System and methods for network native artificial intelligence (AI) task execution control
The described method addresses the lack of control signaling in NWDAF by implementing AI task control policies and direct connections for real-time management of AI tasks, enhancing the efficiency of AI model training and inference in network-native AI service.
Patent Information
- Application Number
- PCT/CN2024/127107
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2024-07-17
- Filing Date
- 2024-10-24
- Publication Date
- 2026-01-22
AI Technical Summary
Existing 5G Network Data Analytics Function (NWDAF) lacks detailed control signaling and procedures for AI model-based tasks, leading to inefficiencies in AI model training and inference processes, particularly in network-native AI service (NET4AI) task execution control.
A computer-implemented method for controlling AI task executions, involving a policy control function that generates and implements AI task control policies, and a task control function that establishes direct connections with participant network resources for real-time control of AI tasks, including model training and inference.
Enhances the control of AI tasks in network-native AI service by providing detailed control parameters and real-time management, improving the efficiency and effectiveness of AI model training and inference processes.
Smart Images

Figure CN2024127107_22012026_PF_FP_ABST
Abstract
Description
SYSTEM AND METHODS FOR NETWORK NATIVE ARTIFICIAL INTELLIGENCE (AI) TASK EXECUTION CONTROL
[0001] CROSS-REFERENCE TO RELATED APPLICATION
[0002] The present application claims priority to United States Provisional Patent Application No. 63 / 672, 419 filed on July 17, 2024, the contents of which are incorporated herein by reference.TECHNICAL FIELD
[0003] The present application relates to artificial intelligence (AI) and, more particularly, to a system and methods for controlling execution of AI jobs / tasks in native-AI network architecture.BACKGROUND
[0004] In 3GPP R19, AI model training and inference operations are supported by the 5G Network Data Analytics Function (NWDAF) . NWDAF leverages AI models for data collection and utilizes collected data for network analytics. A model training logical function (MTLF) in NWDAF trains machine learning models and exposes new training services, while an analytics logical function (AnLF) performs inference and provides analytic results based on the trained models.
[0005] While NWDAF supports the training and inference functionalities of AI models, there are no detailed control signaling and / or procedures defined for AI model-based tasks, either within NWDAF or between NWDAF and other network functions.
[0006] In the context of network-native AI service (i.e., NET4AI) and mission management, the signaling / procedures related to controlling AI model training and inference processes are supported by general task execution control signaling / procedures. An independent network function, the task control function (TCF) , controls the execution of tasks related to AI model training and inference. Task control for NET4AI suffers from a lack of detailed control parameters and a specification of the effects (to the AI model training / inference processes) of parameters defined for the AI task execution control.
[0007] A mission manager generally serves as an intermediate controller between NET4AI participants and NET4AI TCF during AI task execution control. A NET4AI participant may be, for example, a network device with local resources / capability to execute steps involved in an AI task (e.g., model (partition) training, model (partition) inference, training / inference data provision, etc. ) or a network function. During the execution of an AI task, the NET4AI participant conducts data exchange with target data plane functions (e.g., processing service functions, or PSFs) that are managed by the TCF.SUMMARY
[0008] In an aspect, the present disclosure describes a computer-implemented method for controlling AI task executions. The method may include: receiving, by a policy control function, a first request to generate artificial intelligence (AI) task control policies in connection with a target AI task; in response to receiving the first request, generating, by the policy control function, one or more first AI task control policies configured for controlling execution of the target AI task; and controlling, in real-time, execution of the target AI task by a participant network resource based on the one or more first AI task control policies.
[0009] In some implementations, the first request may comprise a request message that indicates task information associated with the target AI task and quality-of-service (QoS) requirements for AI task control.
[0010] In some implementations, the target AI task may be one of AI model training task or AI model inference task.
[0011] In some implementations, the method may further include: sending, by the policy control function to a mission manager, the one or more first AI task control policies, wherein controlling execution of the target AI task by the participant network resource comprises causing the mission manager to: select at least one participant network resource for the target AI task; and conduct real-time control of the selected participant network resource in executing steps of the target AI task based on the one or more first AI task control policies.
[0012] In some implementations, conducting real-time control of the selected participant network resource may include transmitting, by the mission manager to each selected participant network resource, control messages for configuring the selected participant network resource and for executing steps of the target AI task.
[0013] In some implementations, each of the first AI task control policies may include AI task control rules and wherein the AI task control rules comprise one or more of: data feeding control rules; AI model training control rules; or AI model inference control rules.
[0014] In some implementations, the method may further include: requesting, by the policy control function, a mission manager to associate at least one participant network resource with the target AI task; selecting, by the missioner manager, the at least one participant network resource for executing the target AI task; receiving, by the policy control function, participant information of the at least one selected participant network resource; and transmitting, by the policy control function, per-participant AI task control policies to each of the selected participant network resources.
[0015] In some implementations, requesting the mission manger to associate the at least one network resource with the target AI task may include transmitting, by the policy control function, to the mission manager a participant association request that includes participant selection rules.
[0016] In some implementations, the method may further include: receiving, by the policy control function from each of the selected participant network resources, a participant execution status indicator (or PSF / participant execution status information) representing progress of executing the target AI task; and updating, by the policy control function, the per-participant AI control policies based on the participant execution status indicators.
[0017] In another aspect, the present disclosure describes a computer-implemented method for controlling AI task executions. The method may include: receiving, by a mission manager from a task control function, a first request to identify a participant network resource for executing a target AI task, the first request indicating participant requirements for the target AI task; in response to receiving the first request, selecting, by the mission manager, at least one participant network resource based on the participant requirements; establishing, by the task control function, a direct connection between the task control function and the at least one selected participant network resource; and performing, by the task control function via the connection, synchronization of task control parameters for the target AI task with the at least one selected participant network resource.
[0018] In some implementations, the first request may include task information of the target AI task and interface information of available interfaces for establishing the connection between the task control function and the at least one selected participant network resource.
[0019] In some implementations, the method may further include: subsequent to performing the synchronization, triggering, by the task control function, execution of the target AI task by the at least one selected participant network resource; and conducting, by the task control function, real-time control of the at least one selected participant network resource in executing steps of the target AI task.
[0020] In some implementations, the method may further include: receiving, by the mission manager, a registration request to register a first participant network resource for executing AI tasks; in response to receiving the registration request, generating, by the mission manager, a participant profile for the first participant network resource, wherein the at least one participant network resource is selected based on the participant requirements and profile information of participant profiles associated with one or more registered participant network resources.
[0021] In some implementations, the method may further include: receiving, by the task control function from the at least one selected participant network resource, PSF / participant execution status associated with the target AI task; in response to receiving the PSF / participant execution status: determining, by the task control function, task execution control operations; and sending, by the task control function to the at least one selected participant network resource, a task execution control request during execution of the target AI task.
[0022] In some implementations, the task execution control request may include at least one of: data transmission control parameters; resource management control parameters; AI model training control parameters; or AI model inference control parameters.
[0023] In another aspect, the present disclosure describes a computer-implemented method for controlling AI task executions. The method may include: establishing, by a task control function, a direct connection between the task control function and at least one participant network resource selected for executing a target AI task; sending, by the task control function to the at least one selected participant network resource, execution status information of the target AI task; and sending, by the at least one selected participant network resource to the task control function, a task execution control request during execution of the target AI task.
[0024] In some implementations, the execution status information of the target AI task may include at least one of: data transmission information; AI model training status information; AI model inference status information; or available target data plane function information for the at least one selected participant to access.
[0025] In some implementations, the method may further include forwarding, by the task control function, parameters of the task execution control request to a target data plane function.
[0026] In some implementations, the method may further include: receiving, by a mission manager from the task control function, a first request to identify a participant network resource for executing the target AI task, the first request indicating participant requirements for the target AI task; in response to receiving the first request, selecting, by the mission manager, the at least one participant network resource based on the participant requirements.
[0027] In some implementations, the method may further include sending, by the mission manager to the task control function, participant information of the at least one selected participant network resource, the participant information including associated interfaces for connecting with the at least one participant network resource.
[0028] Other aspects and features of the present application will be understood by those of ordinary skill in the art from a review of the following description of examples in conjunction with the accompanying figures. Example embodiments of the present application are not limited to any particular operating system, system architecture, mobile device architecture, server architecture, or computer programming language.
[0029] In the present application, the term “participant network resource” refers to a network resource that executes, or causes to be executed, one or more network-based AI tasks. A participant network resource may, for example, comprise a network device (e.g., UE) or a network function (NF) that is configured to execute steps involved in an AI computing task.
[0030] In the present application, the term “and / or” is intended to cover all possible combinations and sub-combinations of the listed elements, including any one of the listed elements alone, any sub-combination, or all of the elements, and without necessarily excluding additional elements.
[0031] In the present application, the phrase “at least one of …or…” is intended to cover any one or more of the listed elements, including any one of the listed elements alone, any sub-combination, or all of the elements, without necessarily excluding any additional elements, and without necessarily requiring all of the elements.BRIEF DESCRIPTION OF THE DRAWINGS
[0032] Embodiments are described in detail below, with reference to the following drawings:
[0033] FIG. 1 is a schematic diagram illustrating a conceptual structure of an example 6G system;
[0034] FIG. 2 illustrates an example deployment of a 6G system in accordance with the present disclosure;
[0035] FIG. 3 is a high-level schematic diagram of an example apparatus in a communications system;
[0036] FIG. 4 is a sequence diagram illustrating an exemplary workflow of a task execution control process in accordance with the present disclosure;
[0037] FIG. 5 is a sequence diagram illustrating another exemplary workflow of a task execution control process in accordance with the present disclosure;
[0038] FIG. 6 is a sequence diagram illustrating an exemplary workflow of a task execution control process in accordance with the present disclosure;
[0039] FIG. 7 is a sequence diagram illustrating steps for providing PSF / participant execution status updates in NET4AI-triggered AI task execution control procedure;
[0040] FIG. 8 is a sequence diagram illustrating an exemplary workflow of a task execution control process in accordance with the present disclosure;
[0041] FIG. 9 is a sequence diagram illustrating an exemplary workflow of a task execution control process in accordance with the present disclosure; and
[0042] FIG. 10 is a sequence diagram illustrating an exemplary workflow of a task execution control process in accordance with the present disclosure.
[0043] Like reference numerals are used in the drawings to denote like elements and features.DETAILED DESCRIPTION
[0044] The present application discloses a system and methods for controlling AI task executions, such as machine learning training and inference, conducted by NET4AI TCF and participants. In particular, the present application contemplates a number of AI task control scenarios, including:
[0045] 1. PCF-enabled AI task execution control. TCF is implemented as part of policy control function (PCF) . AI task execution control can be conducted by providing AI task control policies to PSF / participants through a policy configuration procedure. Two options are defined: 1) enable AI task execution control by PCF providing AI task control policies to mission manager through a policy configuration procedure; and 2) enable AI task execution control by PCF configuring or updating AI task control policies to PSF / participants directly. This alternative approach to AI task execution control (i.e., through policy configuration) enables PCF to control AI model training and inference processes effectively.
[0046] 2. NET4AI-triggered AI task execution control. TCF has direct control plane (CP) connection with participants, and can directly control the AI tasks executed on participant (s) and PSF (s) . This approach to AI task execution control does not engage a mission manager as the relay / translator of AI task control messages. Detailed parameters for AI task execution control (e.g., content of parameter, effects on PSF / participant execution caused by the parameter) are protected in this approach.
[0047] 3. Participant-triggered AI task execution control. With the direct connection between a participant and the TCF, the participant can send AI task execution control requests to TCF for customizing the steps of AI tasks executed in network. This approach to AI task execution control does not engage mission manager as the relay / translator of AI task control messages. Detailed parameters for AI task execution control (e.g., content of parameter, effects on PSF execution caused by the parameter) are protected in this approach.
[0048] Reference is first made to FIG. 1 which illustrates a conceptual structure of an example 6G system. The present application discloses an evolutionary solution for a 6G system architecture and procedure design. The proposed 6G network architecture represents enhancements of a 5G system and is designed with several fundamental principles and requirements, including: openness, trustworthiness, simplicity in standardization, scalability, rapid deployment of 6G networks and future-proofing. The proposed 6G network architecture design applies modularization strategy and utilizes service-based (XaaS) concepts and network virtualization techniques.
[0049] A procedure of the 6G system of FIG. 1 may include some procedures that can be reused by other procedures. Such a reusable procedure is defined as a “basic” procedure. A complex procedure can include multiple sequential or parallel basic procedures. It is expected that such methodology can simplify designs of procedures. The 6G system leverages a service-based architecture and XaaS concepts.
[0050] XaaS services in the 6G system may be categorized into three layers. The “infrastructure layer” includes infrastructures supporting 6G services. Among them are wireless networks (Radio Access Network (RAN) , Core Network (CN) ) infrastructures, cloud / data center infrastructures, satellite networks, storage / database infrastructures, and sensing networks, etc. These infrastructures can be provided by a single provider or by multiple providers.
[0051] Each XaaS service may be provided by an identified 5G logical function. As shown in FIG. 1, the “service layer” of the 6G system conceptual structure may include:
[0052] Network for AI (NET4AI) is a new type of service in 6G CN / RAN which enables a network to conduct or execute AI model training and inference task (s) by network-based computing and communications resources. As will be described in greater detail below, the proposed system supports NET4AI services by introducing enhancements of the network data analytics function (NWDAF) of a 5G system.
[0053] A NET4Data service provides a de-centralized architecture for data stakeholders to collaboratively manage data lifecycle events. These data lifecycle events include data storage and data sharing. The data could be public, private, sensitive, and / or confidential. In the present application, the NET4Data service could be integrated into the 5GS, or could be enhanced by the 5GS.
[0054] Data analysis and management (DAM) focuses on different types of data: network data (e.g., data collected from network functions, XaaS service) , ISAC data (3GPP-based sensing data (e.g., from user equipment (UE) and RAN) , non-3GPP-based sensing data (e.g., from Radar, LiDAR, Wi-Fi Sensing) ) , sensor data (e.g., data from camera sensor, video sensor, etc. ) , and other data (e.g., digital user data, 3rd party data, synthetization data, and AI data) . DAM provides services for a variety of data consumers, e.g., XaaS service, third-party, NF, UE, etc. 5G system logical functions, such as NWDAF, DCCF, and MFAF of control plane, can be enhanced to support DAM services in the proposed system.
[0055] Network for digital world (NET4DW) as a service provides the capability of intelligent integration / synthesis of information from the physical world and digital world. Customers of NET4DW can be individuals, industries, or governments. The customers have the capability of creation, control, and management of a variety of applications running in the digital world such as virtual reality applications. Digital world services can be supported by enhancing 5G functions and adding new functions where necessary.
[0056] Network for connectivity (NET4CON) as a service provides a capability to support exchange of messages and data among new 6G services. The basic capabilities of NET4CON include managing logical topology among XaaS services and between 6G XaaS services and all types of 6G system customers, to introduce intelligent GWs for controlling dynamic forwarding based on configured procedure principle and to support anonymous interactions among these XaaS services and customers by the introduced intelligent GWs. The NET4CON service is provided via enhancements of 5G system.
[0057] The “control / management layer” of FIG. 1 may include the following services:
[0058] Mission management (MM) as a service provides a capability to program provisioning of XaaS services at service layer to provide mission services. A mission is to achieve a designated goal (i.e., mission goal) which includes providing PDU connectivity and optionally providing data processing. The MM services may include: mission information management service, mission session management service, and mission execution and access management service.
[0059] Resource management (RM) as a service provides a capability of life-cycle management of a variety of slices and over-the-air resource assignment to wireless devices.
[0060] Service provisioning management (SPM) as a service provides a capability of control and management of 6G service access by customers and provisioning of requested services. The capability may be provided by ID management, unified authentication, and anonymous service authorization and key management.
[0061] Connectivity management (CM) as a service provides a capability of reachability management of 6G wireless devices and D-Users in NET4DW in order to support connectivity establishment between wireless devices / D-Users and XaaS services of 6G System. The physical locations of D-Users can be changed. A CM service can be deployed across multiple BAS domains.
[0062] Protocol as a service provides a capability to design service customized protocol stacks for identified interfaces.
[0063] FIG. 2 shows an example deployment of a 6G system in accordance with disclosed embodiments of the present application. The “+” illustrated in FIG. 2 represents “enhanced” ; for example, the 5G AMF-Mobility function is enhanced, denoted as AMF-Mobility+, the 5G RRC function is enhanced, denoted as RRC+, the 5G Network Repository Function (NRF) is enhanced, denoted as NRF+, the 5G Session Management Function (SMF) is enhanced, denoted as SMF+, the 5G Network Exposure Function (NEF) is enhanced, denoted as NEF+, the 5G Authentication Server Function (AUSF) is enhanced, denoted as AUSF+ (other enhanced functions are not described in detail herein) .
[0064] C / M Radio Bearer (C / M RB) of a 6G device comprises over-the-air connection for carrying control signaling for over-the-air interface management and C / M plane messages. A 6G device can have multiple C / M RBs.
[0065] Data Radio Bearer (Data RB) of a 6G device comprises over-the-air connection for carrying Data plane traffic. A 6G device can have multiple Data RBs. RB endpoint comprises an endpoint of an RB at network side. An endpoint of an RB protocol stack (e.g., PDCP) can be in, e.g., a RAN BAS domain, but not limited to. In other words, an RB endpoint can be flexibly deployed / selected for a device.
[0066] RB handler comprises over-the-air interface protocol stack handler. An RB handler is defined as a logical function which perform RB protocol stack operations after getting configurations. A protocol handler is PDCP-only handler or whole protocol stack handler. An RB handler accepts RB configuration from Connectivity Management (CM) service. An RB handler also accepts security configuration, e.g., keying material, from Service Provisioning Management (SPM) service.
[0067] The NET4CON service, a main service impacting on 6G system architecture, is implemented by enhanced 5G Service Communication Proxy (SCP+) as C / M plane GW and enhanced 5G User Plane Function (UPF+) as data plane GW. Proposed per device / D-User C / M session and data session are defined as logical connection between a device / D-User and its serving SCP+(C / M-TW-GW) and serving UPF+ (Data-TW-GW) . All XaaS services are deployed across multiple BAS / clouds.
[0068] The 6G customer can be of various types, such as a device (e.g., electronic device, terminal device) , an apparatus, a chip, an equipment (e.g., user equipment) , etc. For example, the customer may be an individual customer, a business customer, and the like. The 6G customer is used to connect persons, objects, machines, etc. The 6G customer may be widely used in various scenarios including, for example, cellular communications, device-to-device (D2D) , vehicle-to-everything (V2X) , peer-to-peer (P2P) , machine-to-machine (M2M) , MTC, internet of things (IoT) , virtual reality (VR) , augmented reality (AR) , mixed reality (MR) , metaverse, digital twin, industrial control, self-driving, remote medical, smart grid, smart furniture, smart office, smart wearable, smart transportation, smart city, drones, robots, remote sensing, passive sensing, positioning, navigation and tracking, autonomous delivery and mobility, etc.
[0069] Each 6G customer represents any suitable end user device for wireless operations and may include such devices (or may be referred to but not limited to) as a user equipment (UE) or a user device or a terminal device, a wireless transmit / receive unit (WTRU) , a mobile station, a fixed or mobile subscriber unit, a cellular telephone, a station (STA) , a MTC device, a personal digital assistant (PDA) , a smartphone, a laptop, a computer, a tablet, a wireless sensor, a consumer electronics device, a smart book, a vehicle, a car, a truck, a bus, a train, or an IoT device, wearable devices (such as a watch, a pair of glasses, head mounted equipment, etc. ) , an industrial device, or an apparatus in (e.g. module, modem, or chip) or comprising the forgoing devices, among other possibilities. Future generation 6G customers may be referred to using other terms. When a 6G customer performs (or is configured to perform) a method described herein, it may be interpreted as the electronic device, one or more module (or units) in the electronic device, a circuit or chip, or a combination thereof, may perform the method. For example, the circuit or chip may include a modem chip, also referred to as a baseband chip, a system on chip (SoC) including a modem core, or system in package (SIP) ) , and the like, and may be responsible for one or more communication functions in the electronic device.
[0070] FIG. 3 illustrates an example apparatus 320 of a communications system (e.g., the 6G system in FIG. 2) . The apparatus 320 may be an electronic device (e.g. electronic device or other 6G customer) , a network node such as RAN, any components in RAN, CN or any Network Function of CN. As shown in FIG. 3, apparatus 320 may include at least one processor 260. Only one processor 260 is illustrated to avoid congestion in the drawing. The processor 260 may perform (or control the apparatus 320 to perform) operations (or methods) described herein as being performed by the apparatus 320.
[0071] When the apparatus is RAN, components of the RAN or the apparatus is the UE, the apparatus 320 may further include a transmitter 252 and a receiver 254 coupled to one or more antennas. One, some, or all of the antennas may alternatively be panels. The transmitter 201 and the receiver 203 may be integrated, e.g. as a transceiver. The transceiver is configured to modulate data or other content for transmission by at least one antenna or network interface controller (NIC) . The transceiver is also configured to demodulate data or other content received by the at least one antenna. Each transceiver includes any suitable structure for generating signals for wireless or wired transmission and / or processing signals received wirelessly or by wire. Each antenna includes any suitable structure for transmitting and / or receiving wireless or wired signals. In present disclosure, the transceiver (or transmitter 252 and / or receiver 254) may be viewed as an interface circuit.
[0072] The apparatus 320 may include at least one memory 258. The memory 258 stores instructions used to perform operations described herein. The memory 258 may also stores data used, generated, or collected by the apparatus 320. For example, the memory 258 could store software instructions or modules configured to implement some or all of the functionality and / or embodiments described herein and that are executed by one or more processor 260.
[0073] It will be understood by a person skilled in the art that embodiments of this application may be provided as a method, an apparatus (or system) , computer-readable storage medium, or a computer program product. Therefore, this application may describe a form of a hardware-only embodiment, a software-only embodiment, or an embodiment with a combination of software and hardware. Moreover, this application may describe a form of a computer program product that is implemented on one or more computer-usable storage media (including but not limited to a disk memory, an optical memory, and the like) that include computer-usable program code.
[0074] Reference is now made to FIG. 4 which is a sequence diagram illustrating a workflow of an exemplary task execution control process 400. Specifically, FIG. 4 shows an artificial intelligence (AI) task execution control procedure that is enabled by a policy control function (PCF) . The PCF supports taking input from NWDAF into consideration for policies on assignment of network resources and for traffic steering policies. The AI task execution control is enabled by the PCF providing AI task control policies to a mission manager. The mission manager then conducts AI task execution control in accordance with the configured AI task control policies.
[0075] As shown in FIG. 4, the PCF may implement NET4AI task control function (TCF) functionalities. In operation 402, the PCF receives, from a NET4AI user, a first request (i.e., an AI task control policy request) . The first request may be in the form of a request message sent by the NET4AI user for requesting creation of AI task control policies for a target AI task. The target AI task may, for example, be training of an AI model or inference based on a trained model. In at least some implementations, the request message includes target AI task information such as one or more of: ID / name of the target AI task; associated mission ID / name; and an indication of allowed / no device participation for the target AI task.
[0076] Additionally, or alternatively, the request message may include quality-of-service (QoS) requirements for the AI task control. In some implementations, the request message may provide indications of QoS data feeding control such as: minimum / maximum volume; data source requirements (e.g., type / format of data) ; data distribution variation range; minimum / maximum data feeding speed and frequency. If the target AI task is a model training task, the request message may indicate QoS requirements for model training control. For example, the request message may specify convergence criteria (e.g., maximum number of training episodes, accuracy / loss threshold, etc. ) and maximum training time. If the target AI task is a model inference task, the request message may indicate QoS requirements for model inference control. For example, the request message may specify, among others, maximum inference time and inference accuracy threshold.
[0077] In operation 404, the PCF sends, to the NET4AI user, an AI task control policy request response message. The message signals to the NET4AI user that the AI task control policy request was received by the PCF. The operation 404 is optional; in some implementations, the PCF may not send a response message for acknowledging receipt of the AI task control policy request.
[0078] The PCF generates AI task control policies responsive to receiving the AI task control policy request, in operation 406. In particular, the NET4AI TCF functionality in PCF generates the AI task control policies. Each generated policy includes one or more of: ID / name of the target AI task; ID / name of an associated mission; ID / name of the PCF. Additionally, or alternatively, a generated policy may include policy validation information, such as valid location information (e.g., within a RAN / CN, within a city / region, etc. ) and valid time information (e.g., within a few hours or days) , as well as PSF / participant selection rules. The selection rules may indicate, for example, type of PSF / participant (e.g., support GPU / CPU, wireless / wire interface, etc. ) .
[0079] In at least some implementations, a generated policy may also include AI task control rules. The AI task control rules may include data feeding control rules such as, but not limited to. data type / format, minimum data volume from each data source, data filtering criteria (e.g., discard data if exceeding a defined variation range) , minimum and / or maximum data feeding speed and frequency. If the target AI task is a model training task, the generated policy may include AI model training control rules such as convergence criteria (e.g., maximum number of training episodes, accuracy / loss threshold, etc. ) , maximum training time, maximum episode running time, minimum computing / communications resource requirements for training, and dropout rate indicator. If the target AI task is a model inference task, the generated policy may include AI inference control rules such as maximum inference time, minimum inference accuracy threshold, and minimum computing / communications resource requirements for inference.
[0080] The AI task control rules that are included in a generated policy may each have defined effects on a target AI task during its execution. In this way, the AI task control rules serve as rules governing acceptable execution of AI tasks. Table 1 below lists the effects of various AI task control rules that may be included in AI task control policies.
[0081] Table 1
[0082] In operation 408, the PCF requests a mission manager (MM) to create a policy association for the target AI task. More particularly, the generated AI task control policies are sent by the PCF to the MM that manages execution of the target AI task, in operation 410.
[0083] The generated AI task control policies are included in an AI task control policy notification. The MM sends an AI task control policy notification response to the PCF in order to notify the PCF regarding receipt of the generated AI task control policies (operation 412) .
[0084] In operation 414, the MM selects and configures PSF / participant (s) for the target AI task in accordance with the received AI task control policies. During execution of the target AI task, the MM conducts real-time control of participant network resource (e.g., PSF / participant) execution based on the AI task control policies.
[0085] Reference is now made to FIG. 5 which is a sequence diagram illustrating a workflow of another exemplary task execution control process 500. Specifically, FIG. 5 shows an AI task execution control procedure that is enabled by a policy control function (PCF) . The PCF is configured to directly update AI task control policies to PSF / participants. The operations of process 500 may be performed in addition to, or as alternatives of, one or more of the operations of process 400 of FIG. 4.
[0086] As shown in FIG. 5, the PCF may implement NET4AI task control function (TCF) functionalities. In operation 502, the PCF receives, from a NET4AI user, a first request, i.e., AI task control policy request. The first request may be in the form of a request message sent by the NET4AI user for requesting creation of AI task control policies for a target AI task. The target AI task may, for example, be training of an AI model or inference based on a trained model. In at least some implementations, the request message includes target AI task information such as one or more of: ID / name of the AI task; associated mission ID / name; and an indication of allowed / no device participation for the AI task. For details of this first request, refer to the description of “first request” in the implementation of FIG. 4 above.
[0087] In operation 504, the PCF sends, to the NET4AI user, an AI task control policy request response message. The message signals to the NET4AI user that the AI task control policy request was received by the PCF. The operation 504 is optional; in some implementations, the PCF may not send a response message for acknowledging receipt of the AI task control policy request
[0088] The PCF generates AI task control policies responsive to receiving the AI task control policy request, in operation 506. In particular, the NET4AI TCF functionality in PCF generates the AI task control policies.
[0089] In operation 508, the PCF requests a mission manager (MM) to provide information of PSF / participants that will execute the target AI task. Specifically, the PCF transmits, to the MM, a request (i.e., a participant association request) to associate PSF / participants with the target AI task. In at least some implementations, the participant association request includes one or more of: ID / name of the target AI task; ID / name of the associated mission; policy validation information such as valid location and time information; data source selection rules (e.g., supported data type / format, minimum data volume from each data source, minimum / maximum data feeding speed and frequency) ; and participant selection rules. The participant selection rules may specify type of PSF / participant (e.g., support GPU / CPU, wireless or wire interface) , minimum computing or communications resource requirements for training and / or inference.
[0090] The MM selects and configures participant network resource (e.g., PSF / participants (including data sources) ) for executing the target AI task. The MM then provides, to the PCF, a response to the participant association request, in operation 510. The response includes participant information of the at least one selected participant network resource. The participant information is associated with the selected PSF / participant information and may include one or more of: ID / name of PSF / participants; ID name of the target AI task or mission; information regarding configured steps / functionalities of the target AI task to be executed by the selected PSF / participants (e.g., providing data as a data source, training a partition of AI model, etc. ) ; and available computing or communications resources associated with the PSF / participants. Additionally, or alternatively, the participant information may include information regarding interfaces between PCF and the PSF / participants (C2P) (e.g., IP address, port number, tunnel ID) and between the PSF / participants and other connected PSF / participants for executing the target AI task (P2P) .
[0091] In operation 512, the PCF notifies each selected PSF / participant regarding its associated AI task control policies via a configured C2P interface. That is, the PCF transmits, via the C2P interfaces to the selected PSF / participants, AI task control policy notifications. An AI task control policy notification includes per-participant AI task control policies. The per-participant AI task control policies indicate ID / name of the target AI task, ID / name of the associated mission, and ID / name of the PCF. Further, the AI task control policies include policy validation information, such as valid location and time information, and AI task control rules. For a data source, the AI task control rules may include data feeding control rules (e.g., data type / format, data filtering criteria, and data feeding speed / frequency) . If the AI task being executed is a model training task, the AI task control rules may include model training control rules such as convergence criteria (e.g., maximum number of training episodes, accuracy / loss threshold) , maximum training time, maximum episode running time, minimum computing or communications resources for training, and dropout rate indicator. Alternatively, if the task being executed is a model inference task, the AI task control rules may include model inference control rules such as maximum inference time, minimum inference accuracy threshold, and minimum computing or communications resources for inferencing.
[0092] Each PSF / participant transmits, to the PCF, an AI task control policy notification response to notify the PCF regarding receipt of its associated per-PSF / participant AI task control policies (operation 514) .
[0093] During execution of the target AI task, the selected PSF / participants are controlled by configuring and updating AI task control policies directly. In particular, control messages between the MM and PSF / participants are not needed.
[0094] In operation 516, each PSF / participant sends its PSF / participant execution status indicator (period-based or event-based) to the TCF for monitoring the AI task execution progress. The PSF / participant execution status indicator indicates progress of executing the target AI task. In operation 518, the PCF updates related parameters of each PSF / participant’s per-participant AI task control policies to the PSF / participant based on the received PSF / participant execution status indicator.
[0095] Reference is now made to FIG. 6 which is a sequence diagram illustrating a workflow of another exemplary task execution control process 600. Specifically, FIG. 6 shows an AI task control procedure that is based on establishing a direct connection between NET4AI task control function (TCF) and a NET4AI participant.
[0096] In certain scenarios, a TCF can directly control a target AI task executed on PSF (s) that are managed by it. A direct connection between NET4AI TCF and NET4AI participant is established before the target AI task is executed.
[0097] The AI task participant is first registered. In operation 602, the AI participant sends an AI participant registration request to a mission manager (MM) for participating AI tasks. The registration request message may include, at least, ID / name of the participant, information regarding supported AI tasks (e.g., task ID, type, etc. ) , and available interface (s) to connect with NET4AI TCF and PSF (s) . After receiving the AI participant registration request from the participant, the MM creates an AI participant profile for the participant (operation 604) and stores the information included in the registration request message in the profile. The MM then notifies the AI participant regarding the registration via an AI participant registration response, in operation 606. In particular, the MM indicates to the AI participant that an associated participant profile has been created.
[0098] After the participant is registered, a TCF-to-participant (T2P) connection is established. In operation 608, the TCF transmits a first request (i.e., AI participant request) to the MM in order to identify participant network resource (participant (s) ) for a target AI task. The AI participant request message may include one or more of: task information of the target AI task (e.g., ID / name of the AI task) , requirements for the participant (e.g., allowed participant ID, type, location, etc. ) , and available interface (s) to connect with participants on both CP (T2P connection) and DP (PSF-to-participant connection) .
[0099] Upon receiving the AI participant request message, the MM selects one or multiple participants for the target AI task (operation 610) . The participants are selected based on the various participant requirements indicated in the AI participant request and / or the AI participant profiles that are maintained by the MM.
[0100] In operation 612, the MM notifies each of the selected AI task participants regarding the target AI task. Specifically, the MM transmits a participating point notification message to each selected participant. The message includes, among other information, ID / name of target AI task or ID / name of the interfaces (e.g., IP address, port number, tunnel ID) for use in connecting with the TCF and target PSF (s) .
[0101] Each selected AI task participant notifies the MM regarding receipt of participating point notification (operation 614) . The selected participant then monitors the interfaces (referred to in the participating point notification) for messages from the TCF and PSF (s) .
[0102] In operation 616, the MM provides, to the TCF, information regarding the selected AI participant (s) for the target AI task. In particular, the MM transmits an AI participant request response message to the TCF. The message includes one or more of: ID / name of the target AI task, ID / name of each selected AI participant and associated interfaces (e.g., IP address, port number, tunnel ID) to connect with the participant. The interface information is used for establishing a connection between the task control function and the selected participant. Specifically, the task control function establishes a direct connection with the selected participant, for example, by connecting using the participant information (including associated interfaces) , and vice versa. The connection between the task control function and the selected participant enables direct exchange of control messages without an intermediate relay.
[0103] Using the established T2P connection, the AI participant and the TCF can synchronize the AI task control parameters enabled between them during the AI task execution (operation 618) . After synchronizing the control parameters, the TCF may control, in real-time, execution of the target AI task, and the AI task execution control can be conducted between the TCF and the selected AI participants, in operation 620.
[0104] As shown in FIG. 7, during the AI task execution, a PSF can provide a PSF execution status (period-based or event-based) to the TCF (operation 702) , which facilitates monitoring progress of the AI task execution. Similarly, a participant can provide a participant execution status to the TCF (operation 706) . In some implementations, the TCF may perform the optional steps (operations 704 and 708, respectively) of sending a PSF or participant execution status response for acknowledging receipt of the PSF / participant execution status, respectively.
[0105] The PSF / participant execution status may include one or more of: ID / name of the associated AI task and ID / name of the PSF / participant. Additionally, or alternatively, the execution status may indicate: data transmission information (for input and output) , e.g., accumulated data volume, accumulated data distribution / variation, static / real-time data sending speed / frequency (from each data source or all data sources) ; static / real-time communications resource information on associated link (s) , e.g., packet loss ratio, wireless channel quality indicator, bandwidth consumption ratio, data traffic throughput; static / real-time computing / storage resource information, e.g., GPU / CPU consumption volume and / or ratio, storage / RAM consumption volume and / or ratio; and / or real-time location information of the participant.
[0106] If the target AI task is a model training task, the PSF / participant execution status may include AI model training status information, e.g., number of past iteration / episode (s) , accumulated training execution time, accumulated training time of the on-going episode, convergence level of the embedded model (partition) (e.g., variation of weights between two episodes) , achieved training accuracy / lost level.
[0107] If the target AI task is a model inference task, the PSF / participant execution status may include AI model inference status information, e.g., accumulated and / or real-time inference accuracy, inference time of the on-going inferencing, and accumulated inference execution time.
[0108] FIG. 8 illustrates control request operations in the context of NET4AI triggered AI task execution control. The NET4AI TCF receives PSF / participant execution status (operations 802 and 808, respectively) , and determines detailed AI task execution control operations based on the received PSF / participant execution status information. The TCF then sends a PSF / participant execution control request to the designated PSF / participant (s) during execution of the target AI task (operations 804 and 810, respectively) . The PSF / participant execution control request includes, at least, ID / name of the associated AI task and ID / name of the designated PSF / participant. Additionally, or alternatively, the PSF / participant execution control request may include data transmission controls. For example, the PSF / participant execution control request may include control signals to: start / stop forwarding data or receiving input data (i.e., PSF / participant receiving this parameter will start / stop its data forwarding / receiving) ; adjust data sending speed and / or frequency (i.e., PSF / participant receiving this parameter will change its data forwarding speed / frequency to the value indicated in the parameter. By this control parameter, the TCF can constrain the data transmission speed / frequency among data-source / PSF / participants within a pre-configured range) ; and change source / destination PSF / participant / network functions for data transmission (i.e., PSF / participant receiving this parameter will change its data forwarding / receiving target to the new PSF / participant indicated in the parameter) . The PSF / participant may optionally send a PSF / participant execution control request response (operations 806 and 812, respectively) for acknowledging receipt of the task execution control requests.
[0109] In some implementations, the PSF / participant execution control request may include resource management controls, e.g., assign / release computing / communications resources to the executing AI task. In this way, the PSF / participant may allocate / release the required quantity of resources indicated in the parameter to executing the AI task.
[0110] If the target AI task is a model training task, the PSF / participant execution control request may include AI model training controls. The AI model training controls may include control signals to: pause / resume AI model training (FP / BP) (i.e., PSF / participant will stop / resume executing of the AI model training steps configured on it) ; start / stop model training episode (i.e., PSF / participant will start executing a new training episode and discard the on-going training episode, or stop the execution of on-going training episode) ; freezing or de-freezing model (partition) (i.e., PSF / participant will not update the weights / bias of all neurons / links in a model / model-partition in the following training iterations until receiving the de-freezing signal (freezing effects) , or PSF / participant will resume updating the frozen weights / bias of all neurons / links in a model / model-partition in the following training iterations (de-freezing effects) ) .
[0111] The AI model training controls may additionally, or alternatively, include control signals to: customize model (partition) (i.e., PSF / participant will cut-off / mute specific neurons / links in a model / model-partition. The specific neurons / links are specified in the parameters. The cut-off / mute of a neuron / link in an AI model means set the weight / bias of the neuron / link into zero and do not update the weight / bias of the neuron / link during the AI model training) ; update dropout rate indicator (i.e., for each iteration in a training episode, the PSF / participant may randomly mute the weight / bias update on a neurons / links with the probability derived from the updated dropout rate indicator. This dropout mechanism can reduce overfitting in AI model training. ) .
[0112] If the target AI task is a model inference task, the PSF / participant execution control request may include AI inference controls. The AI inference controls may include control signals to: start / stop / resume AI model inference (i.e., PSF / participant will start / stop / resume executing of the AI model inferencing steps configured on it) ; activate or deactivate early-exit inferencing (PSF / participant will execute early-exit inferencing which only inferencing through partial of the AI model and output the intermediate result as early-exit inferencing result (activate effects) , or PSF / participant executing the early-exit inferencing will turn to execute full inferencing which inferencing through the complete AI model (de-activate effects) ) ; activate or deactivate accumulative inferencing (PSF / participant will output accumulative inferencing result which is calculated from multiple inferencing results from previous rounds or other models (activate effects) , or PSF / participant executing the accumulative inferencing will turn to execute full inferencing (de-activate effects) ) ; and / or customize / augment inferencing data samples (e.g., prompts for LLM inferencing) (i.e., PSF / participant will customize the input data structure / contents for each inferencing round, e.g., changes the prompts contents, augment the prompts from multiple contexts received from multiple data sources) .
[0113] An established connection between the TCF and a NET4AI participant can enable the participant to trigger AI task execution control. Referring to FIG. 9, during the AI task execution, the TCF can send NET4AI-side execution status information (period-based or event-based) to the participant for monitoring the AI task execution progress conducted by the NET4AI-side (operation 902) . The NET4AI-side execution status information may include: ID / name of the associated AI task; data transmission information (input to / received from the participant) , e.g., accumulated data volume, accumulated data distribution / variation, static / real-time data sending speed / frequency; and available target-PSF information for participant to access. In response to the received NET4AI-side execution status information, the NET4AI participant may send NET4AI-side execution status response (operation 904) to the TCF.
[0114] If the target AI task is a model training task, the NET4AI-side execution status information may include AI model training status information, e.g., number of past iteration / episode (s) , accumulated training execution time, accumulated training time of the on-going episode, convergence level of the embedded model (partition) (e.g., variation of weights between two episodes) , and achieved training accuracy / lost level.
[0115] If the target AI task is a model inference task, the NET4AI-side execution status information may include AI model inference status information, e.g., accumulated and / or real-time inferencing accuracy, inferencing time of the on-going inferencing, and accumulated inferencing execution time.
[0116] The NET4AI participant determines detailed AI task execution control operations based on the received NET4AI-side execution status information. The participant then sends a participant-triggered NET4AI execution control request to the TCF during the AI task execution (FIG. 10 operation 1002) . The participant-triggered NET4AI execution control request may include: ID / name of the associated AI task; ID / name of the target PSF (if applicable) ; and data transmission control request, e.g., start / stop receiving input data, start / stop sending output data (i.e., TCF receiving this parameter will start / stop the data forwarding / receiving between the target PSF and the participant) ; adjust data sending speed / frequency (i.e., TCF receiving this parameter will change the data forwarding speed / frequency from the target PSF to the participant to the value indicated in the parameter) ; change source / destination PSF for data transmission (i.e., TCF receiving this parameter will change the original target PSF to the new target PSF indicated in the parameter) . In response to the received participant-triggered NET4AI execution control request, the TCF may send participant-triggered NET4AI execution control request response (operation 104) to the TCF.
[0117] If the target AI task is a model training task, the participant-triggered NET4AI execution control request may include AI model training control signals to, for example, pause / resume AI model training (FP / BP) (i.e., TCF will control / notify the target PSF to stop / resume executing of the AI model training steps configured on the target PSF) ; start / stop training episode (i.e., TCF will control / notify the target PSF to start executing a new training episode and discard the on-going training episode (if have) , or stop the execution of on-going training episode) ; (de-) freezing model (partition) (i.e., TCF will control / notify the target PSF to not update the weights / bias of all neurons / links in a model / model-partition in the following training iterations until receiving the de-freezing signal (freezing effects) , or TCF will control / notify the target PSF to resume updating the frozen weights / bias of all neurons / links in a model / model-partition in the following training iterations (de-freezing effects) ) ; customize model (partition) (i.e., TCF will control / notify the target PSF to cut-off / mute specific neurons / links in a model / model-partition. The specific neurons / links are specified in the parameters. The cut-off / mute of a neuron / link in an AI model means set the weight / bias of the neuron / link into zero and do not update the weight / bias of the neuron / link during the AI model training) ; update dropout rate (i.e., TCF will control / notify the target PSF to randomly mute the weight / bias update on a neurons / links with the probability derived from the updated dropout rate indicator for each iteration in a training episode) .
[0118] If the target AI task is a model inference task, the participant-triggered NET4AI execution control request may include AI model inference control signals to, for example, start / stop / pause / resume AI model inferencing (i.e., TCF will control / notify the target PSF to start / stop / resume executing of the AI model inferencing steps configured on it) ; (de-) activate early-exit inferencing (i.e., TCF will control / notify the target PSF to execute early-exit inferencing which only inferencing through partial of the AI model, and output the intermediate result as early-exit inferencing result (activate effects) , or TCF will control / notify the target PSF executing the early-exit inferencing to turn to execute full inferencing which inferencing through the complete AI model (de-activate effects) ) ; (de-) activate accumulative inferencing (i.e., TCF will control / notify the target PSF to output accumulative inferencing result which is calculated from multiple inferencing results from previous rounds, or other models (activate effects) , or TCF will control / notify the target PSF to executing the accumulative inferencing to turn to execute full inferencing (de-activate effects) ) ; and customize / augment inferencing data samples (e.g., prompts for LLM inferencing) (i.e., TCF will control / notify the target PSF to customize the input data structure / contents for each inferencing round, e.g., changes the prompts contents, augment the prompts from multiple contexts received from multiple data sources) .
[0119] The methods and / or processes described above, and steps thereof, may be realized in hardware, software or any combination of hardware and software suitable for a particular application. The hardware may include a general-purpose computer and / or dedicated computing device or specific computing device or particular aspect or component of a specific computing device. The processes may be realized in one or more microprocessors, microcontrollers, embedded microcontrollers, programmable digital signal processors or other programmable devices, along with internal and / or external memory. The processes may also, or instead, be embodied in an application specific integrated circuit, a programmable gate array, programmable array logic, or any other device or combination of devices that may be configured to process electronic signals. It will further be appreciated that one or more of the processes may be realized as a computer executable code capable of being executed on a machine-readable medium.
[0120] The computer executable code may be created using a structured programming language such as C, an object oriented programming language such as C++, or any other high-level or low-level programming language (including assembly languages, hardware description languages, and database programming languages and technologies) that may be stored, compiled or interpreted to run on one of the above devices, as well as heterogeneous combinations of processors, processor architectures, or combinations of different hardware and software, or any other machine capable of executing program instructions.
[0121] Thus, in one aspect, each method described above, and combinations thereof may be embodied in computer executable code that, when executing on one or more computing devices, performs the steps thereof. In another aspect, the methods may be embodied in systems that perform the steps thereof and may be distributed across devices in a number of ways, or all of the functionality may be integrated into a dedicated, standalone device or other hardware. In another aspect, the means for performing the steps associated with the processes described above may include any of the hardware and / or software described above. All such permutations and combinations are intended to fall within the scope of the present disclosure.
Claims
1.An AI task execution method, comprising:receiving, by a policy control function, a first request to generate artificial intelligence (AI) task control policies in connection with a target AI task; andin response to receiving the first request, generating, by the policy control function, one or more first AI task control policies configured for controlling execution of the target AI task.2.The method of claim 1, wherein the first request indicates task information associated with the target AI task and quality-of-service (QoS) requirements for AI task control.3.The method of claim 1, wherein the target AI task comprises one of AI model training task or AI model inference task.4.The method of claim 1, further comprising controlling, by a mission manager in real-time, execution of the target AI task by a participant network resource based on the one or more first AI task control policies.5.The method of claim 4, further comprising:sending, by the policy control function to a mission manager, the one or more first AI task control policies,wherein the controlling, by the mission manager, execution of the target AI task by the participant network resource comprises:selecting, by the mission manager, at least one participant network resource for the target AI task; andconducting, by the mission manager, real-time control of the selected participant network resource in executing steps of the target AI task based on the one or more first AI task control policies.6.The method of claim 5, wherein conducting real-time control of the selected participant network resource comprises transmitting, by the mission manager to each selected participant network resource, control messages for configuring the selected participant network resource and for executing steps of the target AI task.7.The method of claim 1, wherein each of the first AI task control policies includes AI task control rules and wherein the AI task control rules comprise one or more of: data feeding control rules; AI model training control rules; or AI model inference control rules.8.The method of claim 7, wherein the data feeding control rules comprise at least one of:minimum or maximum data volume from each data source;source data type or format requirements;data filtering criteria;data distribution variation range; orminimum or maximum data feeding speed and frequency.9.The method of claim 7, wherein the AI model training control rules comprise at least one of:maximum number of training episodes;accuracy or loss threshold;maximum training time;maximum training episode running time;minimum computing resource requirements for training; ordropout rate indicator.10.The method of claim 7, wherein the AI model inference control rules comprise at least one of:maximum inferencing time;minimum inferencing accuracy threshold; orminimum computing resource requirements for inferencing.11.The method of claim 1, further comprising:requesting, by the policy control function, a mission manager to associate at least one participant network resource with the target AI task;selecting, by the missioner manager, the at least one participant network resource for executing the target AI task;receiving, by the policy control function, participant information of the at least one selected participant network resource; andtransmitting, by the policy control function, per-participant AI task control policies to each of the selected participant network resources.12.The method of claim 11, wherein requesting the mission manger to associate the at least one network resource with the target AI task comprises transmitting, by the policy control function, to the mission manager a participant association request that includes participant selection rules.13.The method of claim 11, further comprising:receiving, by the policy control function from each of the selected participant network resources, a participant execution status indicator representing progress of executing the target AI task; andupdating, by the policy control function, the per-participant AI control policies based on the participant execution status indicators.14.An AI task execution method, comprising:receiving, by a mission manager from a task control function, a first request to identify a participant network resource for executing a target AI task, the first request indicating participant requirements for the target AI task;in response to receiving the first request, selecting, by the mission manager, at least one participant network resource based on the participant requirements;establishing, by the task control function, a direct connection between the task control function and the at least one selected participant network resource; andperforming, by the task control function via the connection, synchronization of task control parameters for the target AI task with the at least one selected participant network resource.15.The method of claim 14, wherein the first request includes task information of the target AI task and interface information of available interfaces for establishing the connection between the task control function and the at least one selected participant network resource.16.The method of claim 14, further comprising:subsequent to performing the synchronization, triggering, by the task control function, execution of the target AI task by the at least one selected participant network resource; andconducting, by the task control function, real-time control of the at least one selected participant network resource in executing steps of the target AI task.17.The method of claim 14, further comprising:receiving, by the mission manager, a registration request to register a first participant network resource for executing AI tasks;in response to receiving the registration request, generating, by the mission manager, a participant profile for the first participant network resource,wherein the at least one participant network resource is selected based on the participant requirements and profile information of participant profiles associated with one or more registered participant network resources.18.The method of claim 14, further comprising:sending, by the mission manager to each selected participant network resource, a participating point notification message, the participating point notification message including at least an identifier of the target AI task;receiving, by the mission manager from each selected participant network resource, a confirmation message in reply to the participating point notification message; andsending, by the mission manager to the task control function, participant information of each selected participant network resource for the target AI task, the participant information including one or more of:the identifier of the target AI task; andan identifier of the selected participant network resource and associated interfaces for establishing a connection with the selected participant network resource.19.An AI task execution method, comprising:receiving, by the task control function from at least one participant network resource, participant execution status information associated with target AI task;in response to receiving the participant execution status information:determining task execution control operations based on the participant execution status information;sending, by the task control function to the at least one selected participant network resource, a task execution control request during execution of the target AI task; andexecuting, by the selected participant network resource, the target AI task in response to receiving the task execution control request.20.The method of claim 19, wherein the participant execution status information associated with the target AI task comprises at least one of:data transmission information;AI model training status information;AI model inference status information.21.The method of claim 20, wherein the data transmission information comprises at least one of:accumulated data volume;accumulated data distribution or variation;static or real-time data sending speed or frequency;static / real-time communications resource information on associated link (s) ;static / real-time computing / storage resource information; orreal-time location information of the participant.22.The method of claim 20, wherein the AI model training status information comprises at least one of:number of past iterations and / or episodes;accumulated training execution time;accumulated training time of a current training episode;convergence level of an embedded model or model partition; orachieved training accuracy or lost level.23.The method of claim 20, wherein the AI model inference status information comprises at least one of:accumulated or real-time inferencing accuracy;inferencing time of the on-going inferencing; oraccumulated inferencing execution time.24.The method of claim 19, wherein the task execution control request comprises at least one of:data transmission control parameters;resource management control parameters;AI model training control parameters; orAI model inference control parameters.25.The method of claim 24, wherein the data transmission control parameters comprise at least one of:control parameter for a data source to start or stop forwarding data;control parameter for a participant network resource to start or stop receiving input data;control parameter to adjust data sending speed or frequency; orcontrol parameter to change source or destination participant network resource for data transmission.26.The method of claim 24, wherein the resource management control parameters comprise control parameter to assign or release computing resources toward execution of the target AI task.27.The method of claim 24, wherein the AI model training control parameters comprise at least one of:control parameter to pause or resume AI model training task;control parameter to start or stop a training episode;control parameter to freeze or de-freeze model or model partition;control parameter to customize model or model partition; orcontrol parameter to update dropout rate indicator.28.The method of claim 24, wherein the AI model inference control parameters comprise at least one of:control parameter to start, stop, or resume AI model inferencing task;control parameter to activate or deactivate early-exit inferencing;control parameter to activate or deactivate accumulative inferencing; orcontrol parameter to customize or augment inferencing data samples.29.An AI task execution method, comprising:sending, by the task control function to the at least one selected participant network resource, execution status information of the target AI task;sending, by the at least one selected participant network resource to the task control function, a task execution control request during execution of the target AI task; andcontrolling execution of data plane functions by the task control function in response to receiving the task execution control request.30.The method of claim 29, wherein the execution status information of the target AI task comprises at least one of:data transmission information;AI model training status information;AI model inference status information; oravailable target data plane function information for the at least one selected participant to access.31.The method of claim 30, wherein the data transmission information comprises at least one of:accumulated data volume;accumulated data distribution or variation; orstatic or real-time data sending speed or frequency.32.The method of claim 30, wherein the AI model training status information comprises at least one of:number of past iterations or episodes;accumulated training execution time;accumulated training time of a current training episode;convergence level of an embedded model or model partition; orachieved training accuracy or lost level.33.The method of claim 30, wherein the AI model inference status information comprises at least one of:accumulated or real-time inferencing accuracy;inferencing time of the on-going inferencing; oraccumulated inferencing execution time.34.The method of claim 30, further comprising forwarding, by the task control function, parameters of the task execution control request to a target data plane function.35.The method of claim 29, wherein the task execution control request comprises at least one of:data transmission control parameters;AI model training control parameters; orAI model inference control parameters.36.The method of claim 35, wherein the data transmission control parameters comprise at least one of:control parameter for a data source to start or stop forwarding data;control parameter for a participant network resource to start or stop receiving input data;control parameter to adjust data sending speed or frequency; orcontrol parameter to change source or destination participant network resource for data transmission.37.The method of claim 35, wherein the AI model training control parameters comprise at least one of:control parameter to pause or resume AI model training task;control parameter to start or stop a training episode;control parameter to freeze or de-freeze model or model partition;control parameter to customize model or model partition; orcontrol parameter to update dropout rate indicator.38.The method of claim 35, wherein the AI model inference control parameters comprise at least one of:control parameter to start, stop, or resume AI model inferencing task;control parameter to activate or deactivate early-exit inferencing;control parameter to activate or deactivate accumulative inferencing; orcontrol parameter to customize or augment inferencing data samples.39.An apparatus, comprising at least one processors, the at least one processors is configured to execute instructions stored in one or more memories to implement the method of any one of claims 1 to 38.40.A computer-readable storage medium having instructions stored thereon which, when executed by a one or more processors cause the one or more processors to perform the method of any one of claims 1 to 38.
Citation Information
Patent Citations
Network resource optimization method and device, electronic equipment and storage medium
CN114021770A
Network mobility management optimization method, base station, device, system and related equipment
CN117560650A
Energy-saving method and system of distributed network, storage medium and electronic equipment
CN117641518A
Method and apparatus for supporting federated learning in wireless communication system
WO2023214806A1
Method and apparatus for supporting hybrid federated learning workload in communication system
WO2024128676A1