Content for  TS 23.482  Word version:  19.3.0

An AIMLE service consumer (e.g., VAL server) makes a request for AIMLE data management assistance. The request includes information as described in Table 8.15.3.1-1.

The AIMLE server authenticates and authorizes the request. If authorized, the AMILE server assigns an identifier for the subscription.

The AIMLE server sends an AIMLE data management assistance subscription response to the AIMLE service consumer. The response includes information as described in Table 8.15.3.2-1.

The AIMLE server sends Client data processing trigger requests to AIMLE clients. The request can be for data preparation or data analysis. The request includes information as described in Table 8.15.3.4-1.

Each AIMLE client sends the requirements to trigger data management for the VAL client to perform the requested data operation. The VAL client performs the operation locally.

After the data operation completes, each AIMLE client sends a response to the AIMLE server with information as described in Table 8.15.3.5-1.

The AIMLE server aggregates the output of each AIMLE client. Aggregation includes combining or performing a statistical operation on the received data. Data received from the AIMLE clients can be numerical or categorical, which allows the AIMLE server to aggregate the output data from the AIMLE clients. If necessary, steps 4-7 is repeated to complete all the required data operations.

The AIMLE server sends an AIMLE data management assistance notification to the AIMLE service consumer with information as described in Table 8.15.3.3-1.

The VAL server sends a Transfer Learning model selection assistance request message to the AIMLE server to provide support for discovering and selecting the appropriate pre-trained model for a given ML task (for ADAE analytics ID or for a certain ML model ID).

The AIMLE Server discovers the possible entities which can provide a pre-trained model for this request. Such entities can be VAL servers or other ADAES or other AIMLE servers.

The AIMLE Server requests one or more pre-trained ML models which can be used for transfer learning for the target ML task. The ML repository identifies the base ML model as a pre-trained ML model that can be mapped to the target ML task and sends information on the models to AIMLE server which are candidate as pre-trained models available for the target task. This step reuses the ML model information discovery procedure as in clause 8.11.3.

The AIMLE Server evaluates with the support of the ML model repository, whether the pre-trained models are applicable to the ML task (ADAE analytics ID or model ID). This can be assisted using historical data or ML model rating based on previous utilization of these models for the certain ML task. Based on the evaluation (which can be based on the rating), the AIMLE Server determines one or more pre-trained models to be used for the ML task.

The AIMLE Server sends to the VAL server a transfer learning selection assistance response to the VAL server, including the information for the pre-trained models (e.g., model ID, profile) which are identified for the ML task. Also, this may include the rating/weight for the pre-trained model if the VAL server needs to select among a list of them.

Based on the selected pre-trained model information, the VAL server retrieves the selected ML model using the procedure as in clause 8.2.2.

The AIMLE client determines a requirement for using a pre-trained ML models for transfer learning for a VAL UE triggered ML task (e.g., UE analytics event).

The AIMLE client sends a UE transfer learning model selection assistance request to the AIMLE server to receive one or more pre-trained ML models which can be used for transfer learning for the target UE-triggered ML task.

The AIMLE Server requests one or more pre-trained ML models which can be used for transfer learning for the target ML task. The ML repository identifies the base ML model as a pre-trained ML model that can be mapped to the target ML task and sends information on the models to AIMLE server which are candidate as pre-trained models available for the target task.

The AIMLE Server sends a UE transfer learning model selection assistance response to the AIMLE client which includes information on the ML models which are candidate as pre-trained models available for the target UE ML task. Such models may be pre-trained for an ADAES analytics task (e.g., VAL server performance analytics) and are applicable to be used for the VAL UE analytics task (e.g. VAL session performance analytics).

The AIMLE client evaluates whether the pre-trained models are applicable to the target ML task. Based on the evaluation, the AIMLE client selects a base model to be used as pre-trained model for the ML task.

Based on the selected pre-trained model information, the AIMLE client retrieves the selected ML model using the procedure as in clause 8.2.2.

The VAL server sends an FL member grouping support request to the AIMLE Server for supporting an FL process. The initial request is to create the FL member grouping support as described in Table 8.17.3.1-1, which may be followed by other requests for querying (as described in Table 8.17.3.1-2), change (as described in Table 8.17.3.1-3), or deletion (as described in Table 8.17.3.1-4) of the FL member grouping support.

The AIMLE Server based on the request, determines the need for creating and processing a group consisting of the needed FL members for a given ML task (i.e., an ML model training/inference job ID). The need for creating an FL member group may be based on an ML task for a given AIMLE service area or for a given AIMLE service area where one or more ML tasks are expected to run, based on step 1 request.

To create, query, or change the FL member grouping support, the AIMLE Server fetches the available FL members for the given ML task (i.e., an ML model training/inference job ID) from the ML repository. Based on this information, AIMLE Server may select one or more FL members for the group for the ML task.

The AIMLE Server creates, configures, and processes the FL member group based on the available or selected FL members by the aggregator which may be the VAL Server or the AIMLE Server, generate group ID and store mapping to group member IDs. The criteria for determining the group members can be the capabilities of the FL participants, or whether the candidate participants are fixed or mobile nodes and their availabilities, the proximity of the participants among them. For the creation of the FL member group, AIMLE may utilize SEAL GMS capability for the FL member group ID generation.

The AIMLE Server interacts with the each candidate FL member from the configured FL member group as shown in step 5a.

The AIMLE Server sends indication to the candidate FL member (the AIMLE client which is deployed on UE) about the group ID and the group member identities for the ML model ID / ADAE analytics ID (based on the request in step 1.

The candidate FL member sends to the AIMLE Server a FL grouping indication acknowledge.

The AIMLE Server sends an FL member grouping support response to the VAL request indicating the group creation or providing the query response, change or deletion (based on step 1 request) and the group information.

An AIMLE server receives an ML model training request from a VAL server as described in clause 8.3.2. If ML model training request is received and the AIMLE server determines to use VFL training (VFL training between different data domains), the procedure continues to step 2.

[Optional] The AIMLE server retrieves the indicated ML model in step 1 using the ML model retrieval procedure as described in clause 8.1. If the retrieved model is already trained and meets the machine learning requirements requested by the consumer, the procedure continues to step 7.

If AIMLE client selection criteria are provided in step 1, then the AIMLE server continuously monitors and selects AIMLE clients for the VFL training as described step 4 of the clause 8.13.2. If a number of the required AIML clients is provided, the AIMLE server ensures the requirement is met when selecting AIMLE clients for VFL training. If a list of AIMLE clients is provided in step 1, the AIMLE server selects the provided AIMLE clients for the VFL training.

The AIMLE Server gets VFL members information (the VFL members information may be included in the request in step 1 (Table 8.3.3.1-1) or be obtained through step 3. The AIMLE Server interacts with the VFL members (AIMLE Clients which are deployed on UEs) for each data domain for ML model training capability evaluation as described in clause 8.19.2.

The AIMLE Server determines the VFL members for this VFL training process based on the information received in step 4 and parameters received in step 1 (Table 8.3.3.1-1). The AIMLE Server determines VFL members for each data domain. The criteria used by the AIMLE Server include:

• Available data and minimum number of data samples for the same sample among the VFL members.
• Feature alignment of the sample/datasets with data labels among the VFL members.
• Available time of the VFL members for support the VFL training operations.
• Capability and minimum number of the VFL members for the VFL training operations.
• AIML model information for the VFL members and for the AIMLE Server.

The AIMLE Server coordinates the selected VFL members for VFL training. During VFL training process, the VFL members send intermediate results to the AIMLE Server, and the AIMLE Server responds to the VFL members with the updated information (e.g. gradients). The information from the AIMLE server can be used to update the model parameters maintained at each VFL member for the different data domains.

The AIMLE server may report to the VAL server with the training status, that includes intermediate training results, The VAL server may adjust its request on the ML model training. If the VAL Server is providing data labels to complete the training, the VAL Server sends updated training parameters for the AIMLE Server to distribute to the VFL members. The updated training parameters apply for models of VFL members associated with each data domain.

The AIMLE server sends a ML model training notification to the VAL server as described in clause 8.3.2. If the training schedule is not complete (e.g., there are remaining training rounds), the AIMLE server configures the next set of training schedules and steps 3 to 6 are repeated for the next training round.