Customer selection for a telecommunications network to implement federated learning

Abstract

Selection of clients from a telecommunications network for the implementation of federated learning. The invention relates to a method for managing federated learning comprising the following steps of a typical phase (30): - connection (300) of a client, obtaining (301) the position and speed of the client and estimating (302) based on the speed and / or position of the client, the connection time of said client to the base station by applying a prediction model; - upon initialization of a federated learning cycle, selection (305) of a subset of clients, each having a remaining connection time greater than an estimate of the participation time required for the federated learning cycle; - transmission (306) of initial parameter values ​​to the subset; - obtaining (308) optimized values ​​from the clients of the subset; - determination (309) of global values, based on the optimized values;- transmission (310) of the overall values ​​to the customers of said set. Figure for the abbreviation: [Fig 3];

Description

Title of the invention: Selection of clients of a telecommunications network for the implementation of federated learning Technical field

[0001] The invention relates to the field of telecommunications networks.

[0002] Its aim is in particular to enable federated learning between client terminals having access to a telecommunications network. State of the art

[0003] Entities connected, hereinafter called clients, to a telecommunications network, such as smartphones, vehicles, connected objects accessing a mobile network of the cellular type for example, generate data during their operation. Such data can be used to train models by machine learning, such models having very varied applications.

[0004] However, this data is often voluminous on the one hand, and may include sensitive data, in particular data from the private lives of the customers' users.

[0005] For example, the data may include emails, recorded phone calls, or the customer's location.

[0006] It is therefore difficult to train a model on a global database bringing together all the data locally obtained by clients connected to a telecommunications network. Indeed, such an approach would generate significant costs linked to the transmission of data collected locally by the clients, to the database, as well as security problems linked to the confidentiality of the transmitted data.

[0007] An alternative approach is called federated learning.

[0008] According to such an approach, learning is distributed among the different clients. The clients thus collaborate to build a global model, the collaboration only involving the transfer, to an aggregator common to several clients, of parameters of a local model, the parameters being optimized locally by each client.

[0009] Clients then do not have to transmit the locally accumulated data, on the basis of which the parameters of the local model are optimized, only the optimized parameters being transmitted to the aggregator.

[0010] Such an approach thus makes it possible to overcome both the cost problems, the quantity of information exchanged being drastically reduced, but also the confidentiality problems, since no personal data is transmitted on the network when aggregating the optimized parameters. The aggregator then aggregates the parameters received from different clients into a set of optimized parameters of a global model, or global parameters. The global parameters are then sent back to the clients for updating their respective local models from the global parameters.

[0011] However, due to the bandwidth limitations of telecommunications networks, as well as the high number of clients simultaneously connected to such a telecommunications network, especially in cellular networks, the aggregator does not involve all connected clients in the learning process but selects some of them.

[0012] The selection of clients participating in supervised learning is implemented by the aggregator according to a given approach. Such an approach may be based on a random selection assigning the same selection probability to all clients connected to the same aggregator.

[0013] The disadvantage of such a random approach is that it does not take into account the characteristics specific to each client, which can considerably influence the learning performance or the operation of the client. Indeed, for a local learning model to converge, the client must have sufficient resources.

[0014] Other approaches propose to select clients based on their respective local resources or based on the respective performances of their local models.

[0015] In the first case, clients with the most resources are selected as a priority for federated learning.

[0016] In the second case, the clients with the highest local learning performance are selected as a priority, considering that this allows the global model to converge more quickly.

[0017] However, both of these approaches assume that clients are static or not very mobile, and that their connection time to the network, or association time to the aggregator, is long enough to allow the implementation of federated learning.

[0018] However, in certain contexts, such as those associated with a cellular mobile network, clients may be mobile, particularly when the clients are vehicles or are on board vehicles.

[0019] In these contexts, taking into account only the local resources of each client or the performance of the respective local models of the clients can lead to selecting clients which cannot participate in all of the federated learning stages, or federated learning cycle, because of their high mobility.

[0020] As a result, the selection of clients of a telecommunications network, for the implementation of federated learning of a model, is unsuitable for contexts in which the clients are mobile.

[0021] The invention offers a solution which does not have the drawbacks of the state of the art. Statement of the invention

[0022] To this end, according to a functional aspect, the invention relates to a method for managing federated learning between several client devices, called clients, capable of communicating with a base station of a telecommunications network associated with a radio coverage area, a set of clients being capable of communicating, at a given instant, with the base station, the clients of said set comprising respective application modules of a given function, the method comprising the following steps, during a current phase: - upon connection of a client to the base station, obtaining at least one piece of information from among the position and the speed of the client, estimating, as a function of said at least one piece of information, a connection duration of said client to the base station, by applying a connection duration prediction model, and adding said client to said set of clients; - upon initialization of a federated learning cycle, selection of a subset of clients from among the clients of said set, so that, for any client of the subset, a remaining connection duration is greater than an estimate of a participation duration necessary for the federated learning cycle; - transmission of initial parameter values ​​to the clients of the subset, for configuration of the application modules and for training the application modules from local data of the clients of the subset; - obtaining optimized parameter values ​​from at least some of the clients in the subset; - determination of global parameter values, based on the optimized parameter values ​​received; - transmission of the global parameter values ​​to at least some of the clients of said set for configuration of the application modules from the global parameter values.

[0023] Such a transmission may be a multicast transmission from the base station to at least some of the clients of said set, and preferably to all the clients of said set (i.e. the clients storing an application module of a given function and capable of communicating with the base station).

[0024] Thus, the use of a connection duration prediction model for each client of said set allows selection of a subset suitable for telecommunications networks in which clients are dynamic. The probability that the selected clients can participate in the entire federated learning cycle (effective participation rate) is thus considerably increased compared to solutions in the prior art, which do not take into account the dynamic aspects related to clients.

[0025] According to embodiments, the method may comprise a phase of training the connection duration prediction model, the method comprising the following steps during the training phase: - obtaining, for several clients, at least one piece of information among speed and position; - determination, for said clients, of respective connection durations to the base station, based on said information obtained; - training the connection duration prediction model from the speed and / or the position, by supervised learning from a training database comprising associations between the speed and / or the position of each of said several clients on the one hand, and the connection duration of said client on the other hand.

[0026] Thus, the connection duration prediction model is trained by machine learning, which considerably improves the accuracy of the prediction model, given the complex and non-deterministic nature of the telecommunications network and the clients connecting to it.

[0027] In addition, the learning phase can be repeated at a given frequency.

[0028] It is thus made possible to regularly update the prediction model, by example every hour.

[0029] According to embodiments, the connection duration prediction model can be obtained by linear regression from the training data.

[0030] Thus, a simple model can be used for the construction of the prediction model, which is compatible with moderate computing powers.

[0031] Alternatively, the connection duration prediction model may be obtained by linear kriging interpolation from the training data.

[0032] Such a prediction model is simple and, in practice, capable of implementing connection duration predictions with high accuracy.

[0033] Alternatively, the connection duration prediction model may be a regression tree trained from the training data.

[0034] Such an embodiment facilitates the prediction of a connection duration in the current phase.

[0035] According to embodiments, during the current phase, the remaining connection duration of a client can be determined from the estimated connection duration of the client and from the time of initialization of the federated learning cycle and the estimation of the duration of participation necessary in the federated learning cycle for a given client of said set can be a sum: - an estimated duration of downlink transmission of the initial parameter values ​​from the base station to the given client, - an estimated local training time, necessary for the given client to train the application module of a given function contained in said given client, from local data stored by the given client, in order to obtain optimized parameter values; and - an estimated uplink transmission time of the optimized parameter values ​​of the given client to the base station. Thus, the accuracy associated with estimating the duration of participation required in the learning cycle is improved.

[0036] In addition, the estimated uplink transmission duration and the estimated downlink transmission duration can be obtained from information relating to a radio link quality between the base station and the given client, the information relating to the quality of the radio link being obtained when the given client connects to the base station.

[0037] Thus, the precision associated with the estimation of the duration of participation necessary for the learning cycle is improved, compared to an embodiment considering fixed values ​​of uplink and downlink transmission duration.

[0038] According to embodiments, the subset of clients may comprise a number of clients less than or equal to a maximum number Ksei>max, and, if the number of clients of said set whose remaining connection duration is greater than the estimate of the participation duration required for the federated learning cycle, is greater than the maximum number Ksei>max, then the selection of the subset of clients may comprise the selection of Ksei>max clients from among the clients of said set whose remaining connection duration is greater than the estimate of the participation duration required for the federated learning cycle.

[0039] Thus, the size of the subset can be limited, which makes it possible to limit the complexity associated with federated learning and to limit exchanges in the telecommunications network.

[0040] In addition, the selection of Ksei>max clients from among the clients of said set whose remaining connection duration is greater than the estimate of the duration of participation necessary for the federated learning cycle, can be random.

[0041] Thus, the selection of the subset of customers is simplified.

[0042] Alternatively, the Ksei>max clients having the longest remaining connection durations among the clients of said set whose remaining connection duration is su greater than the estimate of the duration of participation required in the federated learning cycle, can be selected from the subset.

[0043] Thus, the rate of effective participation in the federated learning cycle is improved.

[0044] According to embodiments, the method may further comprise, during the phase current, upon connection of a client to the base station, adding client information to a current database in association with the predicted connection duration for said client, the current database keeping said set of clients up to date.

[0045] Thus, it is possible to keep said set of clients up to date in real time, which ensures, at the start of the federated learning cycle, that the selection of the subset is carried out on the basis of a set which is up to date.

[0046] According to a hardware aspect, the invention relates to an aggregation device associated with at least one base station of a telecommunications network, said base station being associated with a radio coverage area and being able to communicate with client devices, called clients, the aggregation device comprising a processor configured to: - upon connection of a client to said at least one base station, obtaining at least one piece of information from among the position and the speed of the client, estimating, as a function of said at least one piece of information, a connection duration of said client to the base station, by applying a connection duration prediction model, and adding the client to a set of clients comprising clients capable of communicating, at a given instant, with the base station, the clients of said set comprising respective application modules of a given function; - upon initialization of a federated learning cycle, selecting a subset of clients from among the clients of said set, such that for any client of the subset, a remaining connection duration is greater than an estimate of a participation duration required for the federated learning cycle; - transmitting, by the base station, initial parameter values ​​to the clients of the subset, for configuration of the application modules and for training the application modules from local data of the clients of the subset; - obtaining, by the base station, optimized parameter values ​​from at least some of the clients of the subset; - determine global parameter values, based on the optimized parameter values ​​received; - transmitting, by the base station, global parameter values ​​to at least some of the clients of said set, for configuration of the application modules according to the global parameter values.

[0047] According to another material aspect, the invention also relates to a computer program capable of being implemented on an aggregation device, the program comprising code instructions which, when the program is executed by a processor, carry out the steps of the defined process.

[0048] Such programs may use any programming language. They may be downloaded from a communications network and / or recorded on a computer-readable medium.

[0049] According to another material aspect, the invention relates to a data medium on which at least one series of program code instructions has been stored for the execution of the method defined above.

[0050] According to another material aspect, the invention also relates to a base station of a telecommunications network, comprising the aggregation device defined above. Brief description of the drawings

[0051] The invention will be better understood on reading the following description, given by way of example and with reference to the appended drawings in which:

[0052] [Fig.l] illustrates an example of a telecommunications network according to embodiments of the invention;

[0053] [Fig.2] illustrates the steps of a learning phase of an hourly duration prediction model of a method for managing federated learning, according to embodiments of the invention;

[0054] [Fig.3] illustrates the steps of a current phase of a method for managing federated learning, according to embodiments of the invention;

[0055] [Fig.4] illustrates a device for aggregating a telecommunications network according to embodiments of the invention. Description of the embodiments

[0056] [Fig.l] illustrates an example of an environment for implementing the invention according to embodiments.

[0057] A telecommunications network 100 according to the invention comprises at least a first aggregator 110.1 and a first set of connected entities 111.1 to the telecommunications network 100, hereinafter called the first set of clients 111.1.

[0058] The telecommunications network 100 may be a cellular type mobile network, comprising a plurality of base stations 101.1-101.2, each base station being capable of covering a respective radio coverage area 102.1-102.2, called radio area in the following, i.e. of communicating bidirectionally with clients located in the radio area that it covers. For example, the telecommunications network 100 may be a 3G, 4G, 5G type network or any other cellular network, in particular of later generation.

[0059] Each base station 101.1 and 101.2 is associated with at least one radio cell covered in the radio area of ​​the base station. For example, the first base station 101.1 may be associated with at least one radio cell, or even several radio cells (a 2G cell, a 3G cell and a 4G cell for example).

[0060] In the following, it is considered, for illustrative purposes only, that each base station 101.1-101.2 is associated with only one radio cell.

[0061] The first aggregator 110.1 may be associated with the first set of clients 111.1, i.e. it is capable of receiving local data from the clients 111.1 of the first set. For this purpose, the aggregator 110.1 may be integrated into the base station 101.1 or may be capable of communicating, remotely or locally, with the base station 101.1.

[0062] Aggregator, or aggregation device, is understood to mean a central unit that plays an orchestrator role in federated learning, by coordinating learning between the clients of the set of clients associated with it. In the example considered here, the first aggregator 110.1 is associated with the first radio cell served by the first base station 101.1 in the first radio zone 102.1. In a variant not described, the first aggregator 110.1 may be associated with several first radio cells served by the first base station 101.1 in the first radio zone 102.1. In another variant not described, the first aggregator 110.1 may be associated with several radio cells served by several base stations in several radio zones.

[0063] In the context of the invention, the first aggregator 110.1 is responsible for selecting a subset of clients from the first set of clients, so that the selected subset of clients participates in the federated learning.

[0064] The first set of clients comprises all the clients connected to the first base station at a given time and having an application module of a given function, capable of implementing, in software or hardware, a local model for implementing the function. Thus, the first set varies dynamically depending on whether clients join or leave the first radio zone 102.1.

[0065] Each local model is defined by a set of parameters. Setting the parameter values ​​in each client thus corresponds to the configuration of the application module.

[0066] As detailed in the following, once the subset is selected, the first aggregator sends, to each client of the selected subset, initial values ​​of parameters to be used to configure the local model implemented by the application module of each client. Each client optimizes the parameters obtained from the data stored locally in the client, the locally stored data being used as training data to train the local model.

[0067] Training the local model thus leads, in a known manner, to obtaining optimized parameter values, which may differ from the initial values. Optimized parameter values ​​are parameter values ​​that result in improved performance of the local model compared to the initial parameter values. For example, when the local model is a local predictive model, an average error associated with the prediction is lower for optimized parameter values ​​than for initial parameter values.

[0068] No restrictions are attached to the parameters considered. When the local models are neural networks, the parameters may be weighting coefficients associated with respective neurons of the neural network. The local models may be different models of neural networks, and include any model capable of obtaining output data from input data, from parameters of the local model, the local model being configured by assigning given values ​​to the parameters.

[0069] No restriction is attached to the initial values ​​of the parameters initially transmitted by the first aggregator 110.1 to the selected subset of clients. These may be the current values ​​of the global model, obtained during a previous cycle of federated learning. Alternatively, these may be the parameter values ​​of a local model of a given client, previously sent to the aggregator 110.1, or even the values ​​transferred from a model pre-trained on a large set of data (also called “transfer learning”).

[0070] Upon receiving the optimized parameter values ​​from the selected subset of clients, the first aggregator 110.1 may determine global parameter values ​​from the received optimized parameter values. Thus, federated learning enables optimization of the parameter values ​​locally, and then centralized determination of the global parameter values ​​from the locally determined optimized parameter values.

[0071] The term “global value of a parameter” means a parameter value determined from several values ​​obtained locally for this same parameter (the aforementioned optimized values).

[0072] No restriction is attached to the formula applied to the optimized values ​​of a parameter to obtain the global value of this parameter. For example, the global value of a parameter can be obtained by averaging the locally optimized values ​​obtained for this parameter, or by applying a predetermined weighting to the optimized values ​​of this same parameter.

[0073] For example, for a given parameter, the first aggregator 110.1 receives several optimized values, which may be different, from the clients of the selected subset (one optimized value per client, for this given parameter).

[0074] The global parameter values ​​thus determined can be transmitted to the first subset of 111.1 clients, for updating local models based on global parameter values. We can then understand why such learning is called “federated”.

[0075] In the example shown in [Fig.l] for illustrative purposes, the network 100 further comprises a second aggregator 101.2 associated with a second set of clients 111.2 capable of communicating, in a second radio coverage area 102.2, bidirectionally with a second base station 101.2.

[0076] No restriction is attached to clients 111.1 and 111.2 which may be mobile terminals of the Smartphone type, mobile connected objects such as connected drones, or even connected vehicles.

[0077] Generally, a client designates any entity capable of communicating with the network 100, and of obtaining and storing data locally. In order to obtain data, each client may comprise at least one sensor and / or at least one user interface capable of receiving user inputs.

[0078] For example, a smartphone-type client may obtain data from sensors such as a camera, a microphone, a satellite positioning module, or GPS for "Global Positioning System" in English, or from a touch interface capable of receiving user input. Data relating to user behavior may also come from applications installed on the smartphone.

[0079] An automotive type client may obtain data from vehicle sensors, such as radar, lidar, GPS module, camera, etc.

[0080] Thus, no restrictions are attached to the local data obtained and stored by each client: this may be textual content, data organized in the form of spreadsheets, photos, videos, audio content, data illustrating the behavior of a user, data indicating the quality of the radio network, etc.

[0081] The data of each client 111.1 may be stored in a local memory 112.1.

[0082] As previously described, each client 111.1 may further comprise a local model implemented by an application, hardware or software module of the client 111.1 and which may be trained by the client 111.1 using local resources 113.1, the local model being initially defined from the initial parameters and then trained by receiving as input the locally stored data, when the client is part of the subset selected by the aggregator 110.1, which makes it possible to determine optimized parameter values.

[0083] The client 111.1 may further apply the local model, once configured from the global parameter values ​​received from the aggregator 101.1 (whether the client is part of the selected subset or not), on data acquired in real time by the client 111.1 (outside the federated learning cycle) by means of its application module.

[0084] No restriction is attached to the given function performed by the local model implemented by the client application module 111.1.

[0085] For example, when the clients 111.1 are motor vehicles, the local model may be dedicated to a collision prediction function, driver intention prediction, driving assistance, or image recognition. When the clients 111.1 are smartphones or vehicles, the local model of each client 111.1 may be dedicated to an application optimization function or an infotainment system in order to improve the user experience. As a further variant, the local model of each client 111.1 may be dedicated to a function of predicting a quality of service of the telecommunications network 100 and / or to a function of predicting the quality of the radio signal.

[0086] Thus, no restriction is attached to the function of the local models whose parameter values ​​are optimized by federated learning, according to the invention.

[0087] In the example of [Fig.l], each aggregator is associated with a single base station (corresponding to a single cell) and each base station is associated with a single aggregator. However, as a variant according to the invention: - an aggregator may be associated with two or more base stations, in which case the set of clients associated with the aggregator is the set of clients capable of communicating with at least one base station among the two or more base stations; - a base station may be associated with two or more aggregators, in which case the set of clients associated with each aggregator is only a portion of the clients capable of communicating with the base station.

[0088] As mentioned above, each base station is associated with one or more cells in the radio coverage area it covers. Thus, each aggregator is associated with at least one radio cell of a base station, but may further be associated: - to several radio cells of the same base station; - to several radio cells of several base stations.

[0089] Additionally, in a telecommunications network 100, each radio cell may be associated with an aggregator. Alternatively, only some of the radio cells in the telecommunications network 100 may be associated with respective aggregators.

[0090] The application of the invention to the first aggregator 110.1 and to the first set of clients 111.1 is considered in the following, by way of illustration, the first set of clients 111.1 being the set of clients communicating with the first base station 101.1 and having a local model implemented by an application module and capable of performing a given function (examples of which will be given in the following).

[0091] Thus, in order to minimize exchanges in the network, and the associated costs, federated learning proposes the selection of a subset of clients, among the clients of the first set, for a learning cycle, or learning period.

[0092] As previously described, existing approaches to selecting a subset provide either a random selection, a selection based on the resources available in each client, or a selection based on the performance of the local model of each client.

[0093] However, these approaches do not take into account the dynamic nature of the clients. Indeed, in the telecommunications network 100, the clients on the move do not remain constantly connected to the same base station, and therefore to the same aggregator in the example considered previously, but change base station during the move, in order to remain connected to the network 100, via a transfer procedure or "handover" in English. When the connection duration of a client is less than a learning cycle (or less than a part of the cycle during which the participation of the client is necessary), which is frequent when the client is traveling on a highway, compared to traveling in the city center, the selection of this client according to the approaches of the prior art leads to a degradation of the performance of the federated learning cycle.

[0094] In order to overcome the drawbacks identified in existing federated learning methods, the invention provides for the selection of the subset of clients 111.1 taking into account an estimated connection duration for each client of a first set of clients, the first set of clients comprising all the clients connected to the first base station at a given time, and each storing a local model, in the example considered below.

[0095] In particular, clients having an estimated connection duration that is greater than a learning cycle duration are selected from the first set.

[0096] Such an estimation, or prediction, of the connection duration can be implemented by each aggregator from a connection duration prediction model, obtained in accordance with the description in [Fig.2].

[0097] [Fig.2] presents the steps of a phase 20 of learning a connection duration prediction model, of a method according to embodiments of the invention.

[0098] As detailed previously, phase 20 is described in the context where the first aggregator 110.1 is implemented in the first base station 101.1 in communication with the first set of clients 111.1 at a given time.

[0099] The learning phase 20 comprises a step 200, during which a client 111.1 initiates a connection to the first base station 101.1. No restrictions are attached to such a connection initiation step, which may be according to a procedure dependent on the telecommunications network 100, for example described in a telecommunications standard.

[0100] Following initiation of the connection, the client 111.1 requests a connection to the base station 101.1. During such a request, the first base station 101.1 may receive, at a step 201, one, several, or all of the following information: - a location of the customer 111.1; - a client speed of 111.1; - information on the resources 113.1 of the client 111.1; and / or - information relating to radio quality between the 111.1 client and the 101.1 base station.

[0101] This information is thus available to the first aggregator 110.1 which is integrated into the first base station 101.1 in the example considered.

[0102] In a step 202, upon receipt of the aforementioned information, or of some of the aforementioned information, the first aggregator 110.1 updates a training database with the received information, and the client 111.1 is thus added to the training database in association with the received information. The training database may be internal to the first aggregator 110.1 or may be accessible to the first aggregator 110.1. Such an update in step 202 may in particular comprise the creation of an entry in the database, the entry comprising: - a customer ID 111.1; - a connection time Te, corresponding to the time of receipt of a connection request by the base station 101.1 by the client 111.1; - the aforementioned information, or some of the aforementioned information, such as the location of the 111.1 customer and / or the speed of the 111.1 customer.

[0103] During a step 203, the client 111.1 is moving in the first radio zone 102.1 covered by the first base station 101.1. The client 111.1 being connected to the base station, it can use the telecommunications services to exchange voice data and / or files with other clients of the network 100.

[0104] In a step 204, the client 111.1 initiates the termination of the connection with the first base station 101.1, because the client 111.1 leaves the first radio zone 102.1 covered by the first base station 101.1. The initiation of the termination can be implemented according to a “handover” procedure so as to ensure the continuity of the connection to the network 100, the client 111.1 initiating a connection on a base station of the network other than the first base station 101.1, for example the second base station 101.2 described previously. Alternatively, the client 111.1 leaves the radio coverage zone 102.1 in step 204.

[0105] The first base station 101.1, and consequently the first aggregator 110.1, is informed of the initiation of the end of connection of the client 111.1, at a step 205, corresponding to a termination time Ts.

[0106] At a step 206 following step 205, the first aggregator 110.1 updates the training database by adding the termination time Ts to the entry created for the client 111.1 at step 202.

[0107] Thus, the entry identifies a connection duration to the first base station 101.1 of the client 111.1, corresponding to the difference Ts-Te.

[0108] The aforementioned steps may be repeated in parallel, or sequentially, for all clients 111.1 connecting to the first base station 101.1 during a collection period P, referenced 210, or for some of these clients 111.1, entries being respectively created in the training database. The collection period may be a period of one hour for example, or of several hours.

[0109] At a step 211 following the collection period P 210, the training database thus comprises entries for a plurality of clients 111.1, each entry identifying a connection duration Ts-Te of the client 111.1 to which it corresponds, as well as additional information such as the position of the client 111.1 during the initialization of the connection, its speed, its resources, as described previously.

[0110] The training database can then be used to implement automatic learning at a step 212, which is a supervised learning in particular, of a model for predicting a connection duration. The prediction model thus trained is capable of predicting a connection duration from the information received for a client 111.1 during its connection to the first base station 101.1, namely its position and / or its speed (preferably these two pieces of information).

[0111] The supervised learning of step 212 may consist of: - receive an entry from the training database, the entry corresponding to a client 111.1; - estimate a connection duration D of the client 111.1 from the position and / or speed of the client 111.1 contained in the entry; - compare the estimated, or predicted, connection duration D with the difference Ts-Te; - modify the model so as to minimize a gap between D and Ts-Te; - start again with a next entry (not yet used for training the connection duration prediction model) from the training database.

[0112] Supervised learning can thus be implemented until the connection duration prediction model converges and / or until the training database entries are exhausted.

[0113] The connection duration prediction model can then be exploited by the first aggregator 110.1 for predicting the connection duration of the clients 111.1 of the first set of clients connected to the first base station during a current phase of the method according to the invention, when selecting a subset of clients 111.1 for implementing the federated learning leading to obtaining the global values ​​of the parameters. The steps describing the federated learning are described with reference to [Fig.3].

[0114] It should be noted that the local models of the clients 111.1 previously described are distinct from the connection duration prediction model from phase 20. The connection duration prediction model is used to improve the federated learning leading to the optimization of the local models of the clients 111.1 of the first set.

[0115] In the following, detailed examples corresponding to the implementation of phase 20 are described, for illustrative purposes.

[0116] The training database is denoted py V T- j\}. with: - N; the number of 111.1 customers who connected to the cell during the P210 collection period; ' (.X- y. j • 'a Position of client i, each client is then referenced by an index i varying between 1 and N, when establishing the connection with the first base station 101.1; - v>: the speed of the client 111.1 of index i when establishing the connection with the first base station 101.1; - T': the connection duration of the client 111.1 of index i to the first base station 101.1. This connection duration, observed, corresponds to the difference between the termination time Ts, of the disconnection of the client 111.1 of index i from the first radio cell and the connection time Te, corresponding to the establishment of the connection between the client 111.1 of index i and the first base station 101.1 '■ r= (Ts-Te).

[0117] Once constituted, the training database B is used by the first base station 101.1, or by the first aggregator 110.1, to train the aforementioned connection duration prediction model. Thus, the first aggregator 110.1 has a prediction model making it possible to estimate the connection duration of any new client 111.1 which would connect to the first base station 101.1, the prediction model taking as input the position of the new client and its speed, when connecting to the first base station 101.1.

[0118] The training database B may be separated into training data and test data. For example, 80% of the training database may be dedicated to training data, with the remaining 20% ​​dedicated to test data. The test data may be randomly selected from the database training.

[0119] Three examples of algorithms that can be used to construct the connection duration prediction model are given below for illustrative purposes. Other algorithms, not described, such as neural networks, can however be used for constructing the connection duration prediction model.

[0120] A first example is a prediction model based on a linear regression. According to such a prediction model, Tt is estimated by a linear combination of xà -v; and vî, for each customer of index i, between 1 and N, having an entry in the training database B. More precisely the reals a, b etc are sought during the supervised learning, so as to minimize the mean square error, EQM, on the test data, between the Ti and the aX; + by + cv^: [Om] eqm = Jr(aXf + by. + ) ]

[0122] Alternatively, the mean absolute error may be used on the test data.

[0123] The regression model used in this first example can be derived from the Python™ scikit-leam library. Other languages ​​such as R can be used.

[0124] A second example is a prediction model based on a linear spatial interpolation method called kriging. The connection duration T of a client 111.1 is estimated by the estimation t from at least some of the connection durations of the clients of index i of the training database B, in the following manner:

[0125] = V.

[0126] During training by supervised learning, the coefficients are calculated in such a way as to minimize the variance of the estimation error between the actual connection duration T and its estimate r; = y^f Çt-t)-

[0127] Thus, a connection duration is estimated for each association between a position in the cell and a speed. When a new client 111.1 connects, the estimated connection duration is that which is associated by the kriging method with the position of the new client in the cell and its speed.

[0128] The kriging model used in this first example can be derived from the pykrige library of Python™. Other languages ​​such as the R language can be used.

[0129] A third example is a prediction model based on a regression tree. The space of the input variables (position and speed of the client connecting to the first base station), hereinafter “the space”, is divided into several regions, each region being assigned a connection duration prediction value. The space is divided into different regions recursively. In a first iteration, the space is divided into two regions, and a connection duration prediction value is assigned to the region, so as to minimize a prediction error in each region. The prediction error is a function of the differences between the actual connection durations of the clients in the region (the actual connection durations being taken from the training database B), and the prediction value assigned to the region. For example, at the end of the first iteration, a first region may correspond to speeds less than or equal to 30 km / h, while a second region corresponds to speeds greater than 30 km / h.Each of the two regions is then divided into two sub-regions, in the same way, and such an operation is repeated on each region calculated at the end of an iteration, up to a given stopping criterion, which can be a given depth of the regression tree, or a criterion related to the prediction errors.

[0130] The regression tree used in this first example can be derived from the Python™ scikit-leam library. Other languages ​​such as R can be used.

[0131] Other learning algorithms may be provided for constructing the connection duration prediction model according to the invention. Regardless of the algorithm used, the resulting prediction model is capable of predicting a connection duration from a speed and a client position 111.1 in the first radio zone 102.1.

[0132] Alternatively, the prediction model is developed only for the client's position in the cell, and does not take into account its speed. Alternatively, only the speed is taken into account to estimate the client's connection duration.

[0133] Several training phases 20 can be implemented over several distinct training periods, for example every hour, which makes it possible to regularly update the connection duration prediction model which is specific to the first radio zone 102.1. It should be noted that the first set of clients 111.1, which represents the first clients 111.1 connected to the base station at a given time, evolves dynamically. Thus, the clients added to the training database over a first collection period P are different from the clients added to the training database during a second collection period P subsequent to the first collection period P.

[0134] [Fig.3] illustrates a current phase 30 of implementation of federated learning between several first clients 111.1, of a method according to embodiments of the invention.

[0135] The current phase 30 may follow a phase 20 of learning a connection duration prediction model. In other words, at the start of the current phase 30, the 110.1 aggregator has a connection duration prediction model, constructed in the manner described previously with reference to [Fig.2].

[0136] As previously discussed, the first set of clients 111.1 varies dynamically, the first set during the current phase 30 differs from the clients added to the training database during the phase 20 described with reference to [Fig.2]. In other words, the first client 111.1 illustrated in [Fig.3] is a separate entity from the first client 111.1 illustrated in [Fig.2] (insofar as the first client in [Fig.2] has left the first radio zone 102.1).

[0137] In a step 300, a first client 111.1 initiates a connection to the first base station 101.1, similarly to step 200 previously described.

[0138] In a step 301, following the initiation of the connection in step 300, the client 111.1 requests a connection to the first base station 101.1. During such a request, the first base station 101.1 may receive, in step 301, one, several, or all of the following information: - a location of the customer 111.1; - a client speed of 111.1; - information on the resources 113.1 of the client 111.1; and / or - information relating to radio quality between the 111.1 client and the first 101.1 base station.

[0139] This information is thus available to the first aggregator 110.1 which is integrated into the first base station 101.1 in the example considered.

[0140] In a step 302, the first aggregator 110.1 applies the connection duration prediction model by taking the speed and / or the position of the client 111.1, received in step 301, in order to predict the connection duration of the client 111.1. The connection duration thus predicted is added in association with information of the client 111.1 in a current database which keeps up to date the first set of connected clients 111.1, at a given time.

[0141] Thus, the current database comprises an entry for each client 111.1 forming part of the first set at a current time, the entry comprising the aforementioned information of the client 111.1 as well as the predicted connection duration.

[0142] In a step 303, steps 300 to 302 may be repeated for other clients 111.1 that connect to the first base station 101.1 and store a local model having the given function, and thus join the first set. Entries are respectively created in the current database for each of these other clients 111.1. In the event of a disconnection of a client 111.1, the entry associated with the client 111.1 in the current database is deleted.

[0143] Thus, the current database keeps up to date the clients 111.1 belonging to the first set connected to the first base station 101.1.

[0144] At a time to, a federated learning cycle is initiated. No restriction is attached to the triggering of the federated learning cycle, which may be required by a manager of the network 100, or which may be implemented regularly at a given frequency.

[0145] The instant Lest is therefore an instant of the start of the federated learning cycle.

[0146] In a step 304, the first aggregator 110.1 obtains from the current database, the information relating to the clients 111.1 of the first set as well as the predicted connection durations associated with the clients 111.1.

[0147] In a step 305, the first aggregator 110.1 selects a subset of the clients 111.1 of the first set, according to the predicted connection durations respectively for the clients 111.1 of the first set, so that a remaining connection duration of each client 111.1 of the subset is greater than an estimate of a participation duration necessary for the federated learning cycle.

[0148] Selection step 305 is described in detail after the description of steps 306 to 312 which follow.

[0149] In a step 306, the first aggregator 110.1 transmits to all the clients 111.1 of the subset selected in step 305, initial values ​​of parameters to be used to parameterize the local model of each client 111.1 of the selected subset.

[0150] In a step 307, each client 111.1 of the selected subset, upon receipt of the initial parameter values, configures, according to an initial configuration, the local model of its application module with the initial parameter values ​​and trains the application module thus configured by automatic learning by processing the local data stored in its local memory 112.1. At the end of the training, the client 111.1 of the selected subset obtains optimized parameter values, as previously described.

[0151] In a step 308, the client 111.1 of the selected subset transmits the optimized parameter values ​​to the aggregator 110.1.

[0152] It should be noted that [Fig.3] only illustrates transmission from a single client 111.1. In practice, all clients 111.1 of the selected subset determine their own optimized parameter values ​​in step 307 and transmit the respective optimized parameter values ​​to the first aggregator 110.1 in step 308.

[0153] Upon receiving the optimized parameter values ​​from the clients 111.1 of the selected subset, the first aggregator 110.1 aggregates in step 309 the optimized parameter values ​​in order to obtain global parameter values. Each global value of a parameter can be determined by averaging or weighting the optimized values ​​of this same parameter, obtained locally by the clients 111.1 of the selected subset.

[0154] In a step 310, the global parameter values ​​are transmitted from the first aggregator 110.1 at least to the clients 111.1 of the selected subset, and preferably to all the clients 111.1 of the first set.

[0155] Thus, each client 111.1 receiving the global parameter values ​​can update (parameterize or configure) the local model implemented by its application module, from the global parameter values, at a step 311. The federated learning cycle is thus completed.

[0156] Steps 306 to 311 are consistent with the principles of federated learning, and are not further described in this description.

[0157] At a step 312, the current phase may be iterated in order to implement a new federated learning cycle. It should be noted that for a next federated learning cycle, the connection duration prediction model may have been updated compared to the previous federated learning cycle, following a new iteration of the learning phase described with reference to [Fig.2].

[0158] In the following, the selection step 305 is described in detail.

[0159] Let K £ 1; The number of clients 111.1 connected to the first base station 101.1 at time t0 and storing a local model having the given function, or the number of clients of the first set;

[0160] Let E = ...,} be the set of clients 111.1 connected to the first base station 101.1 at to and storing a local model having the given function. E therefore corresponds to the state of the current database at time t0. In other words, E is the first set of clients 111.1 at time t0.

[0161] Let G )}: For each customer Q e £, the vector T indicates the instant f(*) of establishment of the connection between ek and the first base station 101.1. For example, if customer ek is customer 111.1 shown in Figure 3, the instant (k) corresponds to the execution of step 301 requiring the connection to the first base station 101.1.

[0162] Let seljnax — be the maximum number of customers that the first aggregator 110.1 can select at each learning cycle, or the maximum size of the subset selected at the end of step 305. The maximum number of customers can be between 8 and 20, for example equal to 10.

[0163] Let Ksel = min(K, Kse / jnax), be the number of clients who are actually selected by the first aggregator 110.1 during step 305, to participate in the learning cycle.

[0164] Let SC {cq, ..., eK} be the subset of clients 111.1 selected to participate in the federated learning cycle during step 305, with |S| = Ksel.

[0165] Let Seff GS be the set of clients 111.1 which were selected by the first aggregator 110.1 to participate in the federated learning cycle and which had effective participation, i.e. from which optimized parameter values ​​were received in step 308.

[0166] Let ref f = | Sef f | / 1S | be the effective participation rate.

[0167] Let t —\\jf f(>) \\\ I • For each client Q e E of the first set, 7 do wn " t \ ' down ) " ' ]T . the Tdown vector indicates the estimated downlink transmission time required 'clown to the base station 101.1 to send the initial values ​​of the parameters to the client. This duration can be estimated by the first aggregator 110.1 when establishing the connection between the client and the base station 101.1, from the information obtained in step 301, in particular from the position of the client and / or from the quality of the radio link. This is thus an estimate of the execution duration of step 306, for each client 111.1 of the first set. Alternatively, the estimated downlink transmission durations are considered equal and constant, whichever is the higher. that is k, in which case the vector Tdown can be replaced by a single value.

[0168] Let: For each customer ek e £ of the first set, the vector Tup indicates the estimated uplink transmission duration f(k) required by the client ek to send the parameters of its local model to the first base station 101.1. This duration can be estimated when establishing the connection between the client ek and the first base station 101.1, from the information obtained in step 301, in particular from the position of the client ek and / or from the quality of the radio link. It is thus an estimate of the execution duration of step 308, for each client 111.1 of the first set. Alternatively, the estimated uplink transmission durations are considered equal and constant, regardless of k, in which case the vector Tup can be replaced by a single value. In this variant, Tup can be equal to Tdown. For example, Tup and Tdown can be of the order of one second, or even equal to 1 second.

[0169] Let j\\]. For each customer e E of the first set, the vector indicates the estimated local training time fW required by the client ek 1 train to train its local model. This duration can be estimated from the information obtained in step 301, in particular the information relating to the resources 113.1 indicated by each client 111.1 during step 301 of connection to the base station 101.1. In particular, the greater the resources of the client ek, the shorter the local training duration / £). train

[0170] Let D = \{(Tj) W}._}: For each GE client of the first set, the vector D indicates the connection duration Tk, predicted by application of the prediction model by the first aggregator 110.1 during step 302.

[0171] From the above notations, the selection of the subset by the first aggregator 110.1 during step 305 can be implemented by the algorithm corresponding to the following sub-steps 1 to 5:

[0172] 1- Let Sq be the set which will contain all the customers of the first set (or of a type given in the first set) whose remaining connection duration is sufficient to participate in the learning cycle, in particular to reach step 308 of transmission of the optimized values ​​of the parameters to the first aggregator 110.1.

[0173] 2- SQ = 0 S{} is initialized to the empty set;

[0174] 3- The vectors T, T lip, Tdmvn and D are obtained;

[0175] 4- for each client e E, check if the remaining connection time Tk- ) is greater than an estimate of the duration of the learning cycle until the optimized parameter values ​​are obtained by the first aggregator 110.1 (called “estimation of a necessary duration of participation of the client in the federated learning cycle”), i.e. tA*) , yA^) . yA^) : J train 'J up

[0176] SiyAL) T(k) < ((k)\ then: down train UP — ' k \ mm /

[0177] Add to: So = 5'0 U {ej

[0178] Otherwise ek is excluded from the selection.

[0179] 5- If So includes more than Ksei clients 111.1, determine the subset 5, 5 G. clients that will participate in federated learning from the set So. Clients can be selected from Sq randomly or with priority to clients whose ) ) are the highest, which allows prioritizing clients that have the highest probabilities of participating in the entire federated learning cycle, which therefore allows maximizing the effective participation rate reff.

[0180] Otherwise S=S0.

[0181] S thus corresponds to the subset selected during step 305 previously described.

[0182] It should be noted that, for federated learning of a driving assistance model, the first set includes clients corresponding to motor vehicles storing such a local driving assistance model and connected to the aggregator at a given time. The subset is therefore selected based on the estimated remaining connection times for vehicle-type clients only, the other clients not being part of the first set considered. Similarly, in the In the context of federated learning of a model for improving the use of an application on a mobile phone, only mobile phones having such a local improvement model are part of the first set and are considered for the selection of the subset involved in federated learning.

[0183] The client type may be known to the first aggregator from the information received upon connection in step 301.

[0184] Additionally, the clients of the selected subset may further be selected according to at least one additional criterion. For example, the resources of each client of the first set, the performance of the local model of each client of the first set and / or the amount of local data of each client of the first set may also be taken into account, in addition to the remaining connection time, when selecting the subset in step 305.

[0185] The selection of the subset according to the invention makes it possible to improve the effective participation rate reff compared to the prior art solutions described previously.

[0186] By way of example, in a case where the clients 111.1 considered in the first set are motor vehicles storing local classification models of images acquired by respective cameras of the vehicles, allowing the detection of traffic signs and / or pedestrians, the effective participation rate reff obtained is approximately 85% whereas the solutions of the prior art are of the order of 30%.

[0187] The improvement of the effective participation rate also makes it possible to improve the precision of the global model after the aggregation step 309 described previously, leading to obtaining the global values ​​of the parameters, at the end of each federated learning cycle.

[0188] Thus, the invention advantageously makes it possible to take into account the dynamic aspect inherent in the majority of entities connected to a telecommunications network. Such consideration is made possible by predicting the connection duration of each client by means of a prediction model derived from machine learning based on the principles of artificial intelligence. Artificial intelligence makes it possible to address the complex and non-deterministic nature of the network 100 and the clients 111.1 which connect to it.

[0189] In the following, four examples of applications of the method according to the invention are described, for illustrative purposes.

[0190] In a first application example, each local model is dedicated to driving assistance, and the clients 111.1 considered in the first set are motor vehicles. In this first application example, each vehicle can maintain an up-to-date local model which allows it to predict, based on its perception of the environment around it, captured by one or more sensors, whether it can initiate a action safely such as accelerating or changing lanes. The local model can correspond to a classification model that returns a binary output, for example zero if the action can be performed safely and one otherwise. The local vehicle data that can be used is historical data stored in the local memory 112.1 of each vehicle, representing the actions performed by the vehicle during past journeys.Taking for example the case of lane change assistance, the supervised learning implemented in step 307 can be based on environmental data such as videos or images describing the state of the traffic near the vehicle and the structure of the road, data on a driving state of the vehicle, such as its acceleration and its speed, these data being labeled by data relating to the situation after the lane change, to allow supervised learning (accident or non-accident situation, risk level following the lane change for example). These latter data correspond to the variables to be predicted by the local model. When a vehicle driver wishes to initiate a lane change, the local model predicts whether or not the lane change can be carried out safely.The local model takes as input data on its environment and / or data on the vehicle's driving state and predicts whether the lane change can be made or not. Federated learning in this case consists of determining optimized parameter values ​​for each local model of each vehicle in the selected subset, then determining global parameter values ​​from the optimized parameter values, the global parameter values ​​being returned to all vehicles in the first radio zone.

[0191] In a second application example, the clients of the first set are still motor vehicles, but each local model aims at real-time traffic prediction in order to adapt a vehicle route. The vehicles selected in the subset collaborate to update their local models with the global parameter values ​​resulting from the federated learning. Thus, each vehicle is able to find the best route to its destination. The local model in each vehicle is thus a classification model which can take as input the position of the vehicle, its destination, preference parameters (desired road type for example, toll or no toll), the set of possible routes and the traffic conditions, and which predicts as output a class for each route, which represents a score of the route, the route with the highest score being selected to assist the vehicle navigation.The local model can be applied at a given frequency, in order to adapt in real time to the traffic conditions which evolve along the route. In this second example, and as described with reference to the figures, each aggregator can be associated 1 to 1 with a radio zone of the network 100.

[0192] In a third example of application, the local models are able to predict the quality of the radio link with the first base station 101.1, for different positions in the first cell. In this case, any type of clients (vehicles, telephones, drones, etc.) can participate in the federated learning and can thus be part of the first set. Thus, the clients 111.1 of the first set can be the clients able to communicate with the first base station 101.1 and each of these clients 111.1 has an application making it possible to probe the quality of the radio signal received at different locations different from a current position of the client 111.1. The application makes it possible to build a local database in the local memory 112.1 of each client 11.1, with information on the geographical position of the client and the quality of the radio signal received at each geographical position.The local database is enriched as the client 111.1 moves in the first radio zone. The local model trained locally by each client, to obtain the optimized parameter values, thus consists of a model for predicting the quality of the radio signal for a given position or for a given set of positions. Such predictions can then be used by autonomous vehicles to eliminate routes whose successive positions include positions associated with a radio signal quality that is too low. It is thus possible to ensure a permanent connection to the network 100. Federated learning thus makes it possible to determine global parameter values ​​to configure the local model, obtained by clients 111.1 occupying different positions successively in the cell, which makes it possible to improve the accuracy of the local model thus parameterized in each client 111.1.

[0193] In a fourth example of application, the local model can be associated with a tourist recommendation application, installed on some of the clients 111.1 capable of communicating with the first base station 101.1. The local model can predict, in real time, to a tourist traveling in a city, the nearby places which may interest him, in particular restaurants, leisure activities or any other point of interest.

[0194] The application takes as input data from the user of the client 111.1, such as preferences or habits, the current position of the client and the places he has already visited, data relating to the city (calendar of events, places of interest). From all of this data, the application proposes the most relevant place(s) for the user of the client 111.1.

[0195] The local data used to train the local model in each client 111.1 can thus associate the application's past recommendations with user feedback, positive or not, relating to these recommendations. The recommendations can thus be continuously improved by federated learning between the different customers with the application in a given radio zone.

[0196] [Fig.4] shows the structure of the first aggregator 110.1, according to embodiments of the invention.

[0197] The first aggregator 110.1 comprises a processor 401 configured to communicate unidirectionally or bidirectionally, via one or more buses or via a direct wired connection, with a memory 402 such as a “Random Access Memory” type memory, RAM, or a “Read Only Memory” type memory, ROM, or any other type of memory (Flash, EEPROM, etc.). Alternatively, the memory 402 comprises several memories of the aforementioned types.

[0198] The memory 402 comprises at least one non-volatile memory in which the data used and / or resulting from the implementation of steps 201, 202, 205, 206, 210 to 213 of the phase 20 of learning the prediction model, as well as the implementation of steps 301 to 305, 306, 308 to 310 and 312 of the current phase 30 of federated learning described with reference to [Fig. 3] are stored, temporarily or permanently.

[0199] In particular, the memory 402 can store the connection duration prediction model obtained at the end of step 212, the training database updated at the end of step 206, the current database storing the entities of the first set at each instant, in association with the connection durations predicted at step 302. Alternatively, the prediction model, the training database and / or the current database may be external but accessible to the first aggregator 110.1.

[0200] The processor 401 is able to execute instructions, stored in the memory 402, for the implementation of the steps 201, 202, 205, 206, 210 to 213 of the phase 20 of learning the prediction model described with reference to [Fig.2], as well as the steps 301 to 305, 306, 308 to 310 and 312 of the current phase 30 of federated learning described with reference to [Fig.3].

[0201] The first aggregator 110.1 comprises an interface 403 capable of communicating bidirectionally with the first base station 101.1, when the first aggregator 110.1 is distinct from the first base station 101.1. When the first aggregator 110.1 is integrated into the first base station 101.1, the interface 403 may be a radio interface making it possible to exchange bidirectionally with the clients of the first set in the first radio cell.

Claims

Claims

1. Method for managing federated learning between several client devices (111.1), called clients, capable of communicating with a base station (101.1) of a telecommunications network (100) associated with a radio coverage area (102.1), a set of clients being capable of communicating, at a given instant, with the base station, the clients of said set comprising respective application modules of a given function, the method comprising the following steps, during a current phase (30): - upon connection (300) of a client to the base station, obtaining (301) at least one piece of information from among the position and the speed of the client, estimating (302), as a function of said at least one piece of information, a connection duration of said client to the base station, by applying a connection duration prediction model, and adding said client to said set of clients; - upon initialization of a federated learning cycle, selection (305) of a subset of clients from among the clients of said set, so that for any client of the subset, a remaining connection duration is greater than an estimate of a participation duration necessary for the federated learning cycle; - transmission (306) of initial parameter values ​​to the clients of the subset, for configuration of the application modules and for training the application modules from local data of the clients of the subset; - obtaining (308) optimized parameter values ​​from at least some of the clients of the subset; - determination (309) of global parameter values, as a function of the optimized parameter values ​​received; - transmission (310) of the global parameter values ​​to at least some of the clients of said set, for configuration of the application modules from the global parameter values.

2. The method of claim 1, further comprising a training phase (20) of the connection duration prediction model, the method comprising the following steps during the training phase: - obtaining (201), for several clients, at least one piece of information among speed and position; - determining (206), for said clients, connection durations to the base station, as a function of said information obtained; - training (212) the connection duration prediction model from the speed and / or the position, by supervised learning from a training database comprising associations between the speed and / or the position of each of said several clients on the one hand, and the connection duration of said client on the other hand.

3. Method according to claim 2, in which the learning phase (20) is reiterated (213) at a given frequency.

4. Method according to one of the preceding claims, in which the connection duration prediction model is obtained by linear regression from the training data.

5. Method according to one of claims 1 to 3, in which the connection duration prediction model is obtained by linear kriging interpolation from the training data.

6. The method of one of claims 1 to 3, wherein the connection duration prediction model is a regression tree trained from the training data.

7. Method according to one of the preceding claims, in which, during the current phase (30), the remaining connection duration of a client is determined from the estimated connection duration of the client and from the initialization time of the federated learning cycle and the estimation of the participation duration necessary for the federated learning cycle for a given client (111.1) of said set is a sum of: - an estimated duration of downlink transmission of initial parameter values ​​from the base station to the given client, - an estimated duration of local training, necessary for the given client to train the application module of a given function contained in said given client, from local data stored by the given client, in order to obtain optimized parameter values; and - an estimated duration of uplink transmission of optimized parameter values ​​from the given client to the base station.

8. The method of claim 7, wherein the estimated uplink transmission duration and the estimated downlink transmission duration are obtained from information relating to a radio link quality between the base station (101.1) and the given client (111.1), the information relating to the quality of the radio link being obtained during the connection (300) of the given client to the base station.

9. Method according to one of the preceding claims, in which the subset of clients (111.1) comprises a number of clients less than or equal to a maximum number Ksei>max, in which if the number of clients of said set whose remaining connection duration is greater than the estimate of the participation duration necessary for the federated learning cycle, is greater than the maximum number Ksei>max, then the selection (305) of the subset of clients comprises the selection of K sei.max clients from among the clients of said set whose remaining connection duration is greater than the estimate of the participation duration necessary for the federated learning cycle.

10. Method according to claim 9, in which the selection (305) of K Sei,max clients (111.1) among the clients of said set whose remaining connection duration is greater than the estimate of the participation duration necessary for the federated learning cycle, is random.

11. The method of claim 9, wherein the Ksei>max clients (111.1) having the longest remaining connection durations among the clients of said set whose remaining connection duration is greater than the estimate of the participation duration required for the federated learning cycle, are selected (305) from the subset.

12. Method according to one of the preceding claims, further comprising, during the current phase (30), upon connection (300) of a client (111.1) to the base station (101.1), adding information of the client to a current database in association with the predicted connection duration for said client, the current database keeping said set of clients up to date.

13. Aggregation device (110.1) associated with at least one base station (101.1) of a telecommunications network (100), said base station being associated with a radio coverage area (102.1) and being able to communicate with client devices (111.1), called clients, the aggregation device comprising a processor (401) configured to: - upon connection of a client to said at least one base station, obtain at least one piece of information from among the position and the speed of the client, estimate, as a function of said at least one piece of information, a connection duration of said client to the base station, by applying a connection duration prediction model, and add the client to a set of clients comprising, at a given instant, clients able to communicate with the base station, the clients of said set comprising respective application modules of a given function; - upon initialization of a federated learning cycle, selecting a subset of clients from among the clients of said set, such that for any client of the subset, a remaining connection duration is greater than an estimate of a participation duration required for the federated learning cycle; - transmitting, by the base station, initial parameter values ​​to the clients of the subset, for configuration of the application modules, and for training the application modules from local data of the clients of the subset; - obtaining, by the base station, optimized parameter values ​​from at least some of the clients of the subset; - determine global parameter values, based on the optimized parameter values ​​received; - transmitting, by the base station, global parameter values ​​to at least some of the clients of said set, for configuration of the application modules according to the global parameter values.

14. A computer program implementable in an aggregation device (110.1) as defined in claim 13, the program comprising code instructions which, when executed by a processor (401), performs the steps of the method defined in one of claims 1 to 12.

15. Base station (101.1) of a telecommunications network, comprising the aggregation device (110.1) according to claim 13.

Images