Devices and methods for supporting a dynamic distributed ai group in a mobile network
The AI controller in mobile networks manages DAI group dynamics by using dummy and idle agent statuses, ensuring seamless transitions and continuous AI inference, addressing inefficiencies in existing DAI systems.
Patent Information
- Application Number
- PCT/EP2024/086558
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2024-12-16
- Publication Date
- 2026-06-25
AI Technical Summary
Existing dynamic distributed AI (DAI) systems in mobile networks face interruptions and inefficiencies when the size of the DAI group changes due to agents joining or leaving, leading to prolonged training times and disruptions in inference operations.
Implementing an AI controller that manages a dynamic DAI group by utilizing dummy and idle agent statuses, allowing seamless transitions and continued use of trained ML models without interruption, through status-aware operations and lifecycle management.
Ensures uninterrupted AI inference and reduced training times by dynamically managing agent statuses, enabling efficient collaboration and adaptation to changes in the DAI group without affecting ongoing operations.
Smart Images
Figure EP2024086558_25062026_PF_FP_ABST
Abstract
Description
[0001] DEVICES AND METHODS FOR SUPPORTING A DYNAMIC DISTRIBUTED Al GROUP IN A MOBILE
[0002] NETWORK
[0003] TECHNICAL FIELD
[0004] The present invention relates to devices and methods for communication and data processing. More specifically, the present invention relates to devices and methods for supporting a dynamic distributed Al group in a mobile network, in particular a 3GPP mobile network.
[0005] BACKGROUND
[0006] Since 5G, Al is used in mobile networks to assist network operation / management decisions. Some network operations, such as dynamic task migration among UPFs would require very fast network decision on the operation. Distributed Al (DAI) is beneficial in such scenarios due to its ability to take fast decisions based on the local monitoring data. Considering DAI agent is deployed as part of a Network Entity (NE), and each agent collaborates with other agents in a DAI group to assist the operation / management decision at that NE, the overall performance of these NEs (e.g., the amount of treated traffic in the network) can be maximized by appropriate training of the DAI models at these DAI agents.
[0007] SUMMARY OF THE INVENTION
[0008] It is an objective of the present disclosure to provide improved devices and methods for supporting a dynamic distributed Al group in a mobile network, in particular a 3GPP mobile network.
[0009] The foregoing and other objectives are achieved by the subject matter of the independent claims. Further implementation forms are apparent from the dependent claims, the description and the figures. In the following one or more of the following acronyms and abbreviations may be used:
[0010] 5GC / 5GS 5G Core / 5G System
[0011] Al Artificial Intelligence
[0012] CP Control Plane
[0013] DAI Distributed Al
[0014] DAI Group Distributed Al Group
[0015] DQN Deep Q Learning
[0016] ML Machine Learning
[0017] NW Network
[0018] NE Network Entity
[0019] NF Network Function
[0020] OAM Operation Administration and Maintenance
[0021] SBA Service Based Architecture
[0022] SBI Service Based Interface
[0023] SME Session Management Function
[0024] UP User Plane
[0025] UPF User Plane F unction As used herein, distributed Al refers to a framework of several Al agents (or Al models), where each Al agent autonomously monitors system variables, trains and infers its models, takes an action and exchanges monitoring information with other distributed Al agents.
[0026] As used herein, the life cycle of distributed Al may comprise configuration, training, model deployment, inference and model maintenance.
[0027] As used herein, Al agent (also referred to as Al model) may refer to a computing program / sub network functionality of a NE / NF, with Distributed Al capability (i.e., local data monitoring, monitoring data exchange with neighbours and perform DAI model training / inference based on collected monitoring data).
[0028] As used herein, a DAI group refers to a group of network entities / network functions cooperating with each other on distributed Al training and / or inference.
[0029] As used herein, an Al controller refers to a network entity / network function managing and controlling one or more DAI groups for configuration and life cycle management of distributed Al.
[0030] As used herein, an agent status refers to a status of an Al agent in a DAI group (i.e., Dummy, Active, Idle).
[0031] According to a first aspect an Al controller network entity is provided for controlling a plurality of Al agents of a distributed Al, DAI, group of Al agents for jointly performing an Al task in a mobile network. The Al controller network entity according to the first aspect is configured to determine for each of the plurality of Al agents a current Al agent status of a plurality of possible Al agent statuses, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status. In the active status an Al agent is active in the Al task (e.g. training and / or inference) of the DAI group of agents. In the idle status an Al agent is temporally not active in the Al task of the DAI group of agents. The DAI group of agents comprises at least one a virtual Al agent (herein referred to as dummy Al agent or de-associated agent), wherein the dummy Al agent is not capable of being active in the Al task of the DAI group, because it is virtual and has no training and inference capability. The Al controller network entity according to the first aspect is further configured to control for one or more of the plurality of Al agents a transition from the current Al agent status of the Al agent to a different Al agent status of the plurality of possible Al agent statuses during the Al task. Thus, the Al controller network entity according to the first aspect allows supporting a dynamic distributed Al group in a mobile network without inference interruption in case of a dynamic change of the size of the Al group of Al agents. By using dummy and idle Al agent status in a DAI group, the size of the DAI group does not change when an Al agent joins / leaves the Al task of the DAI group. In this way, the trained ML can continue to be used in the active Al agents of the DAI group for the inference.
[0032] In a further possible implementation form, the Al controller network entity according to the first aspect is further configured to control one or more of the plurality of Al agents to cooperate with the at least one dummy Al agent during the Al task. This allows to ensure a seamless interaction between the plurality of Al agents.
[0033] In a further possible implementation form, the Al controller network entity according to the first aspect is further configured to obtain information about the availability of each of the plurality of Al agents from one or more of the plurality of Al agents. This allows to properly control the status transition of the plurality of Al agents.
[0034] In a further possible implementation form, the Al controller network entity according to the first aspect is further configured to provision a feature of the at least one dummy Al agent to an associated Al agent in the DAI group of Al agents. An associated Al agent refers to Al agent with idle status or active status. This allows a new Al agent to join the Al tasks of a DAI group seamlessly.
[0035] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to provide to each of the Al agents in the DAI group information indicative of the respective status of one or more neighboring Al agents. This allows status aware operation among the plurality of Al agents.
[0036] In a further possible implementation form, the Al controller network entity is configured to associate a further Al agent with the DAI group of Al agents. This allows a new Al agent to join the DAI group seamlessly.
[0037] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to provide a trained Machine Learning, ML, model of an Al agent to the further Al agent and to set the current Al agent status of the further Al agent to the idle status. This allows the further Al agent not to train from the very beginning and in turn shorten the training time of the further Al agent.
[0038] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to trigger (for instance, under certain trigger conditions, such as a load threshold) the further Al agent to operate the trained ML Model and to control for the further Al Agent a transition from the idle status to the active status, wherein in the active status the Al agent is active in the Al task of the DAI group of agents. This allows to ensure the accuracy of the Al inference when using the trained ML model.
[0039] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to trigger the Al agents of the DAI group of agents, including the further Al agent, to perform a retraining phase, wherein in the retraining phase the further Al agent retrains the trained ML model provided by the Al controller network entity. This allows to increase the accuracy of the Al inference when using the trained ML model.
[0040] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to deassociate an Al agent of the DAI group of agents from the DAI group of agents or is configured to de-activate an Al agent of the DAI group of agents. This allows an Al agent to leave the DAI group or leave the Al tasks of the DAI group seamlessly.
[0041] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to change the status of the de-associated Al agent from the active Al agent status or the idle Al agent status to the dummy Al agent status, or to change the status of an Al agent from the active Al agent status to the idle Al agent status. This allows an Al agent to be removed from the DAI group or leave the Al tasks of a DAI group without affecting the Al operation of the other Al agents in the DAI group.
[0042] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to store a ML model of the de-associated Al agent or deactivated Al agent in a storage accessible by the Al controller network entity or one of the Al agents of the DAI group of Al agents. This allows a trained ML model in a DAI group to be retrieved when necessary.
[0043] In a further possible implementation form, the Al controller network entity according to the first aspect is configured to store a ML model identifier, a storage location, and / or one or more use conditions of the ML model of the de-associated Al agent or deactivated Al agent in the Al controller network entity according to the first aspect and / or in one of the Al agents of the DAI group of agents. This allows to retrieve the right trained ML model at the right moment. In a further possible implementation form, the Al controller network entity according to the first aspect is implemented as a network function of the mobile network. This allows the proposed Al controller to be used in a mobile network.
[0044] According to a second aspect a method is provided for operating an artificial intelligence, Al, controller network entity for controlling a plurality of Al agents of a distributed Al, DAI, group of Al agents for jointly performing an Al task in a mobile network. The method according to the second aspect comprises: determining for each of the plurality of Al agents a current Al agent status of a plurality of possible Al agent statuses, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status an Al agent is active in the Al task of the DAI group of agents and wherein in the idle status an Al agent is temporally not active in the Al task of the DAI group of agents, wherein the DAI group of Al agents comprises at least one virtual Al agent (herein referred to as dummy Al agent) and wherein the virtual dummy Al agent is not capable of being active in the Al task of the DAI group, because it is virtual and has no training and inference capability; and controlling for one or more of the plurality of Al agents, based on the determined current Al agent status of each Al agent jointly performing the Al task, a transition from the current Al agent status of the Al agent to a different Al agent status of the plurality of possible Al agent statuses during the Al task.
[0045] The method according to the second aspect can be performed by the Al controller network entity according to the first aspect. Thus, further features of the method according to the second aspect result directly from the functionality of the Al controller network entity according to the first aspect, as well as its different implementation forms described above and below.
[0046] According to a third aspect an artificial intelligence, Al, agent is provided for jointly performing with a plurality of further Al agents of a distributed Al, DAI, group of Al agents an Al task in a mobile network. The Al agent according to the third aspect is configured to receive control information from an Al controller network entity of the mobile network. Moreover, the Al agent according to the third aspect is configured to transition, based on the control information, from a current Al agent status of the Al agent to a different Al agent status of a plurality of possible Al agent statuses during the Al task. The plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status. In the active status the Al agent is active in the Al task of the DAI group of Al agents and in the idle status the Al agent is temporally not active in the Al task of the DAI group of Al agents. The DAI group of Al agents comprises at least one virtual Al agent (herein referred to as dummy Al agent), wherein the virtual, i.e. dummy Al agent is not capable of being active in the Al task of the DAI group, because it is virtual and has no training and inference capability. This allows supporting a dynamic distributed Al group in a mobile network without inference interruption in case of a dynamic change of the size of the Al group of Al agents.
[0047] In a further possible implementation form, the Al agent according to the third aspect is configured to receive from the Al controller network entity information indicative of the respective status of one or more neighboring Al agents of the Al agent. This allows to ensure a seamless interaction between the plurality of Al agents.
[0048] In a further possible implementation form, the Al agent according to the third aspect is configured to operate depending on the information indicative of the respective status of the one or more neighboring Al agents of the Al agent. This allows status aware operation among the plurality of Al agents.
[0049] In a further possible implementation form, the Al agent according to the third aspect is configured to be associated with the DAI group of Al agents. This allows a new Al agent to join the DAI group seamlessly. In a further possible implementation form, based on the control information, the Al agent according to the third aspect is configured to be de-associated from the DAI group of Al agents from the DAI group of Al agents or to be de-activated. This allows an Al agent to leave the DAI group or leave the Al tasks of the DAI group seamlessly.
[0050] In a further possible implementation form, based on the control information, the Al agent according to the third aspect is configured to transition from the dummy status to the idle status, from the active status or from the idle status to the dummy status, from the idle status to the active status, or from the active status to the idle status. This allows an Al agent to join or leave the Al tasks of a DAI group seamlessly based on a trigger from the Al controller.
[0051] According to a fourth aspect a method is provided for operating an artificial intelligence, Al, agent for jointly performing with a plurality of further Al agents of a distributed Al, DAI, group of Al agents an Al task in a mobile network, wherein the method comprises: receiving control information from an Al controller network entity of the mobile network; and transitioning, based on the control information, from a current Al agent status of the Al agent to a different Al agent status of a plurality of possible Al agent statuses during the Al task, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status the Al agent is active in the Al task of the DAI group of Al agents and wherein in the idle status the Al agent is temporally not active in the Al task of the DAI group of Al agents, wherein the DAI group of Al agents comprises at least one virtual Al agent (herein referred to as dummy Al agent) and wherein the dummy Al agent is not capable of being active in the Al task of the DAI group, because it is virtual and has no training and inference capability.
[0052] The method according to the fourth aspect can be performed by the Al agent according to the third aspect. Thus, further features of the method according to the fourth aspect result directly from the functionality of the Al agent according to the third aspect, as well as its different implementation forms described above and below.
[0053] According to a fifth aspect a computer program product is provided, comprising program code which causes a computer or a processor to perform the method according to the second aspect or the method according to the fourth aspect, when the program code is executed by the computer or the processor.
[0054] Details of one or more embodiments are set forth in the accompanying drawings and the description below. Other features, objects, and advantages will be apparent from the description, drawings, and claims.
[0055] BRIEF DESCRIPTION OF THE DRAWINGS
[0056] In the following, embodiments of the present disclosure are described in more detail with reference to the attached figures and drawings, in which:
[0057] Figs, la and lb show schematic diagrams illustrating an Al controller entity according to an embodiment for controlling a plurality of Al agent entities according to an embodiment in a mobile network;
[0058] Fig. 2 shows a schematic diagram illustrating the main functional modules of an Al agent according to an embodiment exchanging monitoring data with another Al agent during an Al task according to an embodiment;
[0059] Fig. 3 shows a schematic diagram illustrating the architecture of a mobile network including an Al controller entity and a plurality of Al agents according to an embodiment; Fig. 4a shows a signalling diagram illustrating a conventional Al agent joining a DAI group with ML model retraining;
[0060] Fig. 4b shows results illustrating the required episodes to train an ML model with a different number of Al agents in a conventional DAI group;
[0061] Fig. 5 shows a signalling diagram illustrating a conventional Al agent joining a DAI group using transfer learning;
[0062] Fig. 6 shows a signalling diagram illustrating the interaction of an Al control entity according to an embodiment with a plurality of Al agents with Al agent status according to an embodiment, when an Al agent joins a DAI group of Al agents;
[0063] Fig. 7 shows a schematic diagram illustrating status transitions of an Al agent according to an embodiment and related triggers;
[0064] Fig. 8 shows a schematic diagram illustrating different aspects of an Al control entity according to an embodiment, more specifically in an upper table an Al agent status maintained at the Al controller and the information used by the Al controller in case of a dummy Al agent and in a lower table neighbour Al agent status and feature of dummy Al agent used by Al agent in the Al task;
[0065] Fig. 9 shows a schematic diagram illustrating different aspects of an Al agent entity according to an embodiment;
[0066] Fig. 10 shows a table illustrating exemplary states and actions at a Re-enforcement Learning agent with QL;
[0067] Fig. 11 shows a schematic diagram illustrating exemplary states and actions at a Re-enforcement Learning Al agent with a neural network;
[0068] Fig. 12 shows a schematic diagram illustrating a deactivation of an Al agent entity according to an embodiment by an Al controller entity according to an embodiment;
[0069] Fig. 13 shows a schematic diagram illustrating an activation of an Al agent entity according to an embodiment by an Al controller entity according to an embodiment;
[0070] Fig. 14 shows a schematic diagram illustrating an association of an Al agent entity according to an embodiment by an Al controller entity according to an embodiment;
[0071] Fig. 15 shows a schematic diagram illustrating a de-association of an active Al agent entity according to an embodiment by an Al controller entity according to an embodiment;
[0072] Fig. 16 shows a schematic diagram illustrating the architecture of a mobile network including an Al controller entity and a plurality of Al agents according to an embodiment;
[0073] Figs. 17a and 17b show schematic diagrams illustrating different implementations of a dummy Al agent entity according to an embodiment;
[0074] Fig. 18 shows a state diagram illustrating late joining of an Al agent entity according to an embodiment; Figs. 19a and 19b illustrate a procedure when a new Al agent according to an embodiment joins the Al tasks of a DAI group in 5G architecture;
[0075] Fig. 20 shows a flow diagram illustrating steps of a method according to an embodiment for operating an Al controller entity; and
[0076] Fig. 21 shows a flow diagram illustrating steps of a method according to an embodiment for operating an Al agent entity.
[0077] In the following, identical reference signs refer to identical or at least functionally equivalent features.
[0078] DETAILED DESCRIPTION OF THE EMBODIMENTS
[0079] In the following description, reference is made to the accompanying figures, which form part of the disclosure, which illustrates specific aspects of embodiments of the present disclosure or specific aspects in which embodiments of the present disclosure may be used. It is understood that embodiments of the present disclosure may be used in other aspects and comprise structural or logical changes not depicted in the figures. The following detailed description, therefore, is not to be taken in a limiting sense, and the scope of the present disclosure is defined by the appended claims.
[0080] For instance, it is to be understood that a disclosure in connection with a described method may also hold true for a corresponding device or system configured to perform the method and vice versa. For example, if one or a plurality of specific method steps are described, a corresponding device may include one or a plurality of units, e.g. functional units, to perform the described one or plurality of method steps (e.g. one unit performing the one or plurality of steps, or a plurality of units each performing one or more of the plurality of steps), even if such one or more units are not explicitly described or illustrated in the figures. On the other hand, for example, if a specific apparatus is described based on one or a plurality of units, e.g. functional units, a corresponding method may include one step to perform the functionality of the one or plurality of units (e.g. one step performing the functionality of the one or plurality of units, or a plurality of steps each performing the functionality of one or more of the plurality of units), even if such one or plurality of steps are not explicitly described or illustrated in the figures. Further, it is understood that the features of the various exemplary embodiments and / or aspects described herein may be combined with each other, unless specifically noted otherwise.
[0081] Figures la and lb show schematic diagrams illustrating an Al controller entity 120 according to an embodiment for controlling a plurality of Al agent entities HOa-n according to an embodiment in a mobile network 100, in particular a 3GPP mobile network 100. Each of the Al agents 1 lOa-n is configured to implement an Al model for performing a distributed Al task in a DAI group. In an embodiment, the Al controller entity 120 may be implemented as a component of a network entity 121, in particular a network function 121 of the mobile network 100. In an embodiment, each of the Al agents HOa-n may be implemented as a component of a respective network entity 11 la-n, in particular a respective network function 11 la-n of the mobile network 100.
[0082] In an embodiment, each Al agent entity 11 Oa-n may implement one or more of the functional modules illustrated in figure 2 and described in the following: a Data Collection and Sharing Function 114a, b that enables the Al agent 110a, b to collect data not only from its managed NF, but also share minimal but essential information with other Al agents in its DAI group to learn cooperatively; an Intelligence Function 113a, b that stores the current status of its Al model e.g. a neural network or a Q-table; a Monitoring Function 115a,b constantly running in the background, that monitors the important AI / ML KPIs, e.g. average reward and loss functions, meaning that the Al agent 110a, b raises a warning to the controller if the AI / ML KPIs deviate by more than a pre-defined threshold value by the controller; and / or an Action Function 112a, b that enables the Al agent 110a, b to autonomously perform an action based on the decision of the Intelligence Function 113a,b. Each of these functions may be implemented in different ways depending on the specific use case.
[0083] The Al controller 120 is responsible for the control and management of the Al agents 1 lOa-n in a DAI group. More specifically, the Al controller 120 configures, reconfigures or updates the computation resources and functionalities of the Al agents 110a- n regarding a specific Al task and triggers the different phases for the life-cycle management of the Al agents 1 lOa-n on a specific Al task. The life-cycle phases of an Al agent 11 Oa-n may consist of 1 ) Training, 2) Deployment of Trained models and Inference, and 3) Reporting and Maintenance. The Al controller 120 may be also aware of the topology information (e.g. Al agent neighbor relationships), Al agent capabilities (i.e. Al algorithm settings) and information relating to infrastructure changes. The Al controller 120 may additionally monitor system KPIs e.g. percentage of tasks violating their deadlines in task migration.
[0084] An example of DAI deployment in a 3GPP mobile network 100 is shown in figure 3, where a group of DAI agents HOa-n (each hosted by an UPF) collaborate with each other to perform task migration among these UPFs. The Al controller 120 hosted at the SMF manages and controls these distributed Al capable UPF s via N4 interfaces. Analytics generated by the Al agents 1 lOa-n may be used to dynamically migrate tasks among UPFs, when UPF task deadlines are violated due to the high load of a UPF.
[0085] In a 3 GPP network, such as the network 100 illustrated in figure 3, a network function (NF) can be dynamically deployed considering the changing load of the network traffic. A NF (e.g., UPF) can also be switched off for energy saving reasons, or newly deployed when the existing NFs in the network are overloaded. When an NF is newly deployed or switched off, the Al agent hosted by that NF needs to be added in a collaboration DAI group or needs to be removed from the collaboration DAI group. When the number of Al agents changes in the collaboration DAI group, Al inference using the old trained ML model in the DAI group may not work anymore.
[0086] In a test scenario using DAI for network slice computation resource auto-scaling, the results show that when a DAI model trained with 5 Al agents is used for inference with 10 Al agents, the overall system KPI is similar to the network operation without using Al. Meanwhile, Al training needs very long time. In the same test scenario, training 8 Al agents using DQN with 30 episodes and 30 minutes per episode requires 15 hours for the training as shown in figures 4a and 4b. This indicates that, when the number of Al agents is changed e.g., from 7 to 8, the Al inference may be interrupted for almost half a day before the model is retrained to achieve sufficient accuracy when used for inference.
[0087] As illustrated by figure 5, initiating training with a pre-trained ML model with a similar number of Al agents instead of starting to train a ML model from zero may shorten the training time. However, for the approach using transfer learning all the Al agents in a DAI group still need to obtain a new trained ML model (as initial ML model) in case the DAI group size changes. This causes additional signaling overhead to retrieve the trained ML model for each of the Al agents in the DAI group. Thus, the DAI inference is still interrupted for all the agents in the DAI group during the ML model update period and further training period.
[0088] Embodiments of the Al controller entity 120 and the plurality of Al agent entities 1 lOa-n disclosed herein allow avoiding DAI inference interruption in case of a dynamic deployment of the Al agents HOa-n, i.e. when an Al agent joins or is removed from the DAI group of Al agents. More specifically, embodiments disclosed herein are related to the lifecycle management and control of a DAI group of Al agents 1 lOa-n (i.e., control of training, inference phases and ML model management), so that Al service interruption can be minimized in case of dynamic group membership of the Al agents 1 lOa-n in the DAI group. As will be described in more detail under further reference to figure 6 below, the Al controller entity 120 is configured to determine for each of the plurality of Al agent entities HOa-n a current Al agent status of a plurality of possible Al agent statuses. The plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status an Al agent is active in the Al task of the DAI group of agents 1 lOa-n and wherein in the idle status an Al agent is temporally not active in the Al task of the DAI group of agents 1 lOa-n. As will be described in more detail further below, the DAI group of Al agents 1 lOa-n comprises at least one dummy Al agent 110a, wherein the dummy Al agent 110a is not capable of being active in the Al task of the DAI group. Moreover, the Al controller entity 120 is configured to control for one or more of the plurality of Al agents 1 lOa-n a transition from the current Al agent status of the Al agent to a different Al agent status of the plurality of possible Al agent statuses during the Al task.
[0089] Thus, according to embodiments disclosed herein the Al controller entity 120 is configured to determine the agent status, and the dummy agent to be used in a DAI group considering the dynamicity of available agents and to control the status transition of each agent 1 lOa-n in the DAI group. As will be described in more detail below, according to embodiments disclosed herein each Al agent entity 1 lOa-n is configured to perform status aware operation and Al computations.
[0090] In the embodiment shown in figure 6, the Al controller 120 configures each agent 1 lOa-n in a DAI group with agent status (dummy, active, idle). The dummy Al agent information is included in the training and inference procedure without having a real Al agent with the Al training and inference capability (see step 1 and step 4 of figure 6). When a new Al agent 110a joins the DAI group, the controller 120 changes the agent status from dummy to active or idle and determine the related (re jeonfiguration of the DAI group (1 lOa-n) accordingly (see step 6 of figure 6). The Al controller initiates the training of the new agent 110a with a trained ML model from the DAI group (see step 7A of figure 6). For the other agents 1 lOb-n in the DAI group (active ones), the controller 120 only updates the agent configuration on the neighborhood information (e.g., IP address, agent status of the new joining agent) (see step 7B of figure 6). In this way, the other agents in the DAI group can continue to use the old trained ML model for the inference. Optionally, the new agent 110a may start to train together with the other Al agents 1 lOb-n in the DAI group using the initial ML model indicated by the Al controller 120 (see step 8 of figure 6). When the training is completed or a certain condition configured by the controller 120 is reached, the new agent 110a reports to the Al controller 120. The Al controller 120 then triggers the new agent 110a to join the inference of the DAI group (see step 10 of figure 6). And the new agent 110a will perform the inference task together with the other Al agents 1 lOb-n (see step 11 of figure 6). As will be appreciated, because only the status of an agent changes when an agent joins or leaves the DAI group and the DAI group size does not change, the trained ML models in a DAI group can be continued to be used for the inference. Thus, according to embodiments disclosed herein the new agent 110a can join the DAI group without interrupting the inference performed by the other DAI agents 1 lOb-n.
[0091] As already mentioned above, the following statuses may be defined for each agent 1 lOa-n: an active, idle and dummy status. The active status refers to an Al agent 1 lOa-n active in the training and inference tasks of a DAI group. The idle status refers to an Al agent 1 lOa-n temporally not active in the training and inference tasks of a DAI group. The dummy status refers to a virtual Al agent 1 lOa-n with no training and inference capability. According to an embodiment, the Al controller entity 120 is configured to determine the agent status, status transitions and the dummy agent to be used in a DAI group considering the dynamicity of available agents, as illustrated in figure 7. As will be appreciated, the dynamicity of available agents HOa-n differentiates the cases where a respective agent 1 lOa-n is temporary not available or not available at all (e.g., not deployed yet). As illustrated in figure 7, the Al controller entity 120 may be configured to perform the following operations resulting in status transitions of the Al agents 1 lOa-n: activate / de-activate; associate; and de-associate. When a real agent 1 lOa-n associated with a DAI group is configured to join / leave the Al tasks of the DAI group, the status of the agent 1 lOa-n is changed by the Al controller 120 from idle to active / active to idle in this DAI group. When a real agent 1 lOa-n is removed from a DAI group the status of the agent 1 lOa-n is changed to dummy in this DAI group. When a real agent 1 lOa-n is added into the DAI group, but does not join the Al tasks of the DAI group immediately, the status of a Dummy agent is changed to idle in this DAI group. According to an embodiment, the Al controller 120 is aware of the agent availability and time schedule of the agent availability (e.g., new deployment of an Al agent, energy saving / maintenance plan) from OAM, by discovery via an NF repository entity or based on report from the agent.
[0092] As alreadv mentioned above, the Al controller entity 120 is configured for status aware control and management of the plurality of Al agent entities HOa-n. More specifically, in an embodiment, the Al controller entity 120 is configured to monitor and maintain the agent status, control the status transitions and manage the trained ML models during the status transitions in a DAI group. Moreover, the Al controller entity 120 may perform AIML KPIs (e.g., rewards, loss, gradients) calculations / estimations considering the agent status (KPIs estimated by the controller itself in case of Dummy and Idle) as illustrated in the table a shown in figure 8. Furthermore, the Al controller entity 120 may be configured to determine the agent configuration in a DAI group considering agent status and status transition.
[0093] In an embodiment, the agent configurations supporting agent status may include one or more of the following configurations: configuration on neighborhood information (e.g., indicate also agent status for each neighbor, address of local / selected agent used for neighbor(s) with Dummy status); configuration on status aware operation (e.g., skip a neighbor agent with certain status from data collection, trained ML model store / load / retrieve during status transfer, and the like); and / or configuration on DAI algorithm (e.g., features of dummy agent, status aware Al calculation, like).
[0094] As illustrated in the embodiments shown in figures 8 and 9, a feature of an agent 1 lOa-n may be used for the Al calculation at the agent 11 Oa-n. In the UPF task migration example mentioned above, a feature of an agent may include the available resources and load information of the hosting NF of that agent (as illustrated in table b in figure 8). For a real agent, such a feature may based on the real monitoring data and may, for instance, be indicated as Y available FLOPs, X% CPU utilization (as shown in figures 8 and 9). For a dummy agent 1 lOa-n, such feature may be estimated by the Al controller 120 since the agent 110a and the hosting NF I l la does not really exist and there is no real monitoring data.
[0095] As further illustrated in figure 9, each Al agent 1 lOa-n may report on agent availability, agent condition (e.g., load), trained ML model ID / version to the controller 120. Moreover, with respect to other agents each agent may be configured for a status aware Al operation and status aware Al calculation (e.g., ignore, or use previously determined feature or estimated feature (e.g., from controller) if neighbor is indicated as dummy / idle). There are different ways to obtain the feature of a dummy agent, as described further below. Optionally, the features may be provided on behalf of the dummy agent. For a joining agent trained ML model parameters (locally stored before the leaving, or from other agent in the DAI group) may be loaded or retrieved based on controller configuration. For a leaving agent the ML model context (i.e., algorithm, configuration, parameters) may be stored before leaving.
[0096] For the status aware Al calculation mentioned above the implementation of the Al calculation depends on the used Al algorithm. When RL algorithm = QL, the Q-table of each QL agent consists of Q-values that are directly learned and improved. Here the state is calculated as a function of monitoring information from different agents, as illustrated by the table shown in figure 10. An exemplary state calculation of each RL agent with QL (with dummy agent) is based on the following equation:
[0097] State s = / (local monitoring data, neighbor monitoring data, dummy agent monitoring data).
[0098] When RL algorithm = DQN / A2C (i.e. neural network based), the neural network of each agent 1 lOa-n consists of weights and biases that are learned and improved in order to improve estimates of Q-values. Here the state may be a vector of the monitoring information from different agents 1 lOa-n. Figure 11 shows exemplary states and actions at a Re-enforcement Learning agent 1 lOa-n with a neural network. The state of each RL agent (with dummy agent) is as follows:
[0099] State s = vector of (local monitoring data, neighbor monitoring data, dummy agent monitoring data).
[0100] Figure 12 shows a schematic diagram illustrating a deactivation of the Al agent entity 110a (referred to as Agent A in figure 12) according to an embodiment by the Al controller entity 120 according to an embodiment such that the Al agent entity 110a temporarily leaves the DAI group further including the Al agent entities 110b and 110c (referred to as Agents B and C in figure 12). If the status of Al agent 110a, i.e. agent A is changed to idle, the Al controller 120 configures the agents in the following way: agent A caches the trained ML model; the neighboring agents B and C have no monitoring data exchange with agent A for the inference, and use X as the monitoring data / features from agent A, wherein X refers to previous / averaged monitored data from A in the past (which may be a value configured by the controller 120 or monitored data from agent C). The controller 120 may estimate by itself the AIML KPIs of an idle agent (e.g., loss, award) without the real report from the agent A.
[0101] Figure 13 shows a schematic diagram illustrating an activation of the Al agent entity 110a (referred to as Agent A in figure 13) according to an embodiment by the Al controller entity 120 according to an embodiment such that the Al agent entity 110a rejoins the DAI group further including the Al agent entities 110b and 110c (referred to as Agents B and C in figure 13). If the status of agent A is changed back to active, the Al controller 120 configures the agents in the following way: a ML model (model ID, model source) is retrieved or reloaded by agent A (the ML model may optionally be also reloaded by the neighboring agents B and C); and agent A resumes collecting monitoring data collection at the neighboring agents B and C, and use that as features from agent A. The Al controller entity 120 may use AIML KPIs reported from agent A again.
[0102] Figure 14 shows a schematic diagram illustrating an association of the Al agent entity 110a (referred to as Agent A in figure
[0103] 14) according to an embodiment by the Al controller entity 120 according to an embodiment such that the new Al agent entity 110a joins for the first time the DAI group further including the Al agent entities 110b and 110c (referred to as Agents B and C in figure 14). Before the association, the Al controller entity 120 maintains features of a dummy agent at the real agents B, C (e.g., with estimated features of a joining agent in the future). The address of dummy agent may be configured to a local address. The Al controller entity 120 may estimate by itself the AIML KPIs of the dummy agent (e.g., loss, award) without the real report from the agent A. If a new agent A joins the DAI group, the Al controller 120 configures the agents in the following way: agents B and C update the address of dummy agent from local address to address of agent A and update the status of the dummy agent to IDLE; and agent A retrieves a trained ML model (model ID, model source) from agent B or C.
[0104] F igure 15 shows a schematic diagram illustrating a de-association of the Al agent entity 110a (referred to as Agent A in figure
[0105] 15) according to an embodiment by the Al controller entity 120 according to an embodiment such that the Al agent entity 110a transitions to the dummy status, i.e. leaves the DAI group further including the Al agent entities 110b and 110c (referred to as Agents B and C in figure 15). If the status of agent A is changed to dummy, the Al controller 120 configures the agents in the following way: the trained ML model of agent A may be cached to an external location, e.g. an external storage; the neighboring agents B and C have no monitoring data exchange with agent A, and use X as the features of a dummy agent, wherein X may be the previously averaged monitored data from agent A in the past (the value may be configured by the controller 120 or monitored data from agent C). The Al controller entity 120 may estimate by itself the AIML KPIs of dummy agent (e.g., loss, award) without the real report from the agent A. The address of dummy agent may be configured to a local address.
[0106] Figure 16 shows an example deployment of DAI infrastructure in 3GPP 5GC. In this example, the Al controller entity 120 is deployed as a dedicated NF or as a component / sub-functionality of an existing NF (e.g., SMF) connected to 5GC using SBI, while the plurality of Al agents 1 lOa-n is deployed as a component / sub-functionality of AMF / UPF / PCF. The Al controller 120 may control multiple DAI groups (e.g., one or more AMF groups, and / or UPF groups, and / or PCF groups, as shown in the figure). Similar infrastructure applies also to 3GPP MP architecture or future SBA based mobile network (e.g., 6G) architecture, where 5GC NF is replaced by a MP NF or a 6G NF.
[0107] Figures 17a and 17b show schematic diagrams illustrating different implementations of a dummy Al agent entity according to an embodiment. In the embodiment shown in figure 17a (referred to as option A.l), the dummy agent is implemented as a virtual agent. The Al controller entity 120 provides features of the dummy agent (e.g., load and available resources) to each real agent in the DAI group. In the embodiment shown in figure 17b the Al controller entity 120 is configured to provide estimated features of the dummy agent (e.g., load and available resources) to one or more real agent in the DAI group, which may be configured to act on behalf of the dummy agent with respect to neighbor information exchange.
[0108] In an embodiment, the Al controller entity 120 may decide to include a new agent at a later until the new agent reaches a certain condition (e.g., a certain load) suitable to use a trained ML model. To this end, the Al controller entity 120 may implement the procedure illustrated in figure 18. In step 0 (not shown), the Al controller entity 120 may configure an agent to report its condition (e.g., load) when leaving. When a new agent is available, the Al controller entity 120 may configure the new agent to get a trained ML model and report when a certain condition (e.g., load, available CPU) is reached (see step 1 of figure 18). When a certain condition is reached, the Al controller entity 120 triggers the new agent to join the DAI group for inference (see step 2 of figure 18).
[0109] Concerning the determination of the features of a dummy or idle Al agent 11 Oa-n, according to an embodiment the Al controller entity 120 may be configured to use at least one of the following options: estimated (e.g., based on simulation using digital twin); duplicated from input of a selected neighboring Al agent 11 Oa-n in the same iteration; using the input from the previous iteration of a selected Al agent 11 Oa-n; average over inputs from the previous iterations of a selected Al agent 1 lOa-n; average over inputs of all neighboring Al agents 1 lOa-n in the same iteration; and / or use null.
[0110] Figures 19a and 19b illustrate a procedure when a new Al agent 110a according to an embodiment in the form of a new UPF 110a joins the Al tasks of a DAI group implemented in the 5G architecture. In this exemplary embodiment, the Al task(s) may be load balancing and / or task migration among the UPFs, i.e. Al agents 1 lOa-n, and the Al controller entity 120 is implemented as a component of the SMF.
[0111] Step 1 offigure 19b: The new UPF 110a is not available yet and has the status of dummy. The existing UPFs HOb-n in a DAI group are operating with Al. The existing UPFs 1 lOb-n are provisioned with the feature of a dummy agent (from the controller 120 hosted at the SMF via N4 interface, or from one selected UPF by the controller). At this stage, the existing UPFs 1 lOb-n are performing inference jointly with the new UPF 110 as a dummy agent.
[0112] Associate phase:
[0113] Step 2 of figure 19b. When the new UPF 110a (with agent capability) is available, it reports to the SMF (controller) 120 via N4 interface (step 2a), or SMF discovery that via OAM or NF repository function (e.g., NRF) 140 (step 2b).
[0114] Step 3 of figure 19b: The SMF 120 decides on the configuration of the dummy agent in the DAI group 1 lOa-n considering the information of the new UPF 110a.
[0115] Step 4 of figure 20b: The SMF 120 provides the agent configuration to the new UPF 110a via N4, which includes the neighborhood information (existing UPFs), and initial ML model to be obtained (e.g., model ID and the source UPF ID / address).
[0116] Step 5 offigure 19b: The new UPF 110a requests or retrieves the ML model from the UPF indicated in step 4. Step 6 of figure 19b: The status of the new UPF I la transits from dummy to Idle and the new UPF 110a is associated to the DAI group. At this stage, the new UPF processes without Al. The existing UPF s 11 Ob-n continue to operating with Al.
[0117] Activate phase:
[0118] The SMF 120 decides to include the agent hosted at the new UPF 110a to the DAI task when certain conditions are reached (not indicated in this sequence chart, the condition could be based on the load of the new UPF to reach a certain threshold).
[0119] Step 7 of figure 19b: The SMF 120 updates the configurations of the old UPFs 11 Ob-n in the DAI group on the status update of the New UPF 110a and the address via N4.
[0120] Step 8 of figure 19b: The SMF 120 (optionally) triggers the new UPF 110a to join training of the DAI task via configuration overN4.
[0121] Step 9 of figure 19b: The new UPF 110a (optionally) trains together with the existing UPFs HOb-n in the DAI group. Step 10 of figure 19b: The new UPF 110a (optionally] indicates the training is completed to the SMF 120 viaN4.
[0122] The steps 8-10 are optional depending on whether further training is necessary for the new UPF before it joins the inference task.
[0123] Step 11 of figure 19b: The SMF 120 triggers the new UPF 110a to join the inference of the DAI task via N4.
[0124] Step 12 of figure 19b: The status of the New UPF transitions from idle to active. The new UPF 110a processes with Al (i.e., inferences together with existing UPFs 1 lOb-n in the DAI group).
[0125] As will be appreciated, depending on the actual deployment of agents and the controller as different type of NFs / NEs, the N4 interface in the embodiment shown in figures 19a and 19b may be replaced by a service-based interface (e.g., agent configuration service provided by SMF and agent reporting service provided by the UPF). However, the key IES (i.e., agent status, neighborhood address, ML model info, triggering of different life cycle state, feature of dummy agent) will not change. Similarly, the interface used for agent to agent communication may be based on actual deployment.
[0126] As will be appreciated, according to embodiments disclosed herein a dummy Al agent may be included in the DAI group for a potential new joining agent, so that the other agents can continue to use the trained ML model without interrupting inference in the DAI group. The joining agent may reuse a trained ML model in the DAI group so that the training time for the joining Al agent is reduced. The Al controller only needs to update the configuration of other agents instead of initiating the training with a new ML model. This also leads to a reduction of the signaling overhead at the other Al agents. The Al controller entity 120 may trigger the Al agent to start the inference when certain conditions are reached, i.e. flexible joining time of the new Al agent in the inference.
[0127] Figure 20 shows a flow diagram illustrating a method 2000 according to an embodiment for operating the Al controller network entity 120 for controlling the plurality of Al agents of a distributed Al, DAI, group of Al agents HOa-n for jointly performing an Al task in the mobile network 100. The method 2000 comprises a step 2001 of determining for each of the plurality of Al agents 1 lOa-n a current Al agent status of a plurality of possible Al agent statuses. As already described above, the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status an Al agent is active in the Al task of the DAI group of agents 1 lOa-n and wherein in the idle status an Al agent is temporally not active in the Al task of the DAI group of agents 1 lOa-n, wherein the DAI group of Al agents 1 lOa-n comprises a dummy Al agent 110a and wherein the dummy Al agent 110a is not capable of being active in the Al task of the DAI group. Moreover, the method 2000 comprises a step 2203 of controlling for one or more of the plurality of Al agents 11 Oa-n, based on the determined current Al agent status of each Al agent jointly performing the Al task, a transition from the current Al agent status of the Al agent to a different Al agent status of the plurality of possible Al agent statuses during the Al task. Figure 21 shows a flow diagram illustrating a method 2100 according to an embodiment for operating an Al agent 1 lOa-n for jointly performing with a plurality of further Al agents 1 lOa-n of a DAI group of Al agents 1 lOa-n an Al task in the mobile network 100. The method 2100 comprises a step 2101 of receiving control information from the Al controller network entity 120 of the mobile network 100. Moreover, the method 2100 comprises a step 2103 of transitioning, based on the control information, from a current Al agent status of the Al agent to a different Al agent status of a plurality of possible Al agent statuses during the Al task, wherein the plurality of possible Al agent statuses comprises an active status, an idle status, and a dummy status. As already described above, in the active status the Al agent is active in the Al task of the DAI group of Al agents 1 lOa-n and in the idle status the Al agent is temporally not active in the Al task of the DAI group of Al agents 1 lOa-n, wherein the DAI group of Al agents 1 lOa-n comprises a dummy Al agent 110a and wherein the dummy Al agent 110a is not capable of being active in the Al task of the DAI group.
[0128] The person skilled in the art will understand that the "blocks" ("units") of the various figures (method and apparatus) represent or describe functionalities of embodiments of the present disclosure (rather than necessarily individual "units" in hardware or software) and thus describe equally functions or features of apparatus embodiments as well as method embodiments (unit = step).
[0129] In the several embodiments provided in the present application, it should be understood that the disclosed system, apparatus, and method may be implemented in other manners. For example, the described embodiment of an apparatus is merely exemplary. For example, the unit division is merely logical function division and may be another division in an actual implementation. For example, a plurality of units or components may be combined or integrated into another system, or some features may be ignored or not performed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections may be implemented by using some interfaces. The indirect couplings or communication connections between the apparatuses or units may be implemented in electronic, mechanical, or other forms.
[0130] The units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one position, or may be distributed on a plurality of network units. Some or all of the units may be selected according to actual needs to achieve the objectives of the solutions of the embodiments.
[0131] In addition, functional units in the embodiments of the invention may be integrated into one processing unit, or each of the units may exist alone physically, or two or more units are integrated into one unit.
Claims
CLAIMS1. An artificial intelligence, Al, controller network entity (120) for controlling a plurality of Al agents (HOa-n) of a distributed Al, DAI, group of Al agents (1 lOa-n) for jointly performing an Al task in a mobile network (100), wherein the Al controller network entity (120) is configured to: determine for each of the plurality of Al agents (1 lOa-n) a current Al agent status of a plurality of possible Al agent statuses, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status an Al agent is active in the Al task of the DAI group of agents (1 lOa-n) and wherein in the idle status an Al agent is temporally not active in the Al task of the DAI group of agents (1 lOa-n), wherein the DAI group of Al agents (1 lOa-n) comprises a dummy Al agent (110a) and wherein the dummy Al agent (110a) is not capable of being active in the Al task of the DAI group; and control for one or more of the plurality of Al agents (1 lOa-n) a transition from the current Al agent status of the Al agent to a different Al agent status of the plurality of possible Al agent statuses during the Al task.
2. The Al controller network entity (120) of claim 1, wherein the Al controller network entity (120) is further configured to control one or more of the plurality of Al agents (1 lOa-n) to cooperate with the at least one dummy Al agent (110a) during the Al task.
3. The Al controller network entity (120) of claim 1 or 2, wherein the Al controller network entity (120) is further configured to obtain information about the availability of each of the plurality of Al agents (1 lOa-n) from one or more of the plurality of Al agents (HOa-n).
4. The Al controller network entity (120) of any one of the preceding claims, wherein the Al controller network entity (120) is further configured to provision a feature of the at least one dummy Al agent (110a) to an associated Al agent in the DAI group of Al agents.
5. The Al controller network entity (120) of any one of the preceding claims, wherein the Al controller network entity (120) is configured to provide to each of the Al agents in the DAI group information indicative of the respective status of one or more neighboring Al agents.
6. The Al controller network entity (120) of any one of the preceding claims, wherein the Al controller network entity (120) is configured to associate a further Al agent with the DAI group of Al agents (1 lOa-n).
7. The Al controller network entity (120) of claim 6, wherein the Al controller network entity (120) is configured to provide a trained Machine Learning, ML, model of an Al agent to the further Al agent and to set the current Al agent status of the further Al agent to the idle status.
8. The Al controller network entity (120) of claim 7, wherein the Al controller network entity (120) is configured to trigger the further Al agent to operate the trained ML Model and to control for the further Al Agent a transition from the idle status to the active status, wherein in the active status the Al agent is active in the Al task of the DAI group of agents (1 lOa-n).
9. The Al controller network entity (120) of any one of claims 6 to 8, wherein the Al controller network entity (120) is configured to trigger the Al agents of the DAI group of agents (1 lOa-n), including the further Al agent, to perform a retrainingphase, wherein in the retraining phase the further Al agent retrains the trained ML model provided by the Al controller network entity (120).
10. The Al controller network entity (120) of any one of the preceding claims, wherein the Al controller network entity (120) is configured to de-associate an Al agent of the DAI group of Al agents from the DAI group of Al agents (1 lOa-n) or is configured to de-activate an Al agent of the DAI group of Al agents.
11. The Al controller network entity (120) of claim 10, wherein the Al controller network entity (120) is configured to change the status of the de-associated Al agent from the active Al agent status or the idle Al agent status to the dummy Al agent status, or change the status of an Al agent from the active Al agent status to the idle Al agent status.
12. The Al controller network entity (120) of claim 10 or 11, wherein the Al controller network entity (120) is configured to store a ML model of the de-associated Al agent or deactivated Al agent in a storage accessible by the Al controller network entity (120) or one of the Al agents of the DAI group of Al agents (1 lOa-n).
13. The Al controller network entity (120) of any one of claims 10 to 12, wherein the Al controller network entity (120) is configured to store a ML model identifier, a storage location, and / or one or more use conditions of the ML model of the deassociated Al agent or deactivated Al agent in the Al controller network entity (120) and / or in one of the Al agents of the DAI group of Al agents (1 lOa-n).
14. The Al controller network entity (120) of any one of the preceding claims, wherein the Al controller network entity (120) is implemented as a network function (120) of the mobile network.
15. A method (2000) for operating an artificial intelligence, Al, controller network entity (120) for controlling a plurality of Al agents of a distributed Al, DAI, group of Al agents (HOa-n) for jointly performing an Al task in a mobile network (100), wherein the method (2000) comprises: determining (2001) for each of the plurality of Al agents (HOa-n) a current Al agent status of a plurality of possible Al agent statuses, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status an Al agent is active in the Al task of the DAI group of agents (11 Oa-n) and wherein in the idle status an Al agent is temporally not active in the Al task of the DAI group of agents (11 Oa-n), wherein the DAI group of Al agents (11 Oa-n) comprises a dummy Al agent (110a) and wherein the dummy Al agent (110a) is not capable of being active in the Al task of the DAI group; and controlling (2003) for one or more of the plurality of Al agents (11 Oa-n), based on the determined current Al agent status of each Al agent jointly performing the Al task, a transition from the current Al agent status of the Al agent to a different Al agent status of the plurality of possible Al agent statuses during the Al task.
16. An artificial intelligence, Al, agent (11 Oa-n) for j ointly performing with a plurality of further Al agents (11 Oa-n) of a distributed Al, DAI, group of Al agents (HOa-n) an Al task in a mobile network (100), wherein the Al agent (HOa-n) is configured to: receive control information from an Al controller network entity (120) of the mobile network (100); andtransition, based on the control information, from a current Al agent status of the Al agent to a different Al agent status of a plurality of possible Al agent statuses during the Al task, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status the Al agent is active in the Al task of the DAI group of Al agents (1 lOa-n) and wherein in the idle status the Al agent is temporally not active in the Al task of the DAI group of Al agents (1 lOa-n), wherein the DAI group of Al agents (1 lOa-n) comprises a dummy Al agent (110a) and wherein the dummy Al agent (110a) is not capable of being active in the Al task of the DAI group.
17. The Al agent (HOa-n) of claim 16, wherein the Al agent (HOa-n) is configured to receive from the Al controller network entity (120) information indicative of the respective status of one or more neighboring Al agents of the Al agent.
18. The Al agent (HOa-n) of claim 17, wherein the Al agent is configured to operate depending on the information indicative of the respective status of the one or more neighboring Al agents of the Al agent.
19. The Al agent (11 Oa-n) of any one of claims 16 to 18, wherein the Al agent (11 Oa-n) is configured to be associated with the DAI group of Al agents (1 lOa-n).
20. The Al agent (HOa-n) of any one of claims 16 to 19, wherein, based on the control information, the Al agent is configured to be de-associated from the DAI group of Al agents from the DAI group of Al agents (11 Oa-n) or to be de-activated.
21. The Al agent (11 Oa-n) of any one of claims 16 to 20, wherein, based on the control information, the Al agent is configured to transition from the dummy status to the idle status, from the active status or from the idle status to the dummy status, from the idle status to the active status, or from the active status to the idle status.
22. A method (2100) for operating an artificial intelligence, Al, agent (1 lOa-n) for jointly performing with a plurality of further Al agents (HOa-n) of a distributed Al, DAI, group of Al agents (HOa-n) an Al task in a mobile network (100), wherein the method (2100) comprises: receiving (2101) control information from an Al controller network entity (120) of the mobile network (100); and transitioning (2103), based on the control information, from a current Al agent status of the Al agent to a different Al agent status of a plurality of possible Al agent statuses during the Al task, wherein the plurality of possible Al agent statuses comprise an active status, an idle status, and a dummy status, wherein in the active status the Al agent is active in the Al task of the DAI group of Al agents (11 Oa-n) and wherein in the idle status the Al agent is temporally not active in the Al task of the DAI group of Al agents (11 Oa-n), wherein the DAI group of Al agents (1 lOa-n) comprises a dummy Al agent (110a) and wherein the dummy Al agent (110a) is not capable of being active in the Al task of the DAI group.
23. A computer program product comprising a computer-readable storage medium for storing program code which causes a computer or a processor to perform the method (2000) of claim 15 or the method (2100) of claim 22, when the program code is executed by the computer or the processor.17