Service delivery system and method for producing the service delivery system

Abstract

Personalizing the services offered enables the delivery of services that are tailored to the user's needs. [Solution] When a user is engaging in a conversation with a dialogue model for general conversation (S235-S246, NO in S247, NO in S249), the system switches to the task processing model to engage in conversation depending on the content of the conversation (NO in S247, S248), enabling the use of the task processing model's specialized knowledge in a particular field. Furthermore, through conversations with individual users by the dialogue model, the dialogue model is personalized to suit the individual user (YES in S245, S246).

Description

Technical Field

[0001] The present invention relates to A service delivery system that enables the provision of services using machine learning models, and a method for producing such a service delivery system. .

Background Art

[0002] The underlying technology included the following: By determining the priority of information based on the user's unique basic selection preferences, corrected using the user's actual viewing history data, an information retrieval method and apparatus could be realized that could easily search for the information desired by the user from a vast number of programs. The user model consists of a user-specific initial model and a learned model corrected using the user's information selection history. The learned model is used when searching for desired information. The initial model is created based on general user selection preference data determined from user attributes and basic information selection preference data representing the information selection preferences of a specific user. (For example, Patent Document 1).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0006] The present invention The goal is to make personalized models available for specific individuals. is to achieve. Specific examples of means for solving the problems and their effects

[0007] Next, an example of the correspondence between various means for solving the problems and embodiments is shown in parentheses below.

[0008] The invention according to claim 1 is a service providing system that enables the provision of a service using a machine learning model (for example, after optimizing (personalizing) a general model obtained as a result of machine learning for each user and providing a personalized service to each user, etc.). A standard model created by machine learning using training data collected from a large number of people (for example, classifying a large number of users by profession such as patent attorney, lawyer, physicist, chemist, etc., calculating the average for each profession, and creating initial weights for each piece of knowledge and a knowledge utilization function appropriate to each initial target. An example of a knowledge utilization function G(x) is as follows: Each piece of knowledge is represented as n1, n2, n3, ..., and each Let x1k, x2k, x3k, ... be the number of positive user responses to knowledge, such as "satisfied," and x1h, x2h, x3h, ... be the number of negative responses, such as "not satisfied." Let w1, w2, w3, ... be the initial weights of each piece of knowledge. Also, for convenience, let S(x) represent the sigmoid function 1 / (1+ex). Then, G(x) = w1n1S(x1k-x1h) + w2n2S(x2k-x2h) + w3n3S(x3k-x3h) )+.... In other words, knowledge in which the number of positive responses exceeds the number of negative responses is considered useful knowledge, and it is possible to control the system so that knowledge with larger coefficients (weights w1, w2, w3, ...) for each piece of knowledge n1, n2, n3, ... is used preferentially.) A personalization means (for example, S255~S258, S270~S273, the judgment in S270 is, for example, the AI ​​server 9 determines the need to change the weights of the knowledge it is using when the user says predetermined set phrases such as "satisfied" or "not satisfied" in response to the response from the AI ​​server 9. Alternatively, the judgment may be based on the user's normal conversation, such as "Thank you for the great answer," instead of or in addition to the above set phrases), The system includes a means for making the personalized model generated by the personalization means available for use as a substitute for the specific person, The personalization model is capable of providing the user with responses that reflect the work knowledge of the specific person (e.g., S269), The aforementioned means for making available the RecordThis function enables responses to the server using natural language processing (for example, the artificial intelligence server 9 is equipped with a natural language processing engine, which incorporates not only the text mining processes of "morphological analysis" and "syntactic analysis (processing to determine dependency relationships between clauses)," but also two technologies: "contextual analysis," which is necessary to grasp the structure of a sentence, and "semantic analysis," which is necessary to understand the intent of a sentence). 。 ) has.

[0010] The invention described in claim 2 is a method for producing a service provision system that enables the provision of services using machine learning models (for example, a general model obtained as a result of machine learning is optimized (personalized) for each user, and then personalized services are provided to each user), A standard model created by machine learning using training data collected from a large number of people (for example, classifying a large number of users by profession such as patent attorney, lawyer, physicist, chemist, etc., calculating the average for each profession, and creating initial weights for each piece of knowledge and a knowledge utilization function appropriate to each initial target. An example of a knowledge utilization function G(x) is as follows: Each piece of knowledge is represented as n1, n2, n3, ..., and each Let x1k, x2k, x3k, ... be the number of positive user responses to knowledge, such as "satisfied," and x1h, x2h, x3h, ... be the number of negative responses, such as "not satisfied." Let w1, w2, w3, ... be the initial weights of each piece of knowledge. Also, for convenience, let S(x) represent the sigmoid function 1 / (1+ex). Then, G(x) = w1n1S(x1k-x1h) + w2n2S(x2k-x2h) + w3n3S(x3k-x3h) )+.... In other words, knowledge in which the number of positive responses exceeds the number of negative responses is considered useful knowledge, and it is possible to control the system so that knowledge with larger coefficients (weights w1, w2, w3, ...) for each piece of knowledge n1, n2, n3, ... is used preferentially.) The first step is to generate a personalized model that is personalized for a specific person and reflects that person's work knowledge (for example, S255~S258, S270~S273, the judgment in S270 is, for example, the need to change the weights of the knowledge used by the AI ​​server 9 when the user says predetermined set phrases such as "satisfied" or "not satisfied" in response to the response from the AI ​​server 9. Alternatively, the judgment may be based on the user's normal conversation such as "Thank you for the great answer" instead of or in addition to the above set phrases), The process includes a second step of making the personalized model generated in the first step available for use as a substitute for the specific person (for example, a user can register the personalized weights of their knowledge in the artificial intelligence DB 17 and make them widely available to the public. For example, when efficiently training a subordinate at work to turn them into a full-fledged expert, if the supervisor transfers the personalized weights of the knowledge they have accumulated so far to that subordinate, the artificial intelligence server 9 will respond to the subordinate's questions and consultations on their behalf, and the artificial intelligence server 9 will take over the training of the subordinate, etc.), The personalization model is capable of providing the user with responses that reflect the work knowledge of the specific person (e.g., S269), The second step described above is the previous Record This enables responses to the server using natural language processing (for example, the artificial intelligence server 9 is equipped with a natural language processing engine, which incorporates not only the text mining processes of "morphological analysis" and "syntactic analysis (processing to determine dependency relationships between clauses)," but also two technologies: "contextual analysis," which is necessary to grasp the structure of a sentence, and "semantic analysis," which is necessary to understand the intent of a sentence). 。 ). [Brief explanation of the drawing]

[0012] [Figure 1] This system diagram shows the overall configuration of the IoT device intermediation system and the service provision system utilizing machine learning. [Figure 2] (a) is a block diagram showing the control circuit of an IoT device, and (b) is a block diagram showing the control circuit of a mobile communication device. [Figure 3] This is a block diagram showing the control circuits of PCs and servers. [Figure 4] This is a flowchart showing the main routine program between a mobile communication device and an artificial intelligence server. [Figure 5] This flowchart shows a subroutine program for IoT processing, as well as flowcharts for IoT devices (sensors) and IoT servers. [Figure 6]It is a flowchart showing a subroutine program for IoT processing, and flowcharts of an IoT device (actuator) and an IoT server. [Figure 7] It is a flowchart showing a subroutine program for IoT processing, and flowcharts of an IoT device (actuator) and an IoT server. [Figure 8] It is a flowchart showing a subroutine program for IoT processing, and flowcharts of a wireless sensor network and an IoT server. [Figure 9] It is an explanatory diagram showing an overview of a service providing system using machine learning. [Figure 10] It is a functional block diagram showing a service providing system using machine learning. [Figure 11] It is a flowchart showing a subroutine program for external display processing and display processing, and a flowchart of a digital menu board. [Figure 12] (a) is a flowchart showing a subroutine program for expert data processing, and (b) is a flowchart showing a subroutine program for maintenance data processing and a flowchart of a maintenance specialist's PC. [Figure 13] (a) is a flowchart showing a subroutine program for around data terminal processing and around data server processing, and (b) is a flowchart showing a subroutine program for personal service terminal processing and personal service server processing. [Figure 14] It is a flowchart showing a subroutine program for user learning terminal processing and user learning server processing. [Figure 15] It is an explanatory diagram explaining a general model creation method. [Figure 16] It is a flowchart showing a subroutine program for learning service terminal processing and learning service server processing. [Figure 17]It is a flowchart showing a subroutine program for muscle training data processing and a flowchart of a PC of a muscle training specialist. [Figure 18] (a) is a flowchart showing a subroutine program for muscle training terminal processing and muscle training server processing, and (b) is a flowchart showing a subroutine program for muscle training service terminal processing and muscle training service server processing. [Figure 19] (a) is a diagram showing a specific example of a dialogue model stored in an artificial intelligence DB, and (b) is a diagram showing a specific example of personalized data for dialogue stored in a user DB. [Figure 20] It is a flowchart showing a subroutine program for dialogue terminal processing and dialogue server processing. [Figure 21] (a) is a diagram showing a specific example of a task processing model stored in an artificial intelligence DB, and (b) is a diagram showing a specific example of personalized data for task processing stored in a user DB. [Figure 22] It is a flowchart showing a subroutine program for task terminal processing and task server processing. [Figure 23] It is a flowchart showing a subroutine program for task terminal processing and task server processing.

Mode for Carrying Out the Invention

[0013] [IoT Device Mediation System] First, the IoT device mediation system will be described based on FIGS. 1 to 8. IoT is an abbreviation for Internet of Things, in which things are networked by the Internet protocol and things transmit signals onto the Internet by themselves. The IoT device mediation system is a system for mediating communication between a group 2 of IoT devices (various sensors and actuators for IoT, etc.) that do not have a function for connecting to and communicating via the Internet and a user's mobile communication device 3, and connecting the IoT device group 2 to the Internet 1 via the mobile communication device 3 so that they can communicate.

[0014] Referring to the overall system in Figure 1, the Internet 1 is connected to a mobile communication device 3, such as the user's wearable computer, a robot 6 in the user's home, a personal computer (hereinafter referred to as "PC") 7 in the user's home, an IoT server 8, an artificial intelligence server 9, PCs 10 of various specialized companies, and an SNS server 11, all of which are configured to communicate with each other. The mobile communication device 3 has functions for communicating with IoT device group 2, a wireless sensor network 4, and a display group 5 such as a digital menu board. As for communication methods, for example, Wi-Fi (registered trademark), Bluetooth (registered trademark), Wi-Fi Direct (registered trademark), Zigbee (registered trademark), Z-wave (registered trademark), and Ant+ (registered trademark) are envisioned. In addition, the OS will support iOS, Android (registered trademark), Linux (registered trademark), TIZEN (registered trademark), and other real-time OSs. IEEE 802.15.4 is adopted as the wireless standard for transmission and reception. IPv6 (Internet Protocol Version 6) is adopted as the Internet Protocol.

[0015] A wireless sensor network 4 is a wireless network that allows multiple sensor-equipped wireless terminals to be scattered throughout a space and to cooperate in collecting environmental and physical information. For example, sensor devices can be created using energy harvesting, M2M, or batteries, and pressure sensors or gauge sensors can be used to constantly monitor things like metal fatigue degradation, and notifications can be sent when changes occur. They are mainly installed in structures such as bridges and tunnels. Generally, they include multiple sensor nodes and a gateway sensor node. These nodes usually consist of one or more sensors, a wireless chip, a microprocessor, and a power supply (battery, etc.). The hardware configuration of the node's control circuit is the same as shown in Figure 2(a). Wireless sensor networks usually have ad hoc functionality and a routing function (routing algorithm) for sending data from each node to a central node. In other words, they have the function to autonomously reconstruct an alternative communication path if there is a communication failure between nodes. Because the nodes cooperate as a group, there is also an element of distributed processing. In addition, they have the function to operate for long periods without receiving power from an external source, and for that purpose they have power saving functions or self-generation functions. In this embodiment, the wireless sensor network 4 is a type of IoT device group 2.

[0016] The IoT server 8 can write to and read data from the IoT device DB 15, which is a database (hereinafter referred to as "DB") for IoT devices. The artificial intelligence server 9 can write to and read data from the artificial intelligence DB 17, the user DB 12, and the learning DB 60. The SNS server can write to and read data from the SNS DB 13.

[0017] Next, with reference to Figure 2(a), the hardware configuration of the control circuit of IoT device 2 will be described. The IoT device group 2 consists of sensors and actuators installed throughout the globe, and includes a CPU (Central Processing Unit) 18 for controlling the whole system, a ROM (Read Only Member) 20 for storing programs to execute various functions, a RAM (Random Access Memory) 19 which is the work area of ​​the CPU 18, and a wireless communication interface unit 22 that uses, for example, Wi-Fi, Bluetooth, Wi-Fi Direct, Zigbee, Zwave, Ant+, etc. If IoT device 2 is a sensor, it includes a sensor unit 14. On the other hand, if IoT device 2 is an actuator, an actuator unit 21 is provided instead of or in addition to the sensor unit 14. The CPU 18, RAM 19, ROM 20, wireless communication interface unit 22, sensor unit 14, and actuator unit 21 are connected in a way that allows for signal exchange.

[0018] Next, referring to Figure 2(b), the hardware configuration of the control circuit of an example wearable computer in the mobile communication device 3 will be described. This wearable computer 3 is representative of smart glasses, and consists of a CPU 23, RAM 24 and ROM 25, and EEPROM (Electronically Erasable and Programmable Read Only Memory) 26 connected by a bus 27. The functions of the CPU 23, RAM 24 and ROM 25 are basically the same as those described above for the IoT device 2. The EEPROM 26 stores application programs and the like downloaded via the internet 1.

[0019] Bus 27 has various devices connected to it via the interface unit 28. For example, a digital video camera input unit 29 that captures images of the user's surroundings (such as images in the user's line of sight), a display unit 30 that overlays information on the lenses of smart glasses, a wireless communication processing unit 31 that communicates data with servers etc. via the Internet 1 by wirelessly communicating with a base station, an input operation unit 32 that allows the user to have the CPU 23 of the wearable computer 3 execute a desired function, an audio output unit 33 and an audio input unit 34 for the user to make voice calls, a location information acquisition unit 35 that obtains the current location based on GPS information from satellites, radio waves from base stations and wireless radio waves from wireless LAN access points, a wireless communication interface unit 36 ​​that communicates with IoT device group 2 using Wi-Fi, Bluetooth, Wi-Fi Direct, Zigbee, Zwave, Ant+, etc., a gaze position detection unit 37 that detects the user's line of sight and identifies the position of the user's gaze in the images captured by the digital video camera input unit 29, and various sensors 38 are all connected to the interface unit 28.

[0020] The input operation unit 32 accepts not only manual operations by the user, but also instructions from the user through gestures and specific winks. Voice instructions from the user are accepted by the voice input unit 34. The various sensors 38 are temperature, humidity, illuminance, and ultraviolet light sensors attached to the wearable computer 3. Note that the hardware configuration of the robot's control circuit is obtained by removing the various sensors 38 and adding a movement drive circuit to the hardware configuration shown in Figure 2(a).

[0021] Next, the hardware configuration of the control circuits of the PC and various servers 7-11 will be explained based on Figure 3. As described above, the CPU 40, RAM 41, and ROM 42 are connected by a bus 43. The interface unit 44 to which the bus 43 is connected is connected to a communication unit 45 for the Internet 1, a display unit 46 for displaying images and information to the operator, and an input operation unit 47 for receiving operations from the operator.

[0022] Next, based on Figure 4, the flowchart of the main program of the control processing executed by the CPU 23 of the mobile communication device 3 and the CPU 40 of the artificial intelligence server 9 will be explained. The mobile communication device 3 performs IoT processing in S1. This is the process in which the mobile communication device 3 acts as an intermediary to connect the IoT device group 2 and the internet 1, transmits detection data from the IoT device (sensor) 2 to the IoT server 8, and transmits command signals (actuator control signals) from the IoT server 8 to the IoT device (actuator) 2.

[0023] Next, the mobile communication device 3 performs external display processing via S2, and the artificial intelligence server 9 performs display processing via S3. This is the process of displaying videos, still images, etc., that the user has uploaded to the internet on the display 5, such as a digital menu board. Next, the mobile communication device 3 performs around data terminal processing via S4, and the artificial intelligence server 9 performs around data server processing via S5. This is the process of collecting video and audio from the user's surroundings using the mobile communication device 3, such as a wearable computer, and using this information for machine learning on the artificial intelligence server 9.

[0024] Next, the mobile communication device 3 performs personal service terminal processing via S7, and the artificial intelligence server 9 performs personal service server processing via S8. This is the process for providing personalized services to the user using the results of machine learning by the artificial intelligence. Next, the mobile communication device 3 performs dialogue terminal processing via S9, and the artificial intelligence server 9 performs dialogue server processing via S10. This is the process for the user and the artificial intelligence to engage in various conversations. Next, the mobile communication device 3 performs task terminal processing via S11, and the artificial intelligence server 9 performs task server processing via S12. This is the process for the artificial intelligence, having completed machine learning, to respond to the user's work-related questions and consultations.

[0025] Next, based on Figures 5 to 8, we will explain the flowchart of the subroutine program for the IoT processing shown in S1 above. First, referring to Figure 5, we will explain the processing performed by the IoT device (sensor) 2, the mobile communication device 3, and the IoT server 8.

[0026] The CPU 18 of the IoT device (sensor) 2 performs the process of storing the detection data from the sensor unit 14 in the RAM 19 in S20, and determines whether or not a transmission command signal has been received in S21. If it has not been received, it returns to S20, and the loop of S20→21→S20 is repeated.

[0027] Meanwhile, in S22, the CPU 23 of the mobile communication device 3 transmits location information based on GPS information, radio waves from base stations, and radio waves from wireless LAN access points to the IoT server 8. The IoT server 8, having received this information in S24, searches the IoT device DB 15 in S25 and determines in S26 whether there is an IoT device 2 to be received near the mobile communication device 3 that transmitted the location information. Ideally, the IoT server 8 should periodically collect detection data from IoT devices (sensors) 2, and searches the IoT device DB 15 to identify IoT devices (sensors) 2 for which it is time for periodic reception. If there is an IoT device 2 to be received, S26 determines YES and control proceeds to S27, sending a command signal for reception to the mobile communication device 3.

[0028] On the other hand, if there is no IoT device 2 to receive the signal from, S26 determines NO and the control proceeds to S38 to determine whether there is an IoT device (actuator) 2 nearby to which a command signal (actuator control signal) should be sent. If there is, the control proceeds to S39 and sends the command signal for transmission and transmission information (information for controlling the actuator unit 21) to the mobile communication device 3. Upon receiving the command signal from S27 or S39, the mobile communication device 3 determines YES in S23 and the control proceeds to S28 to determine whether the received command signal is for receiving. If the command signal is for transmission that was sent in S38, the control proceeds to S47, but if the command signal is for receiving that was sent in S27, the control proceeds to S29 and sends the transmission command signal to a nearby IoT device (sensor) 2.

[0029] Upon receiving this in S21, the mobile communication device 3 sends the detection data and device ID stored in RAM back to the mobile communication device 3 via S30, and returns to S20. Upon receiving this again in S31, the mobile communication device 3 sends the received data to the IoT server 8 via S32. Upon receiving this again in S33, the mobile communication device 3 updates the data stored in the IoT device DB 15, specifically the data for the IoT device corresponding to the received device ID, with the received data via S34.

[0030] Next, the IoT server 8 sends points, an example of a reward, to the mobile communication device 3 in S35, and then control transitions to S61. The mobile communication device 3, having received this in S36, adds and updates the points stored in the EEPROM 26 (S37), and then control transitions to S60. These points stored in the EEPROM 26 can be used to receive various services from the company operating the IoT server 8. For example, points can be used to pay highway tolls or purchase JR or private railway tickets and passes. It may also be possible to exchange points for cash. This provides an incentive for users to use IoT devices as intermediaries.

[0031] On the other hand, if the result in S38 is NO, the control proceeds to S40 to determine whether there are any sensors 38 on the mobile communication device 3 that the user wants to receive data from. In other words, the control determines whether there are any sensors 38 on the mobile communication device 3 located at the location information received in S24 that the user wants to receive detection data for that geographical location from. If there are none, the control proceeds to S24; otherwise, the control proceeds to S62.

[0032] Next, referring to Figure 6, the processing for the IoT device (actuator) 2 will be explained. The IoT device (actuator) 2 performs operation control processing based on the transmission information sent from the IoT server 8 and stored in the RAM 19 (S45). Next, in S46, it is determined whether or not a received command signal has been received from the mobile communication device 3. If it has not been received, the control returns to S45, and the loop of S45→S46→S45 is repeated. In this state, the mobile communication device 3, having received the transmission command signal and transmission information, sends the received command signal and transmission information to the IoT device (actuator) 2 in S47. Upon receiving it, the IoT device (actuator) 2 makes a YES determination in S46, and the control proceeds to S48, where the received transmission information (information for controlling the actuator unit 21) is stored in the RAM 19, and in S49, a reception completion signal and device ID are sent to the mobile communication device 3.

[0033] Upon receiving this in S50, the mobile communication device 3 sends a completion signal to the IoT server 8 in S51. The IoT server 8, upon receiving this in S52, searches the IoT device database based on the device ID in S53 and updates the data corresponding to the device ID. For example, it updates it to something like "Transmission information sent at ○:○:△×". Then, after sending the points to the mobile communication device 3 in S54, control transitions to S61. Upon receiving this in S55, the mobile communication device 3 adds and updates the points stored in the EEPROM 26 (S56). After that, control transitions to S60.

[0034] In S60, it is determined whether the IoT server 8 has received the command signal and data identification signal for mobile reception. If they have not been received, control proceeds to S84. On the other hand, if the IoT server 8 determines in S24 that there are various sensors 38 of the mobile communication device 3 that it wants to receive data from at the location where it received the data, it determines YES in S40 and transmits the command signal and data identification signal for mobile reception in S62. Upon receiving this, the mobile communication device 3 determines YES in S60 and proceeds to S63, where it determines whether the data identified by the data identification signal (temperature, humidity, illuminance, ultraviolet light, etc.) requires detection time. In the case of illuminance and ultraviolet light, detection can be done instantaneously and does not require detection time, so it determines NO in S63 and proceeds to S64, where it processes the identified data using the corresponding sensor 38 and transmits the detected data to the IoT server 8 (S65).

[0035] The IoT server 8, upon receiving the data in S66, searches the IoT device DB 15 in S67 and updates it with the newly received detection data. Then, in S68, it transmits the points to the mobile communication device 3. The mobile communication device 3, upon receiving the data in S69, adds to and updates the points stored in the EEPROM 26 (S70). After that, control proceeds to S84.

[0036] On the other hand, if the data identified by the data identification signal, such as temperature or humidity, requires a detection time, a YES determination is made in S63, and control proceeds to S71, where the display unit 30 notifies the user of the detection time (e.g., 30 seconds). Upon seeing this, the user, if willing to cooperate in transmitting the sensor detection data, stays at their current location for 30 seconds and waits for the detection to be completed. If they do not wish to cooperate, they move on without staying. After the notification in S71, S72 performs detection processing on the corresponding sensor 38 for the identified data. S73 determines whether the detection is complete or not. If it is not yet complete, S74 determines whether the user is still within the detection area. If they are still within the area, the loop returns to S72 and is repeated. If the user's mobile communication device 3 moves outside the detection area before the detection is complete, control proceeds to S84. However, if the user's mobile communication device 3 remains within the detection area until the detection is complete, control proceeds to S75, transmits the detection data to the IoT server 8, and then proceeds to S69.

[0037] The IoT server 8, upon receiving the data in S66, searches the IoT device DB 15 in S67 and updates it with the newly received detection data. Then, in S68, it transmits the points to the mobile communication device 3. The mobile communication device 3, upon receiving the data in S69, adds to and updates the points stored in the EEPROM 26 (S70). After that, control proceeds to S84.

[0038] In the control shown in Figures 5 to 7, a determination means (e.g., S26, S38, S49, etc.) for determining whether or not there are IoT devices that need to exchange information with the IoT device group 2 is provided on the IoT server 8 side. However, this determination means may also be provided on the IoT device group 2 side. This reduces the control burden on the IoT server 8 side.

[0039] Next, the processing for the wireless sensor network 4 will be explained with reference to Figure 8. The wireless sensor network 4 collects detection data from each sensor via S80 and stores it, associating it with each sensor ID. Next, S81 determines whether there is any data in the detection data that should be transmitted urgently. If there is no data that should be transmitted urgently, control proceeds to S82, which determines whether it is time for a periodic transmission (for example, every 24 hours). If it is not time for a periodic transmission, control returns to S80, and the loop of S80→S81→S82 is repeated.

[0040] During this loop, if an obvious anomaly that can be determined by the wireless sensor network 4 is found in each detection data, and S81 determines that there is something that should be transmitted urgently, the control proceeds to S83. Also, if S82 determines that it is time for a regular transmission, the control proceeds to S83.

[0041] In S83, the process of sending a transmission signal is performed. Next, in S86, it is determined whether or not a response signal has been received from the mobile communication device 3. If it has not been received, the process returns to S80, and the loop of S80~S83→S86→S80 is repeated.

[0042] In this state, when a user with a mobile communication device 3 passes near the wireless sensor network 4, the mobile communication device 3 receives a transmission signal via S83, makes a YES determination via S84, and proceeds to control S85. The mobile communication device 3 sends a response signal to the wireless sensor network 4 via S85, and upon receiving it, the wireless sensor network 4 makes a YES determination via S86, and proceeds to control S87. The wireless sensor network 4 sends the collected detection data to the mobile communication device 3 for each sensor ID via S80 (S87). The mobile communication device 3, upon receiving this data via S89, sends the received detection data and sensor ID along with the transmission signal to the IoT server 8 via S90.

[0043] Then, the IoT server 8 makes a YES determination in S61, and control proceeds to S91. If there is emergency detection data among the received detection data, it reports the anomaly along with the sensor ID, and then in S92, it searches the wireless sensor DB16 based on the sensor ID of the detection data and updates it with the newly received detection data. Next, in S93, it sends the points to the mobile communication device 3, and control returns to S24. The mobile communication device 3, which receives this in S94, adds and updates the points stored in the EEPROM26 (S95), then returns, and control proceeds to S2.

[0044] In the control shown in Figure 8, a determination means (e.g., S82) for determining whether or not there is an IoT device that needs to exchange information with the IoT server 8 is provided on the wireless sensor network 4 side. However, this determination means may also be provided on the IoT server 8 side.

[0045] In the IoT device intermediation system described above, the user's mobile communication device (for example, a wearable computer such as smart glasses, a smartphone, or a vehicle such as a car equipped with communication functions) acts as an intermediary between the IoT device group 2 and the internet 1. Therefore, the IoT device group 2 does not necessarily need to have an internet connection function, which makes it possible to keep costs down. Moreover, since the user is given a reward (for example, points) for performing the above-mentioned intermediation using their mobile communication device, it is possible to provide an incentive for the user to perform the intermediation and promote the user's intermediary activity.

[0046] Furthermore, because the wearable computer 3 acts as an intermediary computer (e.g., an IoT server 8 or a network cloud) for the IoT device group 2, a determination means (e.g., S26, S38, S49, etc.) is provided to determine whether or not there are IoT devices (including various sensors 38 of the mobile communication device 3) that need to exchange information. As a result, the IoT device group 2 does not need to make the above determination, reducing the processing burden on the IoT device group 2 and further lowering costs. Moreover, the IoT server 8, which has the determination means, can identify the IoT device group 2 that needs to exchange information (e.g., S26, S38, S49, etc.), and notify the wearable computer 3 of the location information of the identified IoT devices 2, thereby informing the user and encouraging the user to perform the above-mentioned intermediary action.

[0047] Furthermore, when using the various sensors 38 installed on the wearable computer 3 to collect detection data from the various sensors 38 at a desired location and upload it to the internet, if it takes a certain amount of time to detect something, the wearable computer 3 will inform the user of that detection time (for example, S71), thus encouraging the user to cooperate in collecting detection data that takes time.

[0048] [A service delivery system utilizing machine learning] Next, we will explain a service delivery system that utilizes machine learning based on Figures 9 to 23. A major challenge in the practical application of machine learning is how to collect large amounts of training data inexpensively. This system aims to create a mechanism where many users proactively provide training data to the system by offering services that utilize the results of machine learning. In particular, as shown in Figure 9, the training data (data from one's surroundings, etc.) provided by each user to the system is also used for personalization, and the general model obtained from the machine learning results is optimized (personalized) for each user, and then a personalized service is provided to each user. This prevents the inconvenience of users enjoying the service while leaving the provision of training data to others.

[0049] First, let's explain the system outline with reference to Figure 9. Each user 70 transmits surrounding data (such as images, gaze position, voice, and GPS location data) collected by a mobile communication device 3, such as a wearable computer, to the machine learning 72, where it is used as training data for the machine learning 72, and this data is also used for personalization 74.

[0050] For example, each user 70 sends information about their learning status at school or home to the system, which is then modeled using machine learning (reinforcement learning or regression) 72 to create a general model that minimizes the total time required for repeated review by a human. In this process, elements that differ from user to user (such as memorization ability) are treated as constants to generate the general model. The data sent from each user 70 to the system is also used for personalization 74, where elements that differ from user to user (such as memorization ability) are calculated for each user and substituted into the constant parts of the general model to generate a personalized model for each user. Using this personalized model, personalized services are provided to each user, such as suggesting the next review date and providing instruction according to a review plan that minimizes the total review time.

[0051] Furthermore, training data is provided not only by users but also by various specialized companies 71. For example, a sports gym specialist 71 sends aggregated results of muscle training (hereinafter referred to as "muscle training") from many members to the system. Based on this aggregated training data, a model is created using machine learning (reinforcement learning or regression) 72 to create a general model that maximizes the effect of muscle training performed repeatedly by humans. In this process, elements that differ from user to user (such as initial load and load increase coefficient for barbells, etc.) are used as constants to generate a general model. In addition, aggregated muscle training results from each user 70 are sent to the system and used for personalization 74. Elements that differ from user to user (such as initial load and load increase coefficient) are calculated for each user and substituted into the constant parts of the general model to generate a personalized model for each user.

[0052] Using this personalization model, personalized services are provided to each user, such as suggesting the timing and intensity of the next strength training session and providing guidance according to a strength training plan that maximizes the effectiveness of the workout. At the same time, advertisements from 71 professional companies (such as gyms) that provided the learning data consisting of aggregated strength training results are displayed along with the strength training schedule. This motivates various professional companies to provide learning data.

[0053] Figure 10 shows a more detailed representation of the overall system described in Figure 9. Referring to Figure 10, specialized data is provided to the artificial intelligence server 9 from PCs 10 of various specialized companies. In addition, data collected by mobile communication devices 3 such as wearable computers of numerous users (around data such as images viewed by users 70, gaze position, audio, and GPS location data) is also provided to the artificial intelligence server 9. The artificial intelligence server 9 converts this provided data into numerical and label sets 51 to generate training data. Deep learning is used to generate this training data by converting raw data into numerical and label sets. Furthermore, "heterogeneous ensemble learning" is used to handle raw data containing multiple regularities. This is a method that generates a model suitable for each pattern when multiple regularities exist in the collected data.

[0054] For example, when collecting time-series data on a building's power consumption and creating a model to predict power consumption, the building's power consumption patterns change depending on the day of the week and time of day. Traditionally, humans would use their expertise to categorize the data and create a model suitable for each case, determining when the patterns would switch. In this "heterogeneous ensemble learning" method, machine learning is used to identify the pattern changes themselves and generate a model for each pattern. When new data is generated, the system checks its fit with the generated model, and if the model does not fit well, it restarts from determining the pattern changes. In other words, the machine repeatedly performs "pattern categorization" and "model creation" to create the optimal model for each pattern.

[0055] Furthermore, in cases where data cannot be converted into numerical and label sets even using deep learning, artificial intelligence companies will manually convert the data into numerical and label sets.

[0056] This artificial intelligence server 9 uses a conventional von Neumann architecture computer, but a neural network processor (NNP) may also be used. The NNP chip is equipped with numerous "artificial neurons" modeled after real neurons, and each neuron interacts with the others in a network.

[0057] Alternatively, a quantum computer employing the "quantum annealing method" may be used. This can significantly reduce the time required for optimization calculations in machine learning. Note that "artificial intelligence" is a broad concept that includes software agents.

[0058] The learning data, which has been converted into a set of numerical labels, is sent to the learning algorithm 52 and modeled 53. Various learning algorithms 52 are available, such as supervised learning methods like regression and classification, unsupervised learning methods like model estimation and data mining, and intermediate methods like reinforcement learning and deep learning. Modeling using the learning algorithm 52 generates a general model by treating elements that differ from user to user (such as individual ability differences and preferences) as constants. This generated general model is stored in the artificial intelligence database 17. Figure 10 shows specific examples of stored general models, such as a building maintenance and inspection model, a memorization learning model, a muscle training model, a dialogue model, and a task processing model. The aforementioned learning data is also stored in the learning database 60.

[0059] The memorization learning model Ti=T0·i is a general model that minimizes the total time spent on repeated review by humans. i represents the number of repetitions, T0 is the period from initial learning to just before forgetting (specifically, the time from initial learning until memory retention reaches 70%), and Ti represents the time from the number of repetitions (i-1) to i. In this general model, T0 is a user-specific element (individual differences in ability such as memory) and is represented as a constant. The method for generating this memorization learning model Ti=T0·i will be explained in detail later.

[0060] The muscle training model Fi = F0 + (i-1) / a is a general model that maximizes the effect of muscle training performed repeatedly by a human, where i is the number of repetitions, F0 is the load in the first muscle training session (e.g., barbell weight), and a is the load increase coefficient that increases with each repetition of the muscle training.

[0061] The dialogue model is a general model used when a user interacts with artificial intelligence, and consists of dialogue templates for initial targets such as male AI systems, female moe characters, and famous celebrities, as well as change functions for emotions such as joy, anger, sadness, and happiness. This will be explained in detail based on Figure 19.

[0062] The task processing model is a general model used by users when processing tasks such as work, and consists of knowledge utilization functions for initial targets such as patent attorneys, lawyers, and famous scientists. This will be explained in detail based on Figure 21.

[0063] Around data sent from the user's mobile communication device 3 is classified by the artificial intelligence server 9 for various models 54, and personalized data for each model is created 55. This personalized data is stored in the user DB 12. Figure 10 shows specific examples of various models, such as those for memorization learning, muscle training, dialogue, and task processing.

[0064] User DB12 stores personalized data and surrounding data associated with each user ID. The personalized data for memorization is TK, which represents the period from initial learning to just before forgetting (specifically, the time from initial learning until memory retention reaches 70%). Human memory decreases according to the forgetting curve of psychologist Hermann Ebbinghaus. TK is defined as the time it takes for memory retention to reach 70% according to this forgetting curve. This TK is the actual time it takes for a user with user ID "1" to reach 70% memory retention, and is represented by an actual numerical value.

[0065] Personalized data for strength training includes initial load FK (such as the weight of the barbell used in the first workout), load increase coefficient az (which increases with each workout repetition), and supercompensation time CT (which represents actual values ​​for the user with user ID "1"). "Supercompensation" refers to the phenomenon where muscle strength levels increase to pre-training levels after 48 to 72 hours of rest following training. The time required for this supercompensation is called "supercompensation time."

[0066] Personalized data for dialogue consists of an initial target selected by the user from a selection of initial targets such as intelligent males, cute females, and famous celebrities, along with personal weights for emotions such as joy, anger, sadness, and happiness. In a typical model, the initial weights for joy, anger, sadness, and happiness are predetermined for each initial target. During the dialogue with the initial target selected by the user, these initial weights are gradually modified in response to user requests, ultimately resulting in a personalized weight that is optimal for that user, enabling dialogue that expresses emotions that match the user's preferences. This will be explained in more detail later.

[0067] Personalized data for task processing consists of an initial target selected by the user from a list of initial targets such as patent attorneys, lawyers, and famous scientists, and personalized weights for each type of knowledge. In a general model, the initial weights of knowledge for each initial target are predetermined, and as the user interacts with the selected initial target for work, the initial weights of the knowledge used are gradually modified in response to the user's questions and requests, ultimately resulting in personalized weights that are optimal for that user, enabling answers and advice that match the knowledge the user requires for their work. This will be explained in more detail later.

[0068] The artificial intelligence server 9 uses the personalized data described above to provide personalized instruction (services) to each user. Specifically, it uses personalized data for memorization to provide instruction according to a review plan (review schedule) that minimizes the user's total review time, uses personalized data for muscle training to provide instruction according to a muscle training schedule that maximizes the effect of repeated muscle training, uses personalized data for dialogue to engage in conversations that express emotions such as joy, anger, sadness, and happiness that match the user's preferences, and uses personalized data for task processing to provide answers and advice that match the knowledge the user requires for their work.

[0069] Next, referring to Figure 11, we will explain the flowchart of the subroutine program for the external display processing (S2), display processing (S3), and the "Like" processing by the SNS server 11 in Figure 4.

[0070] Figure 11 shows the processing for a digital menu board, which is an example of a display 5. A digital menu board is a display used to show menus in restaurants and other establishments. By displaying the menu digitally, it is possible to significantly reduce the time and cost for the establishment by enabling immediate display of sold-out items and menu changes. In this embodiment, as shown in Figure 1, the digital menu board 5 is connected to the internet 1 and configured to communicate with the artificial intelligence server 9 and the SNS server 11. In the flowchart of Figure 11, images and videos taken by users with a wearable computer 3, etc., and uploaded to the SNSDB 13 and user DB 12, etc., are displayed on the digital menu board 5, allowing customers who have come to the restaurant together to view and enjoy them while dining.

[0071] First, S100 determines whether or not the menu has been updated. If the store updates today's menu, control proceeds to S101 and the menu update process is performed. Next, control proceeds to S102 and the menu is displayed.

[0072] Next, if a user wants to display images or videos that they have taken with their wearable computer 3 or other device and uploaded to SNSDB 13 or user DB 12 on the digital menu board 5, they first perform an operation to access the artificial intelligence server 9. Then, S104 determines that it is YES and control proceeds to S105, where the user ID and access request are sent to the artificial intelligence server 9 and access processing is performed. The artificial intelligence server 9, upon receiving the access request, determines that it is YES in S106. The mobile communication device 3 communicates with the artificial intelligence server 9 via S107 and S108 to select and specify the images or videos that the user wants to display on the digital menu board 5. The artificial intelligence server 9 searches user DB 12 via S108 to identify the selected images or videos.

[0073] Once the image to be displayed has been identified, the user (customer) taps the touch screen of the digital menu board 5 to initiate communication. S103 then determines that this is YES, and S109 initiates the communication start process with the wearable computer 3. The user (customer) also initiates communication on their own wearable computer 3. S108 then determines this is YES, and control proceeds to S110, initiating the communication start process and establishing a communication connection between the wearable computer 3 and the digital menu board 5.

[0074] Next, the digital menu board 5 transmits the display ID assigned to it to the wearable computer 3 (S111). The wearable computer 3, having received it in S112, transmits the display ID to the artificial intelligence server 9 (S113). The artificial intelligence server 9, having received it in S114, transmits the data identified in S108 to the display 5 corresponding to the received ID (S105). The digital menu board 5, having received it, makes a YES decision in S116 and displays the received data (S117). As a result, the image or video specified by the user (customer) is displayed on the digital menu board 5. Consequently, images and videos can be viewed on the relatively large display screen of the digital menu board 5.

[0075] A key point to note in the control described above is that when a user selects and specifies data uploaded to the internet, they directly access the artificial intelligence server 9 using their mobile communication terminal 3 without going through the display 5. If the display 5 itself is connected to the internet and the data to be displayed on the display 5 is to be specified, it would be quicker to access the data on the internet via the display 5 and display that data on the display 5. However, the data uploaded by the user to the internet is personal information related to the user's privacy, and if the user's ID etc. were transmitted via the display 5 in order to access that personal information, there is a risk of the user's ID and personal information being leaked. Therefore, in this embodiment, when a user selects and specifies data uploaded to the internet, they directly access the artificial intelligence server 9 using their mobile communication terminal 3 without going through the display 5.

[0076] Next, the digital menu board 5 determines in S118 whether the data display has finished or not. If it has finished, control proceeds to S119, where the process of directing the user (customer) to the store's homepage is initiated. Specifically, the URL of the store's homepage is sent to the mobile communication terminal 3. Upon receiving this, the mobile communication terminal 3 sends an access request to the page at that URL to the SNS server 11 in S120. Upon receiving this in S121, the SNS server 11 sends the store's homepage to the mobile communication device 3 in S122.

[0077] Upon receiving this, the mobile communication device 3 displays the store's homepage via S123. Meanwhile, the digital menu board 5 displays the message "Please tap 'Like' on our store's homepage" via S124. When a user (customer) sees this and taps "Like," S125 determines it's YES, and control proceeds to S126, sending the "Like" signal to the SNS server 11. The SNS server 11 receives this via S127 and processes the "Like" registration via S128. As a result, the store can provide information to users (customers) via SNS. For example, it can introduce seasonal new menu items, distribute coupons, and notify customers of upcoming events, which can encourage repeat visits.

[0078] If you press Enter during the external display processing, the program will proceed to S4; if you press Enter during the display processing, the program will proceed to S5.

[0079] Next, referring to Figure 12(a), the flowchart of the subroutine program for expert data processing shown in S6 of Figure 4 will be explained. Maintenance data processing is performed in S135, muscle training data processing is performed in S136, other processing is performed in S137, and the program returns to proceed to S8. Maintenance data processing receives and processes data on the results of maintenance and inspection of buildings sent from a maintenance specialist, which is an example of various specialist companies. Muscle training data processing receives and processes aggregated data on the muscle training results of many members sent from a sports gym, which is an example of various specialist companies.

[0080] Next, referring to Figure 12(b), the flowchart of the maintenance data processing subroutine program described above will be explained. The maintenance specialist performs maintenance on the building based on the detection data transmitted from the wireless sensor network 4 shown in Figure 8 and stored in the wireless sensor DB 16, and inputs the results (KO, NG, etc.) for each sensor ID into the PC 10 (S140).

[0081] Then, the maintenance specialist PC10 sends the input inspection results to the artificial intelligence server 9 for each sensor ID via S141. The artificial intelligence server 9 determines whether or not it has received the inspection results via S145, and if it has not, control proceeds to S151. On the other hand, if it has received the inspection results, control proceeds to S146, and the inspection result data is added to the learning DB60 for each sensor ID. Next, in S147, wireless sensor data is read from the wireless sensor DB16, the inspection result data is attached, and data mining is performed to find the pattern NGP1 for NG. For example, a pattern NGP1 is found in which there is a high probability of an abnormal (NG) inspection result when the strain and vibration period have a predetermined relationship.

[0082] Next, in S148, it is determined whether NGP1 is a new one that has been found so far. All the patterns of NG found so far are stored in the memory area of ​​the "building maintenance and inspection model" in the artificial intelligence DB17, and it is determined whether the NGP1 that was found this time is already stored in the artificial intelligence DB17. If it is determined that the same thing is already stored, control proceeds to S151, but if it is determined that it is not yet stored, control proceeds to S149, and NGP1 is added to the memory area of ​​the "building maintenance and inspection model" in the artificial intelligence DB17 as an NG pattern.

[0083] Next, in S150, the wireless sensor DB16 is searched to extract sensor IDs that have the NGP1 pattern from among the wireless sensors that have undergone periodic checks, and these are reported. Wireless sensors are checked periodically at regular intervals (for example, every 24 hours) (S151, S152), and the newly found NGP1 is applied to past sensor data that has already undergone periodic checks to perform a re-check.

[0084] Next, S151 determines whether it is time for a periodic check. If it is time for a periodic check, S152 searches the wireless sensor DB16 to extract sensor IDs with NG patterns from the wireless sensor data added since the last periodic check and issues an NG notification. As mentioned above, new sensor data is added to the wireless sensor DB16 as it is stored, and all NG patterns stored in the artificial intelligence DB are applied to check new sensor data that has not yet been periodically checked. After processing in S152, the system returns, and control proceeds to S136.

[0085] Next, referring to Figure 13(a), the flowcharts of the subroutine programs for the Around Data Terminal Processing shown in S4 and the Around Data Server Processing shown in S5 of Figure 4 will be explained.

[0086] S160 executes user learning terminal processing, S161 executes user learning server processing, S162 executes muscle training terminal processing, S163 executes muscle training server processing, and S164 and S165 execute other processing.

[0087] The user learning terminal processing and user learning server processing involve collecting around data from each user during learning to create a general memorization learning model for each user, as well as generating personalized learning data for each user. The muscle training terminal processing and muscle training server processing involve collecting around data from each user during muscle training to generate personalized muscle training data for each user.

[0088] Next, referring to Figure 13(b), we will explain the flowchart of the subroutine program for the personal service terminal processing shown in S7 and the personal service server processing shown in S8 of Figure 4.

[0089] S170 executes the learning service terminal process, S171 executes the learning service server process, S172 executes the muscle training service terminal process, and S173 executes the muscle training service server process.

[0090] The learning service terminal processing provides users with instruction based on a general memorization learning model and personalized learning data, following the most efficient review plan. The muscle training service terminal processing and muscle training service server processing provide users with instruction based on a general muscle training model and personalized muscle training data, following the most efficient muscle training plan.

[0091] Next, referring to Figure 14, we will explain the flowcharts of the subroutine programs for user learning terminal processing and user learning server processing.

[0092] S180 determines whether the user is currently learning or not. This determination may be made based on the user's explicit declaration of intent, or the artificial intelligence server 9 may make the determination autonomously by processing the user's voice using natural language processing or analyzing the video in the user's gaze direction. If it is determined that the user is not currently learning, the system returns and control proceeds to S162. However, if it is determined that the user is learning, control proceeds to S181, and the system sends surrounding data such as the video in the user's gaze direction, gaze position, and audio to the artificial intelligence server 9. Upon receiving this data in S182, the artificial intelligence server 9, in S183, searches a predetermined database for useful information such as the frequency of past entrance examination questions and the schools that have asked questions related to the currently being learned subject, and sends this information to the user (mobile communication device 3 such as a wearable computer).

[0093] The user's mobile communication device 3, which receives the data via S184, displays the received useful data as an overlay using AR (Augmented Reality) via S185. The user will proactively send the learning data to the AI ​​server 9 so that they can receive guidance according to the most efficient review plan by sending the learning data to the AI ​​server 9. Furthermore, the display of useful information such as the frequency of entrance exam questions will further increase the incentive for the user to send the learning data to the AI ​​server 9.

[0094] Next, in S186, the artificial intelligence server 9 classifies the received data for various models and stores it in the user DB 12 and the learning DB 60. Then, in S187, it determines whether the memorization learning model is complete or not. If it is not yet complete, control proceeds to S188, where the received data is converted into numerical values ​​and labels to create data for machine learning, and in S189, it models the data using a learning algorithm (such as reinforcement learning or regression). An example of this modeling is explained below.

[0095] Using a vast amount of historical learning data collected for each user as training data, and under the condition that each user can maintain a certain level of memory retention (e.g., 90%) at a predetermined endpoint such as the entrance exam date, reinforcement learning is used to find a review plan that minimizes the total time spent on repeated reviews from the initial learning to the aforementioned endpoint. For example, if Ti is the period from the i-th review to the next review, and THi is the time taken for the i-th review, then the total review time T = ΣTHi. Reinforcement learning is performed to find the Ti that minimizes T under the condition that the endpoint memory retention rate ≥ 90%.

[0096] In reinforcement learning, when Q(st,at) represents the value of performing action at in state st, a model-based method can be used to estimate the Q-value if knowledge of the environment model, i.e., the probability distribution of state transition probabilities and rewards, is given. However, if the environment model is unknown, TD (Temporal Difference) learning is used. First, since exploration of the environment is necessary, the ε-greedy method is used. In the initial stages of exploration, various actions are tried, and as the system settles, the concept of temperature is introduced to select the most optimal actions. With temperature T, actions are selected according to the probability expressed by the following equation.

[0097] P(a|s)={exp(Q(s,a) / T)} / {Σexp(Q(s,b) / T)} (Note that b∈A is written under the Σ, but this is omitted in the above formula.) Here, a is an action, and Q(s,a) is the value of performing action a in state s.

[0098] T is called the temperature in annealing; a higher T results in actions being selected with nearly equal probability, while a lower T biases the selection towards the optimal action. As learning progresses, decreasing the value of T stabilizes the learning results. Alternatively, the user can be instructed to actually perform the actions selected in this way, thereby actively collecting learning data.

[0099] An example of the results of reinforcement learning is shown in Figure 15. Referring to Figure 15, the vertical axis of the graph is Ti (review interval), and the horizontal axis is i (review count, indicating which review it is). The circles indicate the results of reinforcement learning for user ID:1. The crosses indicate the results of reinforcement learning for user ID:2. The triangles indicate the results of reinforcement learning for user ID:3.

[0100] Next, using a regression algorithm, we take the machine learning results data from these multiple users as input data, assume that this input data outputs a target based on a certain function, and then find that function. The resulting function is: Ti = T0·i This is a linear function of the straight line. The coefficient T0 is an element that differs from user to user (memorization ability, etc.). Specifically, it is the period from initial learning to just before forgetting for each user (specifically, the time from initial learning until the memory retention rate reaches 70%). In this way, the element that differs from user to user (memorization ability, etc.) is represented as a constant T0 in a general model, Ti = T0·i That is the case.

[0101] Returning to Figure 14, S190 causes the artificial intelligence DB17 to memorize the memorization learning model (Ti=T0·i). If an old and incomplete memorization learning model is already memorized, it is updated to a newer and more complete memorization learning model (Ti=T0·i).

[0102] On the other hand, if S187 determines that the memorization learning model is complete, control proceeds to S191, where personalized data is created based on the classified received data and stored in the user DB12. In this case, the "personalized data" is the user-specific element (such as memorization ability) T0, that is, the time from initial learning until the memory retention rate reaches 70%. This can be easily derived from the user's learning-around data stored in the user DB12.

[0103] Next, referring to Figure 16, the flowcharts of the subroutine programs for the learning service terminal processing shown in S170 and the learning service server processing shown in S171 in Figure 13(b) will be explained.

[0104] The mobile communication device 3, such as a wearable computer, determines in S200 whether or not there has been a review request from the user. If there is no review request, it returns and control moves to S172. If there is a review request, it sends the review request to the artificial intelligence server 9 in S201. The artificial intelligence server 9, having received it in S202, determines in S203 whether or not the memorization learning model is complete. If it is not complete, control moves to S205. If it is already complete, control proceeds to S204, where T0=TK is substituted into the general memorization learning model Ti=T0·i to extract learning items from the user DB12's around data that are close to the time for repeated review. Note that TK refers to the period from the user's initial learning until just before forgetting (specifically, the time from initial learning until the memory retention rate reaches 70%).

[0105] Next, in S205, learning items that the user was distracted from during class or other activities are extracted from the around data in the user DB12. This is because the user's gaze position during learning is sent as around data to the artificial intelligence server 9 (S181) and stored in the user DB12, making it possible to determine whether or not the user was distracted. Next, in S206, the learning items extracted in S204 and S205 are sent to the mobile communication device 3. Upon receiving this in S207, the mobile communication device 3 processes the received learning items in S208 to allow the user to review them.

[0106] In addition to or instead of the distraction, control may be implemented to prompt the user to review the learning material while the user is chatting during class or other activities. In other words, the system includes a collection means for collecting information that can identify the user's attitude during learning, a recording means for recording the learning content the user is receiving, a discrimination means for determining parts of the user's attitude that are inappropriate for learning from the information collected by the collection means, an extraction means for extracting learning content corresponding to the parts determined by the discrimination means from the recording means, and a transmission means for sending the learning content extracted by the extraction means to the user terminal. Note that S205 is not necessarily required. After processing S206, the system returns and control proceeds to S173.

[0107] In the control described above, the user's personalized data is substituted into a general model to create a personalized model specifically for that user, and then the personalized service is provided to that user using that personalized model. However, it is also possible to generate a custom-made personalized model for the user and use that to provide the personalized service to that user. For example, at the point where a YES decision is made in S187 of Figure 14, the machine learning results of each of the multiple users (subjects) who contributed to the creation of the general learning model are available (see Figure 15). Therefore, each of these subject users can have their own custom-made personalized model based on the results of the machine learning performed on their learning, and the system may be controlled to provide the personalized service to the user using that custom-made model. Furthermore, even users other than the above-mentioned subjects may be controlled to provide a large amount of their own learning data (review data) to the artificial intelligence server 9 in order to have a custom-made personalized model created for them if they wish to generate one for themselves.

[0108] Next, referring to Figure 17, we will explain the flowchart of the subroutine program for processing muscle training data shown in S136 of Figure 12(a).

[0109] In one example of a specialized company, a sports gym's PC10 uses S210 to collect muscle training data from numerous members (subjects) and transmit it to the artificial intelligence server 9. The muscle training data includes, for example, the subject's initial load, training time and repetitions, the training interval between the previous and current training sessions, the load increase coefficient, and measured muscle thickness. The artificial intelligence server 9 receives this muscle training data in S211 and, in S212, classifies the received data for various models and stores it in the learning DB60. Next, in S213, it converts the received data into numerical data and a set of labels to create training data.

[0110] Next, S214 models the data based on the training data using learning algorithms such as reinforcement learning and regression. This modeling method is the same as the one explained with reference to Figure 15. A general model for muscle training is: Fi = F0 + (i-1) / a Here, i is the number of repetitions indicating which repetition of the strength training it is, Fi is the load at the i-th strength training (e.g., barbell weight), F0 is the load at the first strength training (e.g., barbell weight), and a is the load increase coefficient that increases with each repetition of the strength training. These F0 and a are factors that differ from user to user (e.g., individual differences in muscle strength).

[0111] Next, S215 stores the above general model (Fi=F0+(i-1) / a) in the memory area of ​​the AI ​​DB17's muscle training learning model. If an old and incomplete muscle training learning model is already stored, it is updated with a newer, more complete memorization learning model (Fi=F0+(i-1) / a).

[0112] Next, referring to Figure 18(a), the flowcharts of the subroutine programs for the muscle training terminal processing shown in S162 and the muscle training server processing shown in S163 in Figure 13(a) will be explained.

[0113] The user's mobile communication device 3 determines in S220 whether the user is currently doing strength training. If the user is not doing strength training, it returns and control proceeds to S164. If the user is doing strength training, control proceeds to S221, where the user's strength training data is collected and sent to the artificial intelligence server 9. The artificial intelligence server 9 receives this data in S222 and, in S113, stores the received data in the learning DB 60, and also calculates the user's initial load FK, load increase coefficient az, and supercompensation time CT, and stores them in the user DB 12. After S113, it returns and control proceeds to S165.

[0114] Next, referring to Figure 18(b), the flowcharts of the subroutine programs for the muscle training service terminal processing shown in S172 and the muscle training service server processing shown in S173 in Figure 13(b) will be explained.

[0115] The user's mobile communication device 3, in S226, sends a muscle training menu request signal to the artificial intelligence server 9 in response to the user's request for a muscle training menu. The artificial intelligence server 9, having received this signal in S227, reads the user's initial load FK, load increase coefficient az, and supercompensation time CT from the user DB 12 in S228, substitutes them into a general muscle training model (Fi=F0+(i-1) / a), and creates the current muscle training menu. Then, in S229, it sends the muscle training menu and an advertisement from the muscle training specialist company that provided the muscle training data (see S210) to the user's mobile communication device 3. The mobile communication device 3, having received this signal in S230, notifies the user of the received muscle training menu and the advertisement from the muscle training specialist company in S231. After S231, it returns and control moves to S9, and after S229, it returns and control moves to S10.

[0116] Figure 19(a) shows details (specific examples) of the data stored in the memory area of ​​the dialogue model in the AI ​​DB17 in Figure 10. The memory area of ​​the dialogue model stores data for initial target, dialogue template, initial weights for emotions, and emotion change function. The initial target includes various selection targets such as male intelligent type, female intelligent type, male wild type, female moe type, talent A, talent B, and talent C. The dialogue template, initial weights for emotions, and emotion change function data differ for each type of initial target, and each data is stored in association with each type of initial target. The dialogue template is a large collection of dialogue templates used to find responses to user dialogues through template matching. For example, in the case of male intelligent type, the proportion of intelligent dialogue templates is high.

[0117] The initial weights for joy, anger, sadness, and pleasure, and the change function for joy, anger, sadness, and pleasure, are used to control the change in emotions in response to the interaction. If the user praises, the emotion will change to "joy"; if the user criticizes, the emotion will change to "anger"; if the user is sad, the emotion will change to "sadness"; and if the user is happy, the emotion will change to "happiness".

[0118] However, since the weights of emotions (joy, anger, sadness, and happiness) and the manner of change differ depending on the type of initial target, different initial weights and change functions for emotions are used for each type of initial target. The initial weights and change functions for emotions are found using machine learning, with a large amount of dialogue data collected from conversations with many users as training data. Many users are classified into male intellectual types, female intellectual types, male wild types, and female moe types, and the average for each type is calculated to create initial weights and change functions for emotions appropriate to each initial target. For talents A, B, C, etc., the initial weights and change functions for emotions are found using machine learning, with a large amount of dialogue data collected from conversations with a specific talent (e.g., talent A) as training data. Alternatively, the change function for emotions could be a unified function common to all people, and the differences (variations) between individuals could be controlled by adjusting the initial weights for emotions.

[0119] One example of a function of change in emotions F(x) is as follows: Let K, D, I, and R represent joy, anger, sadness, and happiness, respectively, and M represent a state of no emotion. Let x1 be the number of times the user praised, x2 the number of times they criticized, x3 the number of times they were sad, and x4 the number of times they were happy. Let K, D, I, and R be the initial weights of joy, anger, sadness, and happiness, respectively. Also, the sigmoid function 1 / (1+e -x If we express ) as S(x) for convenience, F(x)=KS(x1-K)+DS(x2-D)+IS(x3-I)+RS(x4-R)+M(1-S(x1-K)-S(2x-D)-S(x3-I)-S(x4-R)) This means that if any of x1, x2, x3, or x4 exceeds their respective thresholds (initial weights) K, D, I, or R, they will change to the emotion that was exceeded (joy, anger, sadness, or happiness). However, if none of them exceed the thresholds, they will remain emotionless (M).

[0120] Figure 19(b) shows the details of the data stored in the personalization data storage area for conversations in the user DB12 of Figure 10. Figure 19(b) shows the case where user ID:1 selects "Talent A" as the initial target for the conversation model. The personal weights for joy, anger, sadness, and happiness are stored as "0.7AK, 1.5AD, 1.3AI, 0.5AR".

[0121] As shown in Figure 19(a), the initial weights for joy, anger, sadness, and happiness of "Talent A" are "AK,AD,AI,AR". However, as the user engages in dialogue based on these initial weights, the initial weights "AK,AD,AI,AR" are gradually modified according to the user's requests, and are currently "0.7AK,1.5AD,1.3AI,0.5AR". As a result, the emotion change function for this user with personal weights "0.7AK,1.5AD,1.3AI,0.5AR" is as follows. F(x)=KS(x1-0.7AK)+DS(x2-1.5AD)+IS(x3-1.3AI)+RS(x4-0.5AR)+M(1-S(x1-0.7AK)-S(2x-1.5AD)-S(x3-1.3AI)-S(x4-0.5AR))

[0122] As a result, compared to the actual talent A, joy (K) and happiness (R) are more likely to appear, while anger (D) and sadness (I) are less likely to appear.

[0123] In this way, users can select and specify an initial target that matches their preferences, and as they continue to interact with the AI, they can further modify the AI's emotional weighting to better reflect the user's preferences. In other words, while it is almost impossible to transform a human being into one's ideal partner, with artificial intelligence, it is possible to shape (nurture) the AI ​​into one's ideal partner. Up until now, we have lived in an era where we searched for and met our ideal lover, best friend, or partner, but in the future world, we will live in an era where we create our ideal lover, best friend, or partner on the network (on the cloud).

[0124] Next, referring to Figure 20, we will explain the flowchart of the subroutine program for the dialogue terminal processing shown in S9 and the dialogue server processing shown in S10 in Figure 4.

[0125] The mobile communication device 3, such as a wearable computer, determines in S235 whether the user has started a dialogue. If it has not, it returns and control proceeds to S11. If the user has started a dialogue, control proceeds to S236, and the user ID and a dialogue start signal are sent to the artificial intelligence server 9. Upon receiving this, the artificial intelligence server 9 makes a YES determination in S241, and determines in S242 whether the initial target has been received. If it has not been received, control proceeds to S244.

[0126] Meanwhile, in the mobile communication device 3, S237 determines whether or not an initial target has been selected. If no selection has been made, control proceeds to S239. However, if a selection has been made, S238 sends the selected initial target to the artificial intelligence server 9. Upon receiving it, the artificial intelligence server 9 determines YES in S242 and, in S243, updates the initial target for interaction corresponding to the user ID in the user DB 12 with the selected target.

[0127] Next, the mobile communication device 3 and the artificial intelligence server 9 interact with each other via S239 and S244, respectively. During this interaction, the artificial intelligence server 9 interacts according to the initial target for dialogue corresponding to the user ID in the user DB 12, the personal weights of emotions (joy, anger, sadness, and happiness), and the corresponding emotion change function in the artificial intelligence DB 17. The artificial intelligence server 9 is equipped with a natural language processing engine and incorporates not only the text mining processes of "morphological analysis" and "syntactic analysis (processing to determine dependency relationships between clauses)," but also two technologies: "contextual analysis," which is necessary to grasp the structure of a sentence, and "semantic analysis," which is necessary to understand the intent of a sentence. This makes it possible to analyze the user's voice "faster" and "more precisely."

[0128] The artificial intelligence server 9 determines during the above dialogue whether it is necessary to change the personal weights of joy, anger, sadness, and other emotions (S245), and if necessary, updates the personal weights of joy, anger, sadness, and other emotions in S246. For example, if the AI ​​server 9 recognizes that the user has uttered a predetermined phrase such as "Please reduce the weight of anger a little," it will perform control to update the personal weights of joy, anger, sadness, and other emotions. Alternatively, instead of or in addition to the above set phrases, the AI ​​server 9 may also perform control to update the personal weights of anger, sadness, and other emotions in response to the user's statements such as "You're getting too angry."

[0129] Next, the artificial intelligence server 9 determines in S247 whether the conversation is work-related or not. If it is not work-related, it proceeds to S249 to determine whether the conversation has ended. If it is work-related, it proceeds to S248 to perform task server processing.

[0130] In the artificial intelligence server 9, if it is determined by S249 that the dialogue has not ended, control is transferred to S244, and steps S244 to S249 are repeatedly executed. When the dialogue is finished, it returns and control is transferred to S12. On the other hand, in the mobile communication device 3, if it is determined by S240 that the dialogue has not ended, control is transferred to S237, and steps S237 to S240 are repeatedly executed. When the dialogue is finished, it returns and control is transferred to S11.

[0131] Figure 21(a) shows details (specific examples) of the data stored in the memory area of ​​the task processing model in the artificial intelligence DB17 in Figure 10. The memory area of ​​the task processing model stores data for the initial target, the initial weights of each piece of knowledge, and the knowledge utilization function. The initial targets include various selection targets such as patent attorneys, lawyers, physicists, chemists, scientist A, scientist B, and scientist C. The initial weights of each piece of knowledge and the knowledge utilization function data differ for each type of initial target, and each piece of data is stored in association with each type of initial target.

[0132] The initial weights and knowledge utilization functions for each piece of knowledge are designed to make it easier for each initial target to use knowledge that matches their area of ​​expertise when performing tasks such as work. These initial weights and functions are determined using machine learning, based on a large amount of dialogue data collected from work-related conversations with numerous users. The users are classified by profession, such as patent attorneys, lawyers, physicists, and chemists, and the average for each profession is calculated to create the appropriate initial weights and functions for each piece of knowledge for each initial target.

[0133] For example, a patent attorney user might ask numerous questions and seek advice from the AI ​​server 9 regarding their work. A large amount of data, consisting of responses such as "satisfied" or "not satisfied" to the AI ​​server 9's responses, is used as training data to model the AI ​​using algorithms such as reinforcement learning or regression. For scientists A, B, C, etc., a large amount of dialogue data collected from actual conversations between scientists is used as training data, and initial knowledge weights and knowledge utilization functions are determined through machine learning. The knowledge utilization function may be a unified function common to all individuals, and the differences (variations) among individuals may be adjusted using the initial knowledge weights.

[0134] One example of a knowledge utilization function G(x) is as follows: Each piece of knowledge is represented as n1, n2, n3, ..., and the number of positive responses by the user to each piece of knowledge, such as "satisfied", is x1k, x2k, x3k, ..., and the number of negative responses, such as "not satisfied", is x1h, x2h, x3h, ..., and the initial weights of each piece of knowledge are w1, w2, w3, ..., respectively. Also, the sigmoid function is 1 / (1+e -x If we express ) as S(x) for convenience, G(x)=w1n1S(x1k-x1h)+w2n2S(x2k-x2h)+w3n3S(x3k-x3h)+... This means that knowledge for which the number of positive responses exceeds the number of negative responses is considered useful knowledge, and it is possible to control the system so that knowledge with larger coefficients (weights w1, w2, w3, ...) for each piece of knowledge n1, n2, n3, ... is used preferentially.

[0135] Figure 21(b) shows the details of the data stored in the personalization data storage area for task processing in the user DB12 of Figure 10. Figure 21(b) shows the case where user ID:1 selects "Patent Attorney" as the initial target of the dialogue model. The personal weights w for each piece of knowledge n are stored as "1.2·1w11n1, 0.8·1w21n2, 0.7·1w31n3, 1.3·1w41n4".

[0136] As shown in Figure 21(a), the initial weights of each piece of knowledge for "patent attorney" are "1w11n1, 1w21n2, 1w31n3, 1w41n4". However, as the user engages in work-related dialogue based on these initial weights, the initial weights of each piece of knowledge are gradually modified according to the user's requests, and currently they are "1.2·1w11n1, 0.8·1w21n2, 0.7·1w31n3, 1.3·1w41n4". In this way, not only can the initial target be selected and specified according to the user's profession, but as work-related dialogue continues, the personal weights of each piece of knowledge can be modified to better reflect the user's questions and consultations. In other words, in the future world, it will be an era where you can create your ideal work partner on the network (on the cloud).

[0137] Next, referring to Figure 22, the flowcharts of the subroutine programs for task terminal processing shown in S11 of Figure 4 and task server processing shown in S12 and S248 of Figure 20 will be explained.

[0138] The user's mobile communication device 3 determines in S255 whether the user has requested to register their personal weights as initial targets in the artificial intelligence DB 17. If not, control proceeds to S259, where it is determined whether there has been a request to update the initial targets. If not, control proceeds to S265.

[0139] On the other hand, if S255 requests that the user register their personal weights as initial targets in the artificial intelligence DB17, S255 will determine that the response is YES, and control will proceed to S256, where the user ID and the initial target registration request will be sent to the artificial intelligence server 9.

[0140] In the artificial intelligence server 9, S257 determines whether or not an initial target registration request has been received. If it has not been received, control proceeds to S261, where it determines whether or not an initial target request has been received. If it has not been received, control proceeds to S267.

[0141] If the user ID and initial target registration request are transmitted from the mobile communication device 3 according to the processing in S256, a YES decision is made in S257, and control proceeds to S258, where the personalized weights of each piece of knowledge corresponding to the user ID are read from the user DB 12 and registered in the artificial intelligence DB 17 along with the user name. In this way, in this embodiment, a user can register the personalized weights of each piece of knowledge of their own in the artificial intelligence DB 17 and make them widely available to the public. For example, when efficiently training a subordinate at work to turn them into a full-fledged expert, if the superior transfers the personalized weights of each piece of work knowledge they have cultivated so far to that subordinate, the artificial intelligence server 9 will respond to the subordinate's questions and consultations on their behalf, and the artificial intelligence server 9 will take over the training of the subordinate. When a user transfers the personalized weights of each piece of knowledge of their own to another person, a settlement method may be provided for settlement with the transferee, and the transfer may be made for a fee.

[0142] If a user requests the mobile communication device 3 to update the initial target and specifies a new initial target (requested initial target), S259 determines YES and control proceeds to S260, where the user ID and the requested initial target are sent to the artificial intelligence server 9. Upon receiving this, the artificial intelligence server 9 determines YES in S261 and control proceeds to S262, where it determines whether the received requested initial target has been used by the user before. If it has not been used before, S264 reads the requested initial target from the artificial intelligence DB 17 and stores it corresponding to the user ID in the user DB 12, and control proceeds to S267.

[0143] If the initial target of a received request has been used previously by that user, that initial target may already be stored in the user DB12, and its personal weight may have been updated as a result of its use (see S271). Therefore, if the initial target of a received request has been used previously by that user, control proceeds to S263, and the personal weight of the previously used knowledge stored in the user DB12 is used.

[0144] Next, referring to Figure 23, the artificial intelligence server 9 reads the initial target and personal weights for each piece of knowledge corresponding to the user ID in the user DB 12 in S267, and reads the knowledge utilization function corresponding to the read initial target from the artificial intelligence DB 17 in S268. Then, when the user uses the mobile communication device 3 to perform a work-related dialogue, a dialogue takes place between the mobile communication device 3 (S265) and the artificial intelligence server 9 (S269). In S269, the system processes the retrieved knowledge utilization function and personal weights to perform the work-related dialogue. On the mobile communication device 3 side, the work-related dialogue control by S265 continues until it is determined by S266 that the work-related dialogue has ended. On the artificial intelligence server 9 side, the control from S267 to S274 is repeated until it is determined by S274 that the work-related dialogue has ended, and the work-related dialogue control by S269 continues.

[0145] If, while the control in S267-S274 is being repeated, S270 determines that it is necessary to change the personal weights, the control proceeds to S271, and the personal weights in the user DB12 are updated. If those personal weights are also registered in the artificial intelligence DB17 by the user (see S255-S258), S272 determines YES, and S273 updates the weights in the artificial intelligence DB17 as well. This determination by S270 is made, for example, by the user saying predetermined set phrases such as "satisfied" or "not satisfied" in response to the response from the artificial intelligence server 9, which determines the need to change the weights of the knowledge used by the artificial intelligence server 9. Alternatively, the determination may be made based on the user's normal conversation, such as "Thank you for the great answer," instead of or in addition to the above set phrases.

[0146] Then, if the mobile communication device 3 determines in S266 that the work-related dialogue has ended, S273 sends a work-related dialogue termination instruction to the artificial intelligence server 9, returns, and control transitions to S1. Upon receiving this, the artificial intelligence server 9 makes a YES decision in S274, returns, and control transitions to S3.

[0147] Modifications and features of the embodiments described above are described below. (1) At least one or all of the IoT server 9 and IoT device DB 15, artificial intelligence server 9, artificial intelligence DB 17, user DB 12 and learning DB 60, and SNS server 11 and SNS DB 13 may be configured in a network cloud. Furthermore, in the embodiments described above, the mobile communication device 3 was mainly used as an example of a user terminal, but the user terminal may also be a user PC 7 or a robot 6. Also, as a specific example of the mobile communication device 3, a wearable computer such as smart glasses was shown, but it is not limited to that, and various other devices can be considered, such as mobile phones, smartphones, and vehicles such as automobiles with communication functions.

[0148] (2) The IoT server 8 is provided with identification means (S24-S26, S38, S40) for identifying IoT devices that need to exchange information. The geographical location of the IoT devices identified by the identification means may be transmitted to the wearable computer and notified to the user, thereby encouraging the user to act as an intermediary. In this case, the system may be controlled to notify the user that high scores (points, etc.) will be awarded to IoT devices that require particularly high levels of intermediation.

[0149] (3) The system may be controlled to grant high rewards (such as points) to only a limited number of IoT devices, thereby promoting mediation activities by utilizing the same psychology as a lottery. Alternatively, the system may be controlled to allow mediation activities for IoT devices installed in geographical locations to be carried out using location-based games.

[0150] (4) The embodiments described above disclose the following inventions. This invention relates to a system that reduces the cost of IoT devices (such as sensors and actuators) laid out across the globe, thereby enabling the implementation of IoT at a low cost.

[0151] As background technologies for systems that enable IoT at low cost, for example, the following was introduced: A border router (gateway device) is introduced between the IoT sensor and the server on the internet. The server address is not set on the sensor, and after its own address is determined, the sensor data information is sent to the border router. The border router relays the sensor information to the sensor data management server and performs data collection. This allows the server to detect the sensor's address and also perform pull-type data collection addressed to the sensor address (Japanese Patent Publication No. 2014-78773).

[0152] However, with this background technology, while it is possible to collect data via a pull-type method from the server side to the sensor address, it is necessary to either provide each sensor with an internet connectivity function such as a network interface, or prepare a proprietary adapter for connecting to the network. This has the drawback of increasing the cost of realizing IoT.

[0153] On the other hand, among the IoT devices that are laid out all over the globe, there are not necessarily any that do not require pull-type data collection from the server side to the sensor address. For example, in the case of an illuminance sensor installed to aggregate illuminance throughout the year, it is sufficient to store the periodic illuminance detection results in memory and send those stored illuminance detection results to the server once a year when the illuminance is aggregated. Also, in the case of IoT devices installed in places with a lot of human traffic, it is conceivable to use human-assisted methods to send data from the IoT device to the server.

[0154] This invention was conceived in light of these circumstances, and its purpose is to provide a system that can reduce the costs involved in realizing IoT.

[0155] This invention relates to a wearable computer owned by the user, A server that communicates with the wearable computer via the internet, Equipped with a group of IoT devices, The aforementioned IoT device group includes communication means for communicating with the wearable computer, The aforementioned wearable computer is A device-to-device communication means for communicating with the aforementioned IoT device group, A server-to-server communication means for communicating with the aforementioned server via the Internet, An IoT system including an intermediary means that facilitates the exchange of information between an IoT device that is a communication partner via the device-to-device communication means and a server that is a communication partner via the server-to-server communication means.

[0156] With this configuration, the IoT devices do not necessarily need internet connectivity, which reduces costs and makes IoT implementation more affordable.

[0157] Preferably, the server further includes a reward granting means for granting predetermined rewards to the wearable computer that has been mediated by the intermediary means.

[0158] This configuration provides users with an incentive to act as intermediaries, thereby promoting their intermediary activities.

[0159] More preferably, the server further includes determination means for determining whether or not there is an IoT device that needs to exchange information.

[0160] With this configuration, the IoT device group does not need to have a function to determine whether or not information exchange is necessary, which reduces the processing burden on the IoT device group and further lowers costs.

[0161] More preferably, the wearable computer further includes location information transmission means for transmitting information that can identify the current location to the server, The aforementioned server, A location determination means that determines whether or not the IoT device identified by the identification means is located near the wearable computer whose location has been identified by the information transmitted by the location information transmission means, The system further includes, when the location determination means determines that the IoT device is present, an intermediary command transmission means that transmits a command signal to the wearable computer for the intermediary means to perform intermediation for the IoT device, The mediating means performs mediation in accordance with the command signal from the mediating command transmission means.

[0162] With this configuration, the wearable computer initiates the mediation process with IoT devices that require it, eliminating the need for the IoT devices to send out a mediation request signal.

[0163] More preferably, the wearable computer is A sensor that requires a predetermined amount of time to detect, The system further includes a notification means (e.g., S71) for informing the user of the time required for detection by the sensor when performing detection using the sensor at a desired geographical location, If the sensor is able to detect the required detection time within the detection area of ​​the desired geographical location, the detection data is transmitted to the server via the intermediary means.

[0164] With this configuration, if detection requires a certain amount of time, the wearable computer informs the user of that time, thus encouraging the user to cooperate in collecting detection data that takes time.

[0165] (5) In the embodiments described above, specific examples of machine learning were shown, including the creation of a building maintenance and inspection model, a memorization learning model, a muscle training model, a dialogue model, and a task processing model. However, the invention is not limited to these, and various other models can be considered, such as a learning model for infants, a model for an artificial secretary, and a model for an artificial tutor. Similarly, personalized data is not limited to memorization learning, muscle training, dialogue, or task processing, but can be various other models, such as learning data for infants, an artificial secretary, or an artificial tutor. These general models and personalized data are controlled to be switched and used according to the user's situation. As a result, from shortly after birth, humans will grow up receiving personalized services based on artificial intelligence using their own personalized data, and will live their entire lives together with artificial intelligence. In other words, in the future world, humans and artificial intelligence will be paired together. This is a world in which humans and artificial intelligence are paired together to form a single personality. Furthermore, regarding the use of general models and personalized data, in this embodiment, the previous personalized data for dialogue was used unless the user indicated a desire to switch. However, when switching between general models and personalized data, the system may be controlled to automatically switch the personalized data for dialogue to another appropriate one. For example, Talent A could be used as the initial target for dialogue during general everyday conversations, and the system could be controlled to automatically switch to a female intellectual type (see Figure 19) during work-related dialogues.

[0166] (6) Recommendation control may be implemented to recommend artificial intelligence or general models that are thought to be a good match for a user based on surrounding data collected from the user by a mobile communication device 3, etc. Furthermore, when a user searches for artificial intelligence or general models, the search results displayed in the list may be controlled to prioritize those that are thought to be a good match for the user based on the surrounding data collected from the user. In addition, user reviews of artificial intelligence and general models may be collected and displayed. In this case, as mentioned above, since a human and an artificial intelligence are paired to form a single personality, the system may be controlled to collect and display reviews of the paired human and artificial intelligence as a set.

[0167] (7) For children who have learned using the above-mentioned general model and personalized data for learning children, a vast amount of data consisting of the provided learning materials (learning information) and learning results may be collected and used as training data for machine learning to replicate human intellectual growth (supervised learning, etc.). The same learning materials (learning information) provided to the above-mentioned children may also be provided to artificial intelligence, and the learning results of the artificial intelligence that has achieved intellectual growth as a result may be compared with those of the actual children, and reinforcement learning or regression may be performed to minimize the differences, thereby creating a learning model for artificial intelligence that enables learning similar to that of humans. In particular, in areas where artificial intelligence is weak (for example, common sense, emotions, creativity, etc.), it is thought that comparing the learning results of both will result in machine learning (supervised learning, etc.) that is beneficial to artificial intelligence.

[0168] (8) In this embodiment, personalization means (e.g., S191, S223, S246, S271) for personalizing the general model to a model suitable for the user based on information sent from the user is provided on the artificial intelligence server 9. However, instead of or in addition to this, it may be provided on the user's terminal (e.g., a mobile communication device 3 such as a wearable computer, a user PC, a robot, etc.).

[0169] The embodiments described above disclose the following inventions. The disclosed invention relates, for example, to a service provision system and program that utilizes machine learning by artificial intelligence. More specifically, it relates to a service provision system that provides services utilizing machine learning by artificial intelligence, and a program executed by a computer for a user to receive services utilizing machine learning by artificial intelligence.

[0170] There are methods for creating training data from multiple information sources in order to perform machine learning using artificial intelligence (for example, Japanese Patent Publication No. 2011-232997).

[0171] However, machine learning requires a massive amount of training data, and creating this vast amount of data necessitates collecting a large amount of information from various sources. Therefore, in order to put machine learning into practical use, it is essential to reduce the effort and cost required for this information collection.

[0172] Furthermore, for example, a general model created by collecting a large amount of information from a large, unspecified group of people and performing machine learning on that vast amount of training data tends to become an average model for the entire large, unspecified group. If such a general model is used to provide services to individual users who have individual differences, there is a risk that the service will not be suitable for users who deviate from that average.

[0173] The present invention was conceived in view of these circumstances, and its purpose is to reduce the effort and cost required to collect the vast amount of information necessary to create training data. A further purpose is to solve the inconvenience that arises when a service using a general model resulting from machine learning does not match the user.

[0174] Next, an example of the correspondence between various means for solving the problem and embodiments is shown below in parentheses.

[0175] The present invention relates to a service provision system that provides services using machine learning by artificial intelligence (for example, Figures 9 and 10, etc.), A machine learning method (e.g., S189, S214, Figure 15, etc.) for generating a general model (e.g., a memorization learning model Ti=T0·I, a muscle training model Fi=F0+(i-1) / a, etc.) by inputting learning data based on information sent from the user (e.g., video, gaze position, audio, GPS location data, etc. of user 70) and modeling it using machine learning, and Personalization means for personalizing the general model to a model suitable for the user based on information sent from the user (e.g., S191 and S204, S223 and S228, etc.), The system includes a service provision means (e.g., S206, S229, etc.) that provides personalized services to the user using the personalized model (e.g., providing learning materials based on the most efficient review plan, providing a muscle training menu based on the most efficient muscle training plan, etc.), The information sent from the user is used for both machine learning and personalization.

[0176] With this configuration, by providing users with services that utilize the results of machine learning, a large number of users will proactively provide information for training data. Moreover, since the information provided by users is also used to personalize general models into models suitable for those users, it is possible to minimize the inconvenience of each user leaving the provision of information for training data to others and only enjoying the service.

[0177] Preferably, the machine learning means generates the general model by using elements that differ from user to user (for example, individual ability differences such as memorization ability, initial load and load increase coefficient for muscle training, preferences, etc.) as constants (for example, T0, F0, a, etc.) (for example, S189, S214, Figure 15, etc.), The personalization means is A value derivation means (e.g., S191, S223, etc.) for deriving the actual values ​​of the user (e.g., TK, FK, az, CT, etc.) that correspond to the constant parts in the general model, based on information sent from the user, The system includes substitution means (e.g., S204, 228, etc.) for substituting the actual values ​​derived by the value derivation means into the constant portion of the general model to personalize the model to suit the user.

[0178] More preferably, the machine learning means also inputs training data based on information sent from specialized companies (e.g., muscle training specialists) and generates a general model modeled by machine learning (e.g., S214).

[0179] Another aspect of the present invention is a service provision system that provides services utilizing machine learning by artificial intelligence, A storage means (e.g., an artificial intelligence database 17, etc.) for storing multiple types of general models (e.g., the dialogue model in Figure 19(a), the task processing model in Figure 21(a), etc.) generated by modeling using machine learning, A service provision means (e.g., S241-S244, S247, S248, S261-S263, S267-S269, etc.) for providing services to the user (e.g., conversational services, responses to work-related questions and consultations, etc.) using a general model selected by the user from among multiple types of general models stored in the storage means, The service provision means includes personalization means (e.g., S246, S271, etc.) for personalizing the general model to a model suitable for the user in response to the user's response to the service provided by the service provision means, The service provision means includes a personal service provision means for providing services to the user using the model personalized by the personalization means (for example, executing S244, S269, etc., according to the updated personal weights).

[0180] With this configuration, users can select their preferred general model from several general models generated using machine learning, and the service will be provided using that general model. As a result, users can enjoy a service that reflects their preferences. Moreover, the general model is personalized to suit the user based on the user's response to the service, making it possible to provide a service that matches the user's needs through personalization.

[0181] Preferably, the personal service provision means includes means for providing services to the user using the personalized model that was used at the time of the general model selected by the user if that model had previously been used to provide services to the user (e.g., YES in S262) (e.g., S263, etc.).

[0182] More preferably, the system further includes model usage means (e.g., S257, S258, etc.) for allowing others to use the model personalized by the personalization means.

[0183] Another aspect of the disclosed invention is a program executed by a computer for a user to receive a service utilizing machine learning by artificial intelligence, The process involves a step in which the user selects and specifies a desired model from among several general models generated by machine learning using artificial intelligence (for example, the dialogue model in Figure 19(a), the task processing model in Figure 21(a), etc.) (for example, steps S237 and S238, S259 and S260, etc.), Processing steps (e.g., S236-S239, S260, S265, etc.) for the user to receive a service using the general model selected by the selection step, The processing step includes providing the user's response to the artificial intelligence (e.g., S239, S265, etc.) in order to personalize the general model into a model suitable for the user in response to the user's response to the service received, The aforementioned computer will execute the following: The processing step includes a personal service receiving step for the user to receive a personalized service using the model personalized by the artificial intelligence (for example, receiving a response according to the updated personal weights in S239, S265, etc.).

[0184] With this configuration, users can select their preferred general model from several general models generated using machine learning, and the service will be provided using that general model. As a result, users can enjoy a service that reflects their preferences. Moreover, the general model is personalized to suit the user based on the user's response to the service, making it possible to provide a service that matches the user's needs through personalization.

[0185] While embodiments of the present invention have been described above, the embodiments disclosed herein should be considered in all respects to be illustrative and not restrictive. The scope of this invention is indicated not by the above description but by the claims, and all modifications within the meaning and scope of equivalents of the claims are intended to be included. [Explanation of symbols]

[0186] 1. Internet, 2. IoT devices, 3. Mobile communication devices, 4. Wireless sensor network, 8. IoT server, 9. Artificial intelligence server, 10. PCs from various specialized vendors, 12. User database, 15. IoT device database, 17. Artificial intelligence database, 60. Learning database, 38. Various sensors.

Claims

1. A service delivery system that enables the provision of services using machine learning models, A personalization method that uses a standard model modeled by machine learning with training data collected from a large number of people to personalize it for a specific person and generate a personalized model that reflects that specific person's work knowledge, The system includes a means for making the personalized model generated by the personalization means available for use as a substitute for the specific person, The personalization model is capable of providing the user with responses that reflect the work knowledge of the specific person, The aforementioned availability means is a service provision system having a function that enables the user to receive a response using natural language processing.

2. A method for producing a service delivery system that enables the provision of services using machine learning models, The first step involves using a standard model modeled by machine learning with training data collected from a large number of people to generate a personalized model that reflects the work knowledge of a specific person, and then personalizing it for that person. The second step includes making the personalized model generated in the first step available for use as a substitute for the specific person, The personalization model is capable of providing the user with responses that reflect the work knowledge of the specific person, The second step is a method for producing a service provision system that enables the user to receive a response using natural language processing.

Images