31, 2016, and the screen output in S 708 It is assumed that the performance deterioration factor radio button 802 is selected as the performance deterioration factor and the determination button 804 is pressed. As a result, the factor estimating unit 207 generates factor improvement data “except for the period from July 1, 2016 to July 31, 2016” based on the received user input, and learns the factor improvement data that has been generated. It returns to the module selection unit 206.

  In S711, the learning module selection unit 206 generates a learning instruction based on the factor improvement data, transmits the generated learning instruction to the learning instruction receiving unit 101 of the learning device 10, and ends the process. In the present embodiment, the learning module selection unit 206 sets “the learning data acquisition condition: a period from January 1, 2016 to December 31, 2016, from July 1, 2016 to July 31, 2016. A learning instruction including “excluding” is transmitted to the learning instruction receiving unit 101.

  In the present embodiment, when it is determined that the outlier pattern is found in S707, the factor improvement data is generated based on the user's input in S710, but S708 to S710 are omitted and the user's input is accepted. Instead, the factor estimation unit 207 may generate factor improvement data based on the found outlier pattern.

  By doing this, the learning device 10 can additionally learn the neural network based on the received learning instruction. In the present embodiment, the learning device 10 can additionally learn the neural network 1 by using the learning data excluding the period from July 1, 2016 to July 31, 2016.

Third Embodiment
In the second embodiment, an example has been described in which, when it is determined that the learning target is not achieved, data that is not desirable to be included in the learning data is excluded and the neural network is additionally learned. In the third embodiment, an example will be described in which, when it is determined that the learning target is not achieved, data to be learned further is added to the neural network to additionally learn the neural network.

  A third embodiment will be described with reference to FIG. Also in the third embodiment, descriptions of matters in common with the first embodiment and the second embodiment will be omitted, and only differences will be described. Since S701 to S705 in FIG. 7 are the same processes as in the first embodiment, detailed description of these processes is omitted.

  In S706, the factor estimating unit 207 estimates factors that could not achieve the learning goal. In the present embodiment, it is assumed that a pattern is found that "the number of external operations with respect to the room temperature setting value" is large in a specific time zone "8 to 9 o'clock every Saturday". In addition, the pattern is not limited to the period, but when the values such as the outside air temperature and humidity included in the input parameter continue to differ from the predetermined values, the pattern in which the “input parameter value” is large or small You may find out. In addition, when the volume of the room included in the input parameter is changed, a pattern may be found that the "room volume" has changed.

  In S 707, the factor estimating unit 207 determines whether a pattern of outliers has been found. In the present embodiment, since the pattern of outliers is found, the process proceeds to S 708, and the factor estimating unit 207 may output the found pattern of outliers. In the present embodiment, a text 801 indicating the found outlier pattern “There are many external operations to the room temperature setting value from 8 o'clock to 9 o'clock every Saturday”, the performance deterioration factor radio button 802, the detail setting button 803, and The factor determination screen including the determination button 804 is output.

  In the present embodiment, the administrator of the additional learning control device 20 preferably sets the room temperature setting value higher than that on weekdays since every Saturday a smaller number of users than weekdays use a room. In the screen output in S709, the performance degradation factor radio button 802 is selected as the non-performance degradation factor, and the detail setting button 803 is pressed. For example, when the detail setting button 803 is pressed, a detail setting screen is displayed, and the administrator may adjust the input parameter on the detail setting screen. Alternatively, the additional learning control device 20 may adjust the input parameters.

  In step S709, the factor estimating unit 207 determines whether a user input has been received. In this embodiment, it is assumed that the administrator adjusts the input parameter on the detail setting screen and then presses the determination button 804. As a result, the process advances to step S710, and the factor estimation unit 207 generates factor improvement data based on the user input, and returns the generated factor improvement data to the learning module selection unit 206. In the present embodiment, the factor improvement data includes the setting regarding the input parameter designated on the detail setting screen.

  In S711, the learning module selection unit 206 generates a learning instruction based on the factor improvement data, transmits the generated learning instruction to the learning instruction receiving unit 101 of the learning device 10, and ends the process. In the present embodiment, the learning module selection unit 206 transmits, to the learning instruction receiving unit 101, a learning instruction including the setting related to the input parameter specified on the detail setting screen.

  In the present embodiment, when it is determined that the outlier pattern is found in S707, the factor improvement data is generated based on the user's input in S710, but S708 to S710 are omitted and the user's input is accepted. Instead, the factor estimation unit 207 may generate factor improvement data based on the found outlier pattern.

  By doing this, the learning device 10 can additionally learn the neural network based on the received learning instruction. In the present embodiment, the learning device 10 uses the additional learning data when the pattern found by the additional learning control device 20 is not a performance degradation factor but a pattern to be made to further learn by the neural network. Can be further trained.

Another Embodiment
As described above, in another embodiment, as illustrated in FIG. 9, the learning system 1 includes the control device 50 including the learned module reception unit 201, the learning module 202, and the control unit 208. Good. In the present embodiment in which the learning system 1 includes the control device 50, the learning system 1 includes the learning target receiving unit 203, the evaluation unit 204, the determination unit 205, the learning module selection unit 206, the factor estimation unit 207, and the DB 209 60 is provided. As described above, the additional learning control device 20 includes the evaluation device 60.

  As shown in FIG. 10, the evaluation device 60 receives a control unit 60a including a CPU and a RAM, a storage unit 60b for storing data of the DB 209, a communication unit 60c for connecting to the network N, and an input from a user. An input unit 60d, a display unit 60e and the like are provided. These components are mutually connected so as to be able to transmit and receive data via a bus. The control unit 60a functions as each unit in FIG. 9 by the CPU developing the program stored in the storage unit in the RAM and interpreting and executing the program expanded in the RAM.

  The evaluation device 60 may be configured, for example, by executing an additional learning program by a CPU of a general personal computer. The additional learning program may be provided by being stored in a computer-readable storage medium such as a RAM or the storage unit 60b, or may be provided via the communication network N connected by the communication unit.

  A program for performing each process described in the present specification may be stored in a recording medium. If this recording medium is used, the computer can be made to function as the evaluation device 60 or the additional learning control device 20 by installing the program in the computer. Here, the recording medium storing the program may be a non-transitory recording medium. The non-transitory recording medium is not particularly limited, but may be, for example, a recording medium such as a CD-ROM.

  Some or all of the above embodiments may be described as in the following appendices, but are not limited to the following.

(Supplementary Note 1)
An evaluation device for evaluating a second learning module additionally learning a first learning module, comprising: at least one memory and at least one hardware processor connected to the memory.
The hardware processor is
Accepting a learning goal to be achieved by the second learning module;
Evaluating the second learning module with respect to at least an evaluation item included in the learning target to generate evaluation data;
Using the learning goal and the evaluation data, determine whether the second learning module has achieved the learning goal,
When it is determined that the learning goal is not achieved, a third learning module different from the second learning module is acquired based on at least the learning goal.
Evaluation device.

(Supplementary Note 2)
An evaluation method for evaluating a second learning module obtained by additionally learning a first learning module, comprising:
The second learning module receives the learning goal to be achieved by the at least one hardware processor;
The hardware processor evaluates the second learning module for at least an evaluation item included in the learning target, and generates evaluation data.
The hardware processor uses the learning goal and the evaluation data to determine whether the second learning module has achieved the learning goal.
When it is determined that the learning goal is not achieved, a third learning module different from the second learning module is acquired based on at least the learning goal.
Evaluation method.

  DESCRIPTION OF SYMBOLS 1 ... Learning system, 10 ... Learning apparatus, 10a ... Control part, 10b ... Storage part, 10c ... Communications part, 10d ... Input part, 10e ... Display part, 101 ... Learning instruction reception part, 102 ... Learning data acquisition part, 103 ... learning control unit, 104 ... learning module, 105 ... learned module output unit, 106 ... learned module extraction unit, 20 ... additional learning control device, 20a ... control unit, 20b ... storage unit, 20c ... communication unit, 20d ... Input unit 20e: Display unit 201: Learned module reception unit 202: Learning module 203: Learning target reception unit 204: Evaluation unit 205: Determination unit 206: Learning module selection unit 207: Factor estimation unit , 208: control unit, 209: DB, 2091: learning target DB, 2092: evaluation data DB, 30: sensor, 40: storage device, 50: control device, 6 ... evaluation device, 60a ... control unit, 60b ... storage unit, 60c ... communication unit, 60d ... input unit, 60e ... display unit, 801 ... text, 802 ... performance degradation factor radio button, 803 ... detailed setting button, 804 ... determination button

Claims (7)

An evaluation apparatus for evaluating a second learning module that additionally learns a first learning module, comprising:
A learning target receiving unit that receives a learning target to be achieved by the second learning module;
An evaluation unit that evaluates the second learning module with respect to at least an evaluation item included in the learning target, and generates evaluation data;
A determination unit that determines whether the second learning module has achieved the learning target, using the learning target and the evaluation data;
A learning module selection unit that acquires a third learning module different from the second learning module based on at least the learning target when it is determined that the learning target is not achieved.   The evaluation device according to claim 1, wherein the third learning module is the first learning module.   The method according to claim 1, further comprising: a factor estimating unit configured to estimate factor that the second learning module could not achieve the learning target and to generate factor improvement data related to learning data used for additional learning of the learning module. The evaluation device according to 1.   The evaluation device according to claim 3, wherein the learning module selection unit generates a learning instruction based on the factor improvement data.   The evaluation device according to claim 3, wherein the factor estimation unit outputs the estimated factor, receives a user input related to the estimated factor, and generates the factor improvement data based on the user input. It is an evaluation method which evaluates the 2nd learning module which made the 1st learning module additionally learn, and a computer provided with a control part is
Receiving a learning goal to be achieved by the second learning module;
Evaluating the second learning module with respect to at least an evaluation item included in the learning target to generate evaluation data;
Determining whether the second learning module has achieved the learning goal using the learning goal and the evaluation data;
Obtaining a third learning module different from the second learning module based on at least the learning target, when it is determined that the learning target is not achieved. A program for causing a computer to evaluate a second learning module that additionally learns a first learning module,
A process of receiving a learning goal to be achieved by the second learning module;
A process of evaluating the second learning module for at least an evaluation item included in the learning target, and generating evaluation data;
A process of determining whether the second learning module has achieved the learning goal using the learning goal and the evaluation data;
A program for executing a process of acquiring a third learning module different from the second learning module when it is determined that the learning goal is not achieved.

Images