WO2023186271A1 - Procédé de sélection d'entités locales pour entraînement d'apprentissage fédéré - Google Patents

Procédé de sélection d'entités locales pour entraînement d'apprentissage fédéré Download PDF

Info

Publication number
WO2023186271A1
WO2023186271A1 PCT/EP2022/058276 EP2022058276W WO2023186271A1 WO 2023186271 A1 WO2023186271 A1 WO 2023186271A1 EP 2022058276 W EP2022058276 W EP 2022058276W WO 2023186271 A1 WO2023186271 A1 WO 2023186271A1
Authority
WO
WIPO (PCT)
Prior art keywords
request
local
network
entities
local entities
Prior art date
Application number
PCT/EP2022/058276
Other languages
English (en)
Inventor
Dario BEGA
Borislava GAJIC
Horst Thomas BELLING
Original Assignee
Nokia Technologies Oy
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Nokia Technologies Oy filed Critical Nokia Technologies Oy
Priority to PCT/EP2022/058276 priority Critical patent/WO2023186271A1/fr
Publication of WO2023186271A1 publication Critical patent/WO2023186271A1/fr

Links

Classifications

    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/16—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks using machine learning or artificial intelligence
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N20/00—Machine learning
    • H—ELECTRICITY
    • H04—ELECTRIC COMMUNICATION TECHNIQUE
    • H04L—TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION
    • H04L41/00—Arrangements for maintenance, administration or management of data switching networks, e.g. of packet switching networks
    • H04L41/50—Network service management, e.g. ensuring proper service fulfilment according to agreements
    • H04L41/5041—Network service management, e.g. ensuring proper service fulfilment according to agreements characterised by the time relationship between creation and deployment of a service
    • H04L41/5051—Service on demand, e.g. definition and deployment of services in real time

Definitions

  • the following disclosure relates to network configuration management . More speci fically, the following disclosure relates to selecting local entities for federated learning training .
  • Modern mobile communication networks are very complicated . In addition to traditional mobile phones they serve a large number of di f ferent types of devices including computers , tablet computers and other consumer computing devices . Furthermore , common consumers may carry several devices that may cause di f ferent data traf fic . In addition to consumer computing devices also other internet connected devices have been introduced . It is common that household devices , cars , working machines and other devices have a mobile communication connection of their own . This has increased a need for data trans fer capacity both in terms of bandwidth and latency . As the networks are more complicated there i s a need for constant monitoring and maintenance for detecting and solving any problems , decreased capacity or any other issues in the network .
  • NWDAF network data analytics function
  • MDAS Management Data Analytics Service
  • OAM Operations , Administration and Maintenance
  • Arti ficial intelligence and machine learning approaches may be used in several functions , for example , for performing continuous assessment of network metrics and statistics .
  • the arrangement provides a continuous learning and evolves normal performance profiles for each service .
  • This continuous operation requires an appropriate model with initial training so that results and predictions given by the arti ficial intelligence and machine learning models are useful . Determining and training arti ficial intelligence and machine learning models is a complicated task and there i s always a continuous need to improve acquisition of the training data that is then used in initial training for machine learning .
  • Federated learning is a machine learning approach where distributed entities train a machine learning model using the local training data .
  • the distributed entities perform the training based on the model acquired from a central entity and report the interim results to the central entity .
  • the central entity aggregates the received interim results , updates a global model accordingly and provides the updated model to the distributed entities .
  • the federated approach provides the means to train the global model in collaborative way by us ing the local data available at end devices without j eopardi zing the privacy of local data .
  • an arrangement for selecting local entities for federated learning training is disclosed .
  • a service consumer responsible for machine learning model requests federated learning training support from a network function .
  • the network function selects local entities for training so that a heterogenous group of local entities is used for training .
  • the service consumer provides the respective instances of a machine learning model to local entities directly or indirectly .
  • Local entities train the received instances and provide an update to the instance as a response .
  • an apparatus comprises at least one processor ; and at least one memory including computer program code , the at least one memory and the computer program code being configured, with the at least one processor, to cause the apparatus to : receive a first request from a network function to select at least one local entity for a federated learning process ; and select at least one local entity in accordance with the received request .
  • the apparatus is configure to receive from the network function information to guide the selection of the at least one local entity, wherein this information comprises at least one of : A group or list of candidate local entities , A desired number of selected local entities , A minimum number of local entities , A maximum number of local entities , A desired percentage of selected local entities relative to the number of candidate local entities , A minimum percentage of selected local entities relative to the number of candidate local entities , A maximum percentage of selected local entities relative to the number of candidate local entities , Desired location of local entities in a form of a target area, duration for learning iteration and desired average latency for federated learning iteration responses , expected number of federated learning iterations , a time when to start federated learning, duration of federated learning, federated learning training support indicator, requirements for capabilities of the local entities , requirements for the hardware or so ftware of the local entities , requirements for the minimum battery level of the local entities , desired response probability for selected local entities , si ze of input model for federated
  • the information to guide the selection o f the at least one local entity comprises a plural ity of sub-areas , and for each sub-area at least one of a desired number, minimum number, maximum number or percentage of local entities that should be selected .
  • the information to guide the selection o f the at least one local entity is received as part of the first request or as part of a request to update this information .
  • the information guiding the selection applies to individual sub-areas or a plurality of sub-areas .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit to at least one NWDAF a second request for subarea information for sub-areas within the area of interest ; and receive from at least one NWDAF information for at least one sub-area as a response to the transmitted second request, wherein the information for the received at least one sub-area is used in selecting the at least one local entity .
  • the received information comprises at least received sub-area types and the respective mapping between the received sub-area type and the responsible NWDAF .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit a second request for information about subareas within the target interest ; receive information for at least one sub-area as a response to the transmitted second request , wherein the information comprises the identity of a network data analytics function being capable of providing additional information related to the sub-area ; in response the received information send a request for additional information related to sub-areas to the network data analytics function; and receive additional information from the network data analytics function .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit identi bombs for the selected at least one local entity as a response to the first request to the network function .
  • the transmitted identi bombs for the selected at least one local entity are associated with identi bombs of the subareas they are located in .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit a third request to update an analytics model to at least one out o f the selected at least one local entity, wherein the third request comprises an input model .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : receive a fourth request from the network function to update at least one analytics model and send the third request in response to the fourth request .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : send the third request in response to the first request .
  • the first request , the third request , or the fourth request comprises an input model for a federated learning cycle such as at least one partial analytics model to be updated .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : send a request for the input model to the network function and receive the input model in response .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the apparatus to : receive in response to the third request at least one update to the analytics model from the at least one out of the selected at least one local entity; transmit the updates to the analytics model to the network function as a response to the received updates to the analytics model .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit to an access management function a fi fth request for one time information about , or for a subscription to noti fications about , at least one of UE connectivity state , UE loss of communication, UE reachability status , registration state changes , UE location, or UE moving in or out of area of intreat ; and receive in response to the fi fth request information about the availability, connectivity state , registration status .
  • Location of local entities wherein the information is used in selecting the at least one local entity .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit to a user data management function a sixth request for one time information about , or for a subscription to noti fications about, at least one of UE reachability, roaming status , CN type change ; and receive in response to the sixth request information about the reachability, roaming status , or core network type , wherein the information i s used in selecting the at least one local entity .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : transmit to a network data analytics function a seventh request for one time information about , or for a subscription to noti fications about , at least one of analytics about local entities , analytics about one or several second network functions serving the local entities , Observed Service Experience related network data analytics , Network Performance Analytics , UE related analytics , UE mobility analytics , User Data Congestion Analytics , or Dispersion Analytics ; and receive in response to the seventh request information about local entities or about the one or several second network function serving the local entities , wherein the information is used in selecting the at least one local entity .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the apparatus to : receive a request to perform a supplementary selection of local entities ; select at least one additional local entity; and transmit identi bombs for the selected additional local entities as a response to the received request .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the apparatus to : receive an eighth request from the network function for subscription about noti fication about changes of the selected at least one local entity; monitor in response to the eighth request information relevant for the selection of the at least one local entity; re-select at least one local entity based on the monitored information; and provide in response to the eight request and the re-selection a noti fication to the network function about the re-selected at least one local entity .
  • the eighth request is combined with the first request .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to : receive a ninth request from the network function for the termination of subscription about noti fication about changes of the selected at least one local entity; terminate the monitoring of information relevant for the selection of the at least one local entity .
  • the terminating the monitoring comprises at least one of : Sending a request for terminating the subscription about event noti fications to a network data analytics function; Sending a request for terminating the subscription about event noti fications to an access management function; and Sending a request for terminating the subscription about event noti fications to a uni fied data management function .
  • the first network function is an application function or a network exposure function .
  • the apparatus is network data analytics function or a network exposure function .
  • the local entity is a user equipment .
  • an apparatus comprising at least one processor ; and at least one memory including computer program code , the at least one memory and the computer program code being configured, with the at least one processor, to cause the apparatus to : transmit to a network function a first request to select at least one local entity for a federated learning process ; receive a response to the transmitted first request , wherein the response comprises at least one local entity identi bomb ; transmit at least one second request to update an analytics model to at least one o f the local entities corresponding to the received at least one local entity identi fiers ; and receive at least one update to the analytics model as a response to the second request .
  • an apparatus comprising : at least one processor ; and at least one memory including computer program code , the at least one memory and the computer program code being configured, with the at least one processor, to cause the apparatus to : transmit to a network function a first request to select at least one local entity for a federated learning process ; transmit at least one second request to update an analytics model to the network function; and receive at least one update to the analytics model as a response to the second request .
  • an apparatus comprising : at least one processor ; and at least one memory including computer program code , the at least one memory and the computer program code being configured, with the at least one processor, to cause the apparatus to : transmit a first request to a network function select at least one local entity for a federated learning process ; Receive a request for at least one input model from the network function; provide in response the input model to the network function; and receive at least one update to the input model .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the network element to send to the network function information to guide the selection of local entities for the federated learning process , wherein this information comprises at least one of : A group or list of candidates local entities , A desired number of selected local entities , A minimum number of local entities , A maximum number of local entities , A desired percentage of selected local entities relative to the number of candidate local entities , A minimum percentage of selected local entities relative to the number of candidate local entities , A maximum percentage of selected local entities relative to the number of candidate local entities , Des ired location o f local entities in a form of a target area, duration for learning iteration and desired average latency for federated learning iteration responses , expected number of federated learning iterations , a time when to start federated learning; duration of federated learning, federated learning training support indicator, requirements for capabilities of the local entities , requirements for the hardware or software of the local entities , requirements for the minimum
  • the information to guide the selection o f the at least one local entity comprises a plural ity of sub-areas , and for each sub-area at least one of a desired number, minimum number, maximum number or percentage of local entities that should be selected .
  • the information guiding the selection applies to individual sub-areas or a plurality of sub-areas .
  • the information to guide the selection of local entities is sent as part of the first request or as part of a request to update this information .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the apparatus to : determine i f the received plurality of local entity identi bombs is suf ficient ; and in case the determination is positive the apparatus is further configured to transmit an acknowledgement ; and in case the determination is negative , the apparatus is further configured to : transmit a request to perform a supplementary selection of local entities ; and receive identi bombs for the selected additional local entities as a response to the received request .
  • the at least one memory and the computer program code are further configured, with the at least one processor, to cause the apparatus to : transmit an eighth request to the network function for subscription about noti fication about changes of the selected at least one local entity; receive in response to the eighth request at least one noti fication about the re-selected at least one local entity .
  • the eight request is combined with the first request .
  • the apparatus is an application function .
  • the network function is network data analytics function or a network exposure function .
  • the local entity is a user equipment .
  • the arrangements discussed above provide di f ferent approaches for performing federated learning process in a mobile communication network or other network .
  • the arrangements discussed above have a benefit of providing a heterogenous selection of local devices for training .
  • the training performed according to the principles discussed above provide an improved representation of statistically di f ferent types of local devices .
  • the selection may be extended so that more or even al l statistically di f ferent groups are well represented even i f they are not available in the initial selection .
  • Fig . 1 is an example of a method in an entity for federated learning
  • Fig . 2 is an example of a method for an arrangement using a federated learning
  • Fig . 3 is a signaling chart of an example ;
  • Fig . 4 is a signaling chart of another example ;
  • Fig . 5 is a signaling chart of another example
  • Fig . 6 is a signaling chart of an example .
  • Fig . 7 is a signaling chart of another example
  • FL is a form of machine learning where instead of model training at a single entity, di f ferent versions of the model are trained at the di f ferent distributed hosts .
  • the obj ective of thi s approach i s to keep the training dataset where it is generated and perform the model training locally at each individual learner in the federation .
  • each individual FM member After training a local model , each individual FM member trans fers its local model parameters , instead of raw training dataset , to the application server or application function acting as an aggregation entity .
  • the application server utili zes the local model parameters to update a global model which may eventually be fed back to the local learners ( i . e . FL members ) for further iterations until global model converges .
  • each local learner benefits from the datasets of the other local learners only through the global model , shared by the aggregator, without explicitly accessing high volume of privacy-sensitive data available at each of the other local learners .
  • an area of interest is an area that the network operator desires to apply a federated learning model .
  • This is an area, which may be also called as a target area or similar area to which the federated learning model could be applied to .
  • the network operator is acquiring information from the area of interest for federated learning purposes . This may include measuring the environmental properties , predicting expected environmental properties , detecting abnormal environmental properties and similar functions at the area of interest and further processing the information before transmitting it to the requesting entity .
  • the environmental properties comprise environmental statistical properties relating to data trans fer capacity, actual data statistics and similar .
  • An example of an area of interest is a city or a portion of a city . The area of interest may be small or large .
  • some embodiments are implanted in a network, wherein geographical areas have been divided into sub-areas that are statistically similar to each other .
  • these areas are used to provide a bas is for heterogenous selection of local entities , that may be any type of user equipment including conventional mobile phones and other mobile network connected devices , such as loT- devices , computers , tablet type devices and similar .
  • the use of di f ferent sub-area types is provided as an example and also di f ferent selection criteria may be used without or in addition the sub-area assisted selection .
  • a sub-area is an area that is partitioned from a larger area belonging to the area of interest .
  • partitioning areas into sub-areas based on environmental properties will be explained .
  • the area of interest typically comprises a plurality of subareas , wherein the environmental properties are di f ferent to the rest of the area . These areas can be detected, for example , from data statistics .
  • the area of interest is partitioned into a plurality of sub-areas using the detected information .
  • Each of the sub-areas are assigned an area type identi fier, wherein the identi fier is associated with particular environmental properties .
  • the sub-area may be divided in di f ferent ways .
  • the sub-area may be partitioned according to detected environmental properties , according to landmarks or similar .
  • One possibility is to partition the area of interest into small areas determined by absolute values , such as the area of the sub-area .
  • the si ze o f the subareas may be described with an area granularity measure , wherein the area granularity describes the si ze and shape of the granules used to determined sub-areas .
  • the known granularity may be used in implementing a service relating to detection of abnormal environmental properties .
  • the analytics model may be an arti ficial intelligence or machine learning based analytics model , wherein the model parameters and the training data is configured to produce analytics and predictions .
  • the analytics model may be used for predicting the load of the area, which the analytics model is covering .
  • the analytics model may be applied per sub-area so that the sub-areas having the same sub-area type identi bomb may share the same analytics model .
  • the di f ference between sub-area type identi bombs may be limited to di f ferent training data of the analytics model or the model as whole may be di f ferently configured .
  • Fig . 1 is explaining a basic principle of an arrangement with federal learning arrangement is disclosed .
  • the method is implemented in an entity, which is responsible for selecting the local entities for federated learning .
  • This may be a speci fic federated learning server, NWDAF (Network Data Analytics Function) or similar entity that has been implemented combining the federated learning server, NWDAF and other similar components .
  • NWDAF Network Data Analytics Function
  • the main di f ference between the entity type is how the messaging is arranged .
  • the federated learning server may have own speci fic messaging, while NWDAF may incorporate this messaging as additional parameters in more conventional analytics requests .
  • These analytics requests may be implemented as separate individual requests or subscriptions , wherein the subscription provides the information, for example , on a regular time interval basis or by any other trigger .
  • the process begins so that an application function transmits a request for federated learning training support .
  • This request is received at a federated learning server, NWDAF or any other entity responsible for providing the local entities , step 100 .
  • NWDAF federated learning server
  • the responsible entity then generates a list local entities , such as user equipments , that will be used in the federated learning process .
  • a service consumer such as NWDAF service consumer or any other application function subscribes or requests federated learning training support . This may be performed, for example , by requesting or subscribing an analytics service using a flag or parameter indicating that federated learning training support is requested, step 100 .
  • the request is transmitted to an entity, which may be the service providing entity, a federated learning server or any similar entity .
  • entity which may be the service providing entity, a federated learning server or any similar entity .
  • the entity acquires , a list of di f ferent subarea types within the area of interest .
  • the di f ferent sub-area types may be acquired by requesting sub-area type identi bombs from Analytics Area Type Properties Function (AATPF) .
  • AATPF transmits the available sub-area type identi bombs to NWDAF .
  • the NWDAF processes the received sub-area type identi bombs and selects user equipments from each subarea type , step 110 .
  • the selection may be done according to additional parameters that have been received at step .
  • the NWDAF has a list of user equipments that match the request and represent statistically di f ferent sub-area types .
  • the list comprises a heterogenous selection of user equipments that may be used for federated learning purposes .
  • the list is provided to the service consumer which is in this example the NWDAF service consumer . In the above it is assumed that the list was success fully generated, however, it is possible that the requirements speci fied by the service consumer cannot be satis fied . In such case , the NWDAF informs the service consumer that the requirements cannot be satis fied .
  • the service consumer When the service consumer has received the list , for each training epoch, the service consumer forwards an instance of the analytics model to the selected user equipments .
  • the user equipments each use their own local data to train the received instance of the analytics model .
  • the further training of the analytics model improves and updates the analytics model .
  • the trained analytics model is transmitted back to the service consumer . This may be done by transmitting a complete instance of the analytics model or instead of a complete model only the updated parts or gradients of the analytics model .
  • a method for an arrangement us ing a federated learning model is disclosed .
  • the method includes in step 210 transmitting to a network function a first request to select at least one local entity for a federated learning process ;
  • the method includes in step 220 receiving a response to the transmitted first request , wherein the response comprises at least one local entity identi bomb ;
  • the method includes in step 230 transmitting at least one second request to update an analytics model to at least one o f the local entities corresponding to the received at least one local entity identi bombs ; and
  • the method includes in step 240 receiving at least one update to the input model as a response to the second request.
  • Fig. 3 discloses a specific example of an arrangement similar to the one of Fig. 1.
  • an NWDAF service consumer subscribes to an analytics service with an attribute "FL training support".
  • the service consumer is an application function (AF) , however, it could also be for example, an 0AM (Operations, Administration and Maintenance) .
  • the NWDAF selects from each sub-area type within the AOI, UEs, i.e., local entities, to participate in the training of the ML model in a federated learning approach. The selection is based on the service consumer' s requirements .
  • Signal 1 indicates that AF subscribes to analytics service information. It sets a plurality of input parameters including the Analytics ID indicating the selected analytics and at least "FL training support" attribute to True. Furthermore, the service consumer specifies at least one area of interest. There may be additional and optional attributes, such as a reporting percentage attribute. In addition, the service consumer may indicate the local entities requirements specifying the requirements to be satisfied when selecting local entities that will participate to the training of the machine learning model.
  • Signal 2 indicates that NWDAF requests from an AATPF the sub-area type identifiers within the area of interest specified by the service consumer.
  • the request is based on the data related to local entities, i.e., the data required to produce the requested analytics.
  • the NWDAF subscribes to the analytics service generating the sub-area types based on local entities related data.
  • the AATPF replies with the subarea type identi bombs within the area of interest and optionally also the list of NWDAFs managing similar subareas type in di f ferent areas of interest .
  • the NWDAF may coordinate with other NWDAFs that manage similar sub-areas , in order to select suitable local entities from other areas of interest but handling data with similar statistics . This may also be leveraged by the NWDAF in order to perform for example load balancing and so selecting local entities from not overloaded subareas .
  • the above is possible in an arrangement , where the selection of additional substitutes can be directed to similar areas or otherwise to devices that share similar statistics .
  • the NWDAF selects at step 4 local entities from each subarea type within the area of interest . This may be done in a way to ensure that each sub-area type which characteristic data statistics is well represented ( takes substantial part ) in the training procedure . Thus , the selected local entities are heterogenous and the selection covers well all di f ferent types of devices within the area of interest . This may be done by selecting suf ficient number of local entities for each sub-area type . As the sub-area types are di f ferent to each other by selecting a suf ficient number from each sub-area type provides a heterogenous selection of the devices .
  • the service consumer may speci fy di f ferent weights for each sub-area type or the NWDAF may automatically decide the best proportion . When the selection is not complete enough, more local entities can be selected as mentioned above .
  • the NWDAF coordinates with other NWDAFs as detailed in signal 3 .
  • the NWDAF may trigger analytics services , such as "UE mobility analytics" to determine the UEs that will be available in that sub-area for the time needed to proces s at least a training cycle .
  • the time window may be speci fied in the service consumer request .
  • the NWDAF may trigger other analytics services , such as , "UE communication analytics" to determine the local entities with higher traf fic volume , or "User data congestion analytics” to determine congestion with regard area/user and so avoid to select overloaded local entities and/or local entities from overloaded subareas .
  • the NWDAF may subscribe to the AATPF in order to be noti fied in case a sub-area of interest changes its data statistics . In this case , the NWDAF informs the service consumer to discard the latest updates received by the impacted local entities and will perform a new UEs selection .
  • the analytic services mentioned above are j ust examples and also others may be used .
  • Signal 5 is sent i f the requirements speci fied by the service consumer can be satisfied .
  • the NWDAF forwards to the service consumer the selected local entities identi bombs that will take part in the training of the machine learning model .
  • the NWDAF informs the service consumer about the sub-area type identi bomb o f each local entity . I f the selection according to the requirements cannot be made, the NWDAF informs the service consumer that the requirements cannot be satis fied .
  • the service consumer may decide to relax the requirements or wait until the selection can be obtained .
  • the service consumer may distribute the machine learning model to the local entities in the received selection list .
  • the input model is used to re fer to this portion o f the complete machine learning model .
  • the corresponding response will be nominated as an update to the input model .
  • the update to the input model may be provided as a complete model , as upgraded portion, gradient or any other means that facilitate merging the update to the input model to the complete machine learning model .
  • the service consumer forwards to the selected local entities an input model .
  • the input model is an instance of the model to be trained by using their local data .
  • the local entities train the input with their local data, step 8 .
  • the trained and updated model is sent in signal 8 back to to the service consumer .
  • step 9 the service consumer veri fies that enough updates have been received from each local entity and checks that each sub-area type identi fied is represented as speci fied by reporting percentange , i f the reporting percentage was provided in the requirements .
  • the veri fication step may include a time interval , which may be required in case o f synchronous federated learning process .
  • I f the selection provided by NWDAF has not been success ful and enough updates have not been received, the service consumer informs the NWDAF with signal 10a that a new selection is needed .
  • step I l a is performed by the NWDAF and a new selection is coordinating, typically with other NWDAFs as detailed in connection of signal 3 .
  • the new local entities selection is forwarded using signal 12a to the service consumer and the training epoch is repeated starting from signaling 6 .
  • the service consumer aggregates them and updates the general machine learning model at step 10b .
  • the service consumer informs the NWDAF using signal 11b that the selection was success ful .
  • the NWDAF optionally veri fies in step 12b i f the local entities are still available for a new training epoch or i f an updated selection is needed .
  • the NWDAF receives the possible additional selection using signal 13b .
  • Fig . 4 examples of signaling using an arrangement with federated learning server has been disclosed .
  • the service consumer is assumed to be an AF .
  • the AF transmits requests relating to federated learning process to an FL-server .
  • the FL- server could be integrated to di f ferent components , such as NWDAF .
  • NWDAF di f ferent components
  • signal AF instructs us ing s ignal 1 the FL- server to propose UEs for federated learning .
  • the AF may speci fy overall response time for learning cycle , expected time for local entities learning cycle , desired locations of local entities and time when to start the training cycle .
  • a more detailed list of input parameters is provided in below .
  • the parameters described there are j ust examples , and the list is not intended to be an exhaustive list .
  • the FL-server may subscribe/reques t analytics from NWDAF, signal 2 , to support selecting UEs as described in the above .
  • the NWDAF provides analytics requested by FL-server by responding using signal 3 . Based on the retrieved information the FL-server selects suitable local entities that ful fill the AF' s requirements .
  • the chart includes two options .
  • the FL- server returns the list of selected local entities to the AF using signal 5a .
  • the AF is responsible to manage the federated learning training task and sends input models to the selected local entities the AI /ML model for training according to signal 6a .
  • the local entities train the input model with their local data, step 7a .
  • the local entities send the updates to the input model back to the AF the updated ML model using signal 8a .
  • the AF provides to the FL- server the AI/ML model to be trained by local entities , signal 5b .
  • the FL-server takes responsibility of di stributing of the partial model s as input models and receiving corresponding updates to the input model .
  • the FL-server manages the federated learning training process and sends the respective input models to selected local entities using signal 6b .
  • the local entities train their instance of the model with their local data as in the other examples in step 7b . After training they send back their updates to the input model to the FL-server using signal 8b .
  • the FL-server sends the updated machine learning models to AF using signal 9 .
  • This submission may be a complete machine learning model , updates to the input model , updated portions , gradients or any other means that facilitate updating the machine learning model so that the AF can have an updated model .
  • the updated model may be del ivered as a complete updated model , as updates to the earlier model or as an update indicating di f ferences , gradient or similar .
  • Steps and signal s 10 - 14 correspond with the steps 9 - 13 of the example of Fig . 3 .
  • the di f ference between examples of Fig . 3 and Fig . 4 is that the signals in the example of Fig . 4 are performed between the AF and the FL-server, while in the example of Fig . 3 the signaling is between the AF and NWDAF . Even i f the signaling chart of Fig . 4 does not show any signaling between the AF and NWDAF there may be signaling between those two when the signaling is not related to the federated learning process .
  • Fig . 5 discloses a more detailed signaling chart , which is in accordance with the second option of the example of Fig . 4 .
  • the example o f Fig . 5 the f irst step 501 is an AF establishing a federated learning session for several federated learning cycles .
  • the AF provides at least some of the above parameters when establishing the session . However, it may also provide a new input model that may be updated for each FL cycle using signal 513 .
  • the server may query or subscribe at AMF to UE related events such as "connectivity state” , "UE loss of communication” , “UE reachability status” , “registration state changes” , "UE location” , or “UE moving in or out of area of intreat” .
  • UE related events such as "connectivity state” , "UE loss of communication” , “UE reachability status” , “registration state changes” , "UE location” , or “UE moving in or out of area of intreat” .
  • This may be done to inquire or monitor availability of UEs and preferably select connected, active , registered, and/or reachable UEs and/or UEs within a desired region .
  • the FL-server may query or subscribe at UDM for UE reachability, roaming status , CN type change , and preferably select reachable , nonroaming UEs at preferred CN type .
  • the FL-server may inquire analytics about UE or network serving the UE . These may be , for example , Observed Service Experience related network data analytics , Network Performance Analytics , UE related analytics , UE mobility analytics , User Data Congestion Analytics , and/or Dispersion Analytics as defined in 3GPP TS 23 . 288 . Using the inquired information the FL-server may select less UEs served by less loaded network components . In a variant the server may inquire predictions about time-dependent network load or UE load in steps 510 and 511 and start federated learning cycle at time with lower load for some or al l selected UEs .
  • the FL-server selects the UEs for the federated learning cycle from the group of UEs , step 512 .
  • the AF transmits the signal 513 instructing the FL- server to perform a federated learning cycle on a group of UEs .
  • the server may inquire or subscribe to UE locations ( e . g . via AMF) and predictions ( e . g . from NWDAF) about network load at UE locations and select UEs located in network areas with lower load .
  • the AF indicates list or group of UEs to choose from, number of UEs required to perform federated learning cycle , input model (AI /ML model for training) , overall response time for learning cycle , possibly expected time for a UE to perform federated learning cycle , possibly desired location of UEs . Furthermore , the AF may indicate time when to start federated learning cycle , possibly si ze of input and/or update to the input model (AI /ML model updated parameters/gradients ) . This indication may be done using signal 501 or 513 .
  • the FL-server After signal 513 has been received at the FL- server, the FL-server sends to each selected UE a request to start federated learning cycle and provides input model in signal 514 . In a variant the FL-server monitors whether selected UEs perform federated learning cycle . In another variant UEs send acknowledgment for request and/or progress reports at regular intervals .
  • Federated learning server After performing the federated learning cycle the UEs transmit the update to the input model .
  • Federated learning server receives the signal 515 update to the input model from each selected UE and forwards it to the AF using signal 516 .
  • FL-server monitors whether each selected UE sends update to the input model within expected response time .
  • the FL-server observes that some UEs are not sending acknowledgment , progress report , or update to the input model within the expected time or is noti fied that some selected UEs are no longer registered, active , or reachable , or left an area of interest , the FL-server selects additional UEs to compensate for those UEs and sends to each selected additional selected UE a request to start federated learning cycle and provides input model .
  • AF requests termination of federated learning session from federated learning server using signal 517 .
  • the FL-server terminates related events and analytics subscriptions at AMF, UDM, and/or NWDAF using signals 518 - 520 .
  • Fig . 6 discloses a more detailed signaling chart , which is in accordance with the first option of the example of Fig . 4 .
  • the steps 601 - 612 and 617 - 620 are similar to the example of Fig . 5 .
  • the signal 613 transmitted from the FL-server to the AF provides a list of suggested UEs to the AF .
  • an optional s ignal 614 is transmitted for updating the list of suggested UEs . This is done i f UEs become unavailable or no longer meet selection criteria .
  • the FL-server may suggest new UEs or remove UEs that are not anymore available .
  • the AF transmits the input model directly to UEs using signal 615 and receives the update to the input model as a response from the UEs using signal 616 after the UEs have trained the input model using the local data .
  • Table 1 Examples of input parameters for federated learning training support
  • the example request is for an implementation, wherein the service consumer communicates with the NWDAF .
  • the NWDAF can provide Observed Service Experience analytics ( statistics or prediction) as described in TS 23 . 288 .
  • Such service experience may be related to speci fic network slice and can be expressed by, for example , average of observed Service MoS and/or variance of observed Service MoS indicating service MOS distribution for services such as audio-visual streaming, V2X and Web browsing .
  • the service consumer requiring the support for federated learning model training which provides the network slice service experience may subscribe to the NWDAF using service attribute " FL training support" in its request .
  • o S-NSSAI identifies the Network Slice for which analytics information is subscribed or requested o Area of Interest - Identifies the Area (i.e. set of TAIs) ,for which the analytics is requested. o NSI ID (Optionally) - Identifies the Network Slice instance (s) for which analytics information is subscribed or requested.
  • the NWDAF collects the required input data unless it is already available.
  • the data may include, for example, UE level Network Data from 0AM related to the QoS profile, QoS flow level Network Data from 5GC NF related to the QoS profile assigned for a particular service etc. cf TS 23.288.
  • Fig. 7 discloses another more detailed signaling chart.
  • the example of Fig. 7 addresses aspects of key issue #7 on 5GS Assistance to Federated Learning Operation and key issue #3 on 5GC Information Exposure to authorized 3rd party for Application Layer Al / ML Operation.
  • the example allows the 5GS to suggest devices that should be part of FL operation, based on requirements expressed by the AI/ML application server.
  • the example relies on introduction of a Federated Learning server to select UEs suitable to participate to the FL operation, when receiving a request from the AI/ML application server (acting as AF) .
  • the AI/ML application server tasks the Federated Learning server to propose UEs for FL. Based on suggestion from the Federated Learning server, the AI/ML application server can trigger FL operations.
  • the Federated Learning server can be integrated in NEF, in NWDAF or be standalone.
  • the Federated Learning server monitors whether suggested UEs remain available (e.g. registered, active, reachable) , or still meet other selection criteria e.g. are still in the area of interest as proposed by the AI/ML application server. If UEs become unavailable or no longer meet selection criteria, the Federated Learning server notifies the AF and may suggests new UEs. The AF then sends to suggested UEs a request to start federated learning cycle (s) and provides input model and receives update to the input model from each UE for each learning cycle.
  • suggested UEs remain available (e.g. registered, active, reachable)
  • other selection criteria e.g. are still in the area of interest as proposed by the AI/ML application server. If UEs become unavailable or no longer meet selection criteria, the Federated Learning server notifies the AF and may suggests new UEs. The AF then sends to suggested UEs a request to start federated learning cycle (s) and provides input model and receives update to the input model from each UE for each learning cycle
  • the Federated Learning server may inquire or subscribe to UE locations (e.g. via AMF) and predictions (e.g. from NWDAF) about network load at UE locations and select the UEs located in network areas with lower load.
  • the Federated Learning server may inquire analytics about UE or network serving the UE (e.g. from NWDAF, e.g. Observed Service Experience related network data analytics, Network Performance Analytics, UE related analytics, UE mobility analytics, User Data Congestion Analytics, and/or Dispersion Analytics as defined in 3GPP TS 23.288) and may preferably select UEs served by less loaded network components.
  • NWDAF Observed Service Experience related network data analytics
  • Network Performance Analytics e.g. Observed Service Experience related network data analytics
  • UE related analytics e.g. Observed Service Experience related network data analytics, Network Performance Analytics, UE related analytics, UE mobility analytics, User Data Congestion Analytics, and/or Dispersion Analytics as defined in 3GPP TS 23.288
  • Dispersion Analytics as defined in 3GPP TS 23.288
  • the Federated Learning server may query or subscribe at AMF to UE related events such as connectivity state, UE loss of communication, UE reachability status, registration state changes, UE location, or UE moving in or out of area of intreat, to inquire or monitor availability of UEs and preferably select connected, active , registered, and/or reachable UEs and/or UEs within a desired region .
  • UE related events such as connectivity state, UE loss of communication, UE reachability status, registration state changes, UE location, or UE moving in or out of area of intreat.
  • the Federated Learning server may query or subscribe at UDM for UE reachability and preferably select reachable UEs .
  • the Federated Learning server shall query or subscribe at UDM for roaming status and shall only select non-roaming UEs .
  • NEE is not shown in the figure , however, the 3rd party AF would first interact with NEE which then would in turn interact with the Federated Learning server .
  • step 1 the AI /ML application server ( acting as AF) prepares for running FL training by collecting information such as potential FL members , FL aggregated model si ze , local model size , model aggregation average latency target for FL iterations , expected number of iterations and time interval between iteration, time window when FL training needs to be performed . It determines an initial list for candidate FL members .
  • the AF tasks Federated Learning Server (which could be integrated in NEF, NWDAF or be standalone ) to propose UEs for FL .
  • the AF indicates a list or group of UEs to choose from, the number of UEs required to perform federated learning cycle , possibly desired location of UEs in a form of a target area for FL, duration for iteration and desired average latency for FL iterations .
  • the AF may provide sub-areas , and provide a percentage o f UEs that should take part in FL from each sub-area, and/or a minimum number of UEs and /or a maximum number of UEs that should take part in FL from each sub-area .
  • the AF may provide time when to start federated learning, duration of federated learning , si ze of input model and/or update to the input model .
  • step 2 is shown as AIML_session Create Request , however, step 2 could also be reali zed via an analytics request / subscription with new analytics type .
  • the Federated Learning Server may query or subscribe at AMF to UE related events such as connectivity state , UE loss of communication, UE reachability status , registration state changes , UE location, or UE moving in or out of area of interest , to inquire or monitor availability of UEs and preferably select connected, active , registered, and/or reachable UEs and/or UEs within a desired region .
  • UE related events such as connectivity state , UE loss of communication, UE reachability status , registration state changes , UE location, or UE moving in or out of area of interest .
  • the Federated Learning Server may query or subscribe at UDM for UE reachability and preferably select reachable UEs .
  • the Federated Learning server shall query or subscribe at UDM for roaming status and shall only select non-roaming UEs .
  • the Federated Learning Server may inquire analytics about UE or network serving the UE (e . g . from NWDAF, e . g . Observed Service Experience related network data analytics , Network Performance Analytics , UE related analytics , UE mobility analytics , User Data Congestion Analytics , and/or Dispers ion Analytics as de fined in 3GPP TS 23 . 288 ) and may preferably select UEs served by less loaded network components .
  • the federated learning Server select UEs for the federated learning from the group of UEs .
  • the Federated Learning Server may select more UEs than requested to prepare for unresponsive UEs .
  • the federated learning server provides a list of suggested UEs to the AF .
  • the AF sends to suggested UEs a request to start federated learning cycle ( s ) with input model and receives update to the input model from each UE for each learning cycle .
  • step 11 and signal 12 the federated learning server monitors whether suggested UEs remain available ( registered, active , reachable ) , or still meet other selection criteria e . g . area of interest . I f UEs become unavailable or no longer meet selection criteria federated learning server noti fies AF .
  • the Federated Learning Server may also suggest new UEs for FL operation .
  • Signals 13 - 16 show when the AF requests termination of federated learning session from Federated Learning Server, the Federated Learning Server terminates related events subscriptions and analytics subscriptions at AMF, UDM, and/or NWDAF .
  • the components of the exemplary embodiments can include a computer readable medium or memories for holding instructions programmed according to the teachings of the present inventions and for holding data structures , tables , records , and/or other data described herein .
  • a computer readable medium can include any suitable medium that participates in providing instructions to a processor for execution .
  • Computer-readable media can include , for example , a floppy disk, a flexible disk, hard disk, magnetic tape , any other suitable magnetic medium, a CD- ROM, CD ⁇ R, CD1RW, DVD, DVD-RAM, DVD1RW, DVD1R, HD DVD, HD DVD-R, HD DVD-RW, HD DVD-RAM, Blu-ray Disc, any other suitable optical medium, a RAM, a PROM, an EPROM, a FLASH-EPROM, any other suitable memory chip or cartridge , a carrier wave or any other suitable medium from which a computer can read .
  • the base station and/or user equipment may comprise a circuitry .
  • circuitry may refer to one or more or all of the following :
  • hardware circuit (s) and/or processor ( s ) such as microprocessor ( s ) or a portion of microprocessor ( s ) that requires software (e.g., firmware) for operation, but the software may not be present when it is not needed for operation.
  • software e.g., firmware
  • circuitry also covers an implementation of merely a hardware circuit or processor (or multiple processors) or portion of a hardware circuit or processor and its (or their) accompanying software and/or firmware.
  • circuitry also covers, for example and if applicable to the particular claim element, a baseband integrated circuit or processor integrated circuit for a mobile device or a similar integrated circuit in a server, a cellular network device, or other computing or network device.
  • An example of an access architecture that may be applied may be e.g. a radio access architecture based on long term evolution advanced (LTE Advanced, LTE-A) or new radio (NR, 5G) , without restricting, however, the example embodiments to such an architecture. It is obvious for a person skilled in the art that the example embodiments may also be applied to other kinds of communications networks having suitable means by adjusting parameters and procedures appropriately.
  • LTE Advanced long term evolution advanced
  • NR new radio
  • the depicted system is only an example of a part of a radio access system and in practice, the system may comprise a plurality of (e/g)NodeBs, the user equipment may have an access to a plurality of radio cells and the system may comprise also other apparatuses, such as physical layer relay nodes or other network elements, etc. At least one of the (e/g)NodeBs may be a Home ( e/g) nodeB . Additionally, in a geographical area of a radio communication system a plurality of different kinds of radio cells as well as a plurality of radio cells may be provided.
  • Radio cells may be macro cells (or umbrella cells) which are large cells, usually having a diameter of up to tens of kilometers, or smaller cells such as micro-, femto- or picocells.
  • the (e/g)NodeBs may provide any kind of these cells.
  • a cellular radio system may be implemented as a multilayer network including several kinds of cells. Typically, in multilayer networks, one access node provides one kind of a cell or cells, and thus a plurality of (e/g)NodeBs are required to provide such a network structure.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Signal Processing (AREA)
  • Theoretical Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Medical Informatics (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Artificial Intelligence (AREA)
  • Data Mining & Analysis (AREA)
  • Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Databases & Information Systems (AREA)
  • Computer And Data Communications (AREA)

Abstract

L'invention concerne un agencement qui sélectionne des entités locales pour un entraînement d'apprentissage fédéré. Dans l'agencement, un consommateur de service responsable d'un modèle d'apprentissage automatique demande une assistance d'entraînement d'apprentissage fédéré à une fonction de réseau. La fonction de réseau sélectionne des entités locales à entraîner de telle sorte qu'un groupe hétérogène d'entités locales est utilisé pour l'entraînement. Ensuite, le consommateur de service fournit les instances respectives d'un modèle d'apprentissage automatique aux entités locales directement ou indirectement. Les entités locales entraînent les instances reçues et fournissent une mise à jour de l'instance en tant que réponse.
PCT/EP2022/058276 2022-03-29 2022-03-29 Procédé de sélection d'entités locales pour entraînement d'apprentissage fédéré WO2023186271A1 (fr)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/EP2022/058276 WO2023186271A1 (fr) 2022-03-29 2022-03-29 Procédé de sélection d'entités locales pour entraînement d'apprentissage fédéré

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/EP2022/058276 WO2023186271A1 (fr) 2022-03-29 2022-03-29 Procédé de sélection d'entités locales pour entraînement d'apprentissage fédéré

Publications (1)

Publication Number Publication Date
WO2023186271A1 true WO2023186271A1 (fr) 2023-10-05

Family

ID=81388869

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/EP2022/058276 WO2023186271A1 (fr) 2022-03-29 2022-03-29 Procédé de sélection d'entités locales pour entraînement d'apprentissage fédéré

Country Status (1)

Country Link
WO (1) WO2023186271A1 (fr)

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021032497A1 (fr) * 2019-08-16 2021-02-25 Telefonaktiebolaget Lm Ericsson (Publ) Procédés, appareils et supports lisibles par ordinateur relatifs à l'apprentissage automatique dans un réseau de communication

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021032497A1 (fr) * 2019-08-16 2021-02-25 Telefonaktiebolaget Lm Ericsson (Publ) Procédés, appareils et supports lisibles par ordinateur relatifs à l'apprentissage automatique dans un réseau de communication

Non-Patent Citations (4)

* Cited by examiner, † Cited by third party
Title
"3rd Generation Partnership Project; Technical Specification Group Services and System Aspects; Study on 5G System Support for AI/ML-based Services (Release 18)", 4 March 2022 (2022-03-04), XP052124357, Retrieved from the Internet <URL:https://ftp.3gpp.org/tsg_sa/WG2_Arch/Latest_SA2_Specs/Latest_draft_S2_Specs/23700-80-010.zip 23700-80-010_rm.doc> [retrieved on 20220304] *
"3rd Generation Partnership Project; Technical Specification Group Services and System Aspects; Study on enablers for network automation for the 5G System (5GS); Phase 2 (Release 17)", no. V17.0.0, 17 December 2020 (2020-12-17), pages 1 - 382, XP051975177, Retrieved from the Internet <URL:https://ftp.3gpp.org/Specs/archive/23_series/23.700-91/23700-91-h00.zip 23700-91-h00.docx> [retrieved on 20201217] *
3GPP TS 23.288
NOKIA ET AL: "KI #7 & #3, New Solution: Federated Learning Server assisting on federated learning members selection", vol. SA WG2, no. Elbonia; 20220406 - 20220412, 29 March 2022 (2022-03-29), XP052133206, Retrieved from the Internet <URL:https://ftp.3gpp.org/tsg_sa/WG2_Arch/TSGS2_150E_Electronic_2022-04/Docs/S2-2202364.zip S2-2202364-Federated-Learning-server.docx> [retrieved on 20220329] *

Similar Documents

Publication Publication Date Title
US11418413B2 (en) Sharable storage method and system for network data analytics
US20230090022A1 (en) Method and device for selecting service in wireless communication system
EP2204056B1 (fr) Mise en grappes de service de mobilité à l&#39;aide de segments de service de réseau
EP4087193A1 (fr) Procédé, appareil et système d&#39;analyse de données
US20220103443A1 (en) Methods and devices for operation of a network data analytics function
EP3955523A1 (fr) Composant d&#39;analyse de réseau et procédé pour la fourniture d&#39;une analyse de réseau et/ou d&#39;informations de prédiction concernant des instances de tranche d&#39;un réseau de communication mobile
CN109995845B (zh) 一种控制面资源迁移的实现方法、装置及网络功能实体
CN114616846A (zh) 用于管理边缘应用服务器的发现的方法和系统
US20220103644A1 (en) Method and apparatus for renewing subscription for network data analysis in wireless communication system
US20230239680A1 (en) Method and device for supporting mobility for collecting and analyzing network data in wireless communication network
US11564156B2 (en) Provision of data analytics in a telecommunication network
KR20200116844A (ko) Nwdaf를 위한 nf 장치로부터의 네트워크 데이터 수집 방법
US20220330303A1 (en) Network monitoring in service enabler architecture layer (seal)
EP4222935A1 (fr) Technique de commande de rapport d&#39;événements de réseau
WO2023186271A1 (fr) Procédé de sélection d&#39;entités locales pour entraînement d&#39;apprentissage fédéré
CN113596932A (zh) 信息提供、生成、目标基站确定方法及设备、介质
US11706093B2 (en) Auto switching for enterprise federated network slice
WO2023052010A1 (fr) Collecte de données de trajectoire dans des systèmes de télécommunication mobile
JP7389919B2 (ja) ネットワーク分析構成要素とネットワーク分析情報の提供方法
EP4135413A1 (fr) Agencement de réseau de communication et procédé de sélection de composant de réseau
EP4254896A1 (fr) Coordination d&#39;entraînements de modèles pour un apprentissage fédéré
CN116938747A (zh) 一种通信方法及装置
US20240040366A1 (en) Methods, systems, and computer readable media for rebalancing subscriber location function (slf) subscriber data
WO2023061593A1 (fr) Procédé de détection et de gestion de propriétés de réseau par zone
JP2023527499A (ja) 通信ネットワーク分析を実行する機械学習モデルを提供する通信ネットワーク配置及び方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22719265

Country of ref document: EP

Kind code of ref document: A1