Content for  TS 23.482  Word version:  19.3.0

The AIMLE Server sends ML model training capability evaluation request to the FL members (AIMLE Clients which are deployed on UEs). The request message includes information as described in Table 8.19.3.1-1.

The FL members evaluate their capability and availability to join the FL training process. The FL members (AIMLE Clients which are deployed on UEs) run the test task contained in the request in step 1 and determine if it can join the FL process. For VFL, as part of the test task, data alignment between the datasets of the different domains are determined. The VAL server may also provide data labels for the data alignment.

The FL members (AIMLE Clients which are deployed on UEs) send ML model training capability evaluation response to the AIMLE Server. The response message contains the information as described in Table 8.19.3.2-1.

The VAL server sends AIML service operations control and management request to the AIML Enablement server as per the Table 8.20.3.1-1.

The AIMLE server determines the required service operation mode to manage the AIML service operation lifecycle based on the AIML service operation mode, AIML service operation information. For example, the AIMLE server can perform client discovery/selection using the procedure defined in clause 9.3 and model training using the procedure defined in clause 8.3. Based on the AIML client identifier the AIMLE server sends the AIML Enablement client service operation request as per Table 8.20.3.3-1. The AIML service operation mode includes start and stop operation. Start indicates the initiation of the AIML service and stop indicates termination of the AIML service. The AIML Enablement client receiving the AIML service operation mode performs the service operation mode for the AIML service operation. The AIMLE client can configure and monitor the AIML service operation as per the AIML service operation mode configuration. Based on the AIML service operation mode status reporting configuration (periodic or event-triggered), the AIML client reports the service operation mode status to the AIML Enablement server. The AIML Enablement client sends a response indicating the success or failure of the AIML Enablement client service operation as per Table 8.20.3.4-1.

The AIML Enablement server provides the AIML service operations control and management response to the VAL server. The message includes AIML service operation ID and the reporting status of the AIML service operation.

The AIMLE Consumer sends an ML model update request to the AIMLE Server that includes the ID of the model and performance degradation information.

Based on the performance degradation information, the AIMLE Server determines whether to update the ML model. If the AIMLE Server does not update the model, steps 3 and 4 are skipped.

The AIMLE Server retrieves the ML model information from the ML Repository as described in clause 8.11.3. The AIMLE Server may also perform ML model discovery to determine whether an existing ML model stored by the ML Repository can be used to replace the degraded model or train the new model (e.g., using Transfer Learning). If an existing model can be used to replace the degraded model, step 4 is skipped, and the identified model is provided in step 5. The AIMLE Server can also discover models that are related due to Transfer Learning or the use of the same training data, to identify additional models that may require an update.

The AIMLE Server performs ML model re-training, which corresponds to the ML model training procedure as described in clause 8.3. The updated model is stored in the ML repository once re-training is complete. If the degraded model is linked to other models (e.g., due to Transfer Learning, or the same training data has been used), the AIMLE Server may trigger the re-training and update of those related models.

The AIMLE Server provides the updated ML model to the AIMLE Consumer either by sending it directly, or by providing endpoint information to retrieve it from the ML Repository.

The VAL server sends an ML model performance monitoring subscription request to AIMLE server, requesting to assist in monitoring the ML model performance. This request consists of ML model information and optionally the AIML service information, etc.

The AIMLE server checks whether the VAL server is authorized to perform the ML model performance monitoring request.

If the VAL server is authorized, AIMLE server returns the success response, otherwise a failure response indication the reason for failure.

The AIMLE server identifies the AIMLE services which are utilizing the requested ML model. Such AIMLE services can be an ML model training service or an HFL service at the AIMLE server or clients. The identification of the AIMLE services may be performed via fetching ML model information from the ML repository using ML model management procedure as in clause 8.11.3. Then, the AIMLE server then starts monitoring the AIMLE service performance (e.g. accuracy, a KPI or QoS metric related to the AIML operation). This step includes receiving information from one or more AIMLE clients performing an operation based on the target ML model (e.g., based on step 7 of procedure in clause 8.12.2 or step 6 of procedure in clause 8.15.2), with an expected or experienced deviation of the required performance of the AIMLE service.

The AIMLE server detects an expected ML model degradation (e.g., model drift, data drift) based on the deviation of the performance of the AIMLE service as indicated in step 4.

The AIMLE server based on the expected ML model degradation, it may also indicate and execute a trigger action to ensure meeting the AIMLE service requirement. Such trigger action may be either an adaptation of the AIMLE service, such as training of a new ML model for the AIMLE by the same or a different AIMLE client, or re-training of the ML model by the same or different AIMLE client; or termination of the AIMLE service and initiating a new AIMLE service with a new ML model.

The AIMLE server notifies the VAL server on the expected ML model degradation and if requested the triggered adaptation of the AIMLE service.