Predictive model in a system

Abstract

Embodiments of the present disclosure relate to a prediction model in a system. The present disclosure provides an apparatus comprising means for: determining a sustainability cost associated with enabling of a prediction model in a system; determining a gain associated with enabling of the prediction model in the system; and conditionally controlling the enabling of the prediction model in the system in dependence on the determined sustainability cost associated with the enabling of the prediction model in the system, and the determined gain associated with the enabling of the prediction model in the system.

Description

Technical Field

[0001] The examples disclosed herein relate to predictive models in a system. Background Technology

[0002] Predictive models can be used in a system to make future decisions based on past learning.

[0003] Training data is used to train a model to create a predictive model. The predictive model can then be deployed for use in the system. When inference data is fed to the predictive model, it produces output.

[0004] While predictive models can be useful, they can also be expensive. Summary of the Invention

[0005] According to various, but not necessarily all, examples, an apparatus is provided that includes components for: determining a sustainability cost associated with enabling a predictive model in a system; determining a gain associated with enabling a predictive model in a system; and conditionally controlling the enabling of a predictive model in a system based on the determined sustainability cost associated with enabling the predictive model in a system and the determined gain associated with enabling the predictive model in a system.

[0006] According to various, but not necessarily all, examples, a computer program is provided that, when run by one or more processors of a device, causes the device to conditionally enable a predictive model in the system based on a determined sustainability cost associated with enabling the predictive model in the system and a determined gain associated with enabling the predictive model in the system.

[0007] Based on various, but not necessarily all, examples, a method is provided that includes: determining sustainability costs associated with enabling a predictive model in a system; determining gains associated with enabling a predictive model in a system; and conditionally controlling the enabling of a predictive model in a system based on the determined sustainability costs associated with enabling a predictive model in a system and the determined gains associated with enabling a predictive model in a system.

[0008] In some, but not all, examples, sustainability costs are estimated energy costs.

[0009] In some, but not all, examples, energy costs include the energy costs of moving data within the system and the energy costs of moving the predictive model within the system.

[0010] In some, but not all, examples, the system is a cellular radio network.

[0011] In some, but not all, examples, conditional control enable includes performing one of the following with respect to the prediction model: conditionally deploying the prediction model; conditionally activating the prediction model; or conditionally deactivating the prediction model.

[0012] In some, but not all, examples, deployment conditions based on estimated sustainability costs and estimated gains associated with the use of the predictive model in the system are used to enable the deployment of the predictive model; activation conditions based on estimated sustainability costs and estimated gains associated with the use of the predictive model in the system are used to enable the activation of the deployed predictive model; and deactivation conditions based on estimated sustainability costs and estimated gains associated with the use of the predictive model in the system are used to enable the deactivation of the activated, deployed predictive model.

[0013] In some, but not all, examples, the deployment conditions, activation conditions, and deactivation conditions are different.

[0014] In some, but not all, examples, conditionally controlling the enabling of the predictive model can be reconfigured to include one of the following: conditionally deploying the predictive model; conditionally activating the predictive model; or conditionally deactivating the predictive model.

[0015] In some, but not all, examples, enabling the predictive model is conditionally controlled based on a reconfigurable comparison threshold used to compare the determined sustainability costs associated with enabling the predictive model in the system with the determined gains associated with enabling the predictive model in the system.

[0016] In some, but not all, examples, the apparatus includes components for conditionally enabling the predictive model in the system based on the integrated sustainability costs associated with enabling the predictive model in the system and the integrated gains associated with enabling the predictive model in the system, by integrating the expected user integration of the sustainability costs associated with enabling the predictive model in the system and the integrated gains associated with enabling the predictive model in the system.

[0017] In some, but not all, examples, the apparatus includes components for: obtaining combined sustainability costs, including multiplying the per-user sustainability cost associated with enabling the predictive model in the system by the expected number of users; and obtaining combined gains, including multiplying the per-user gain associated with enabling the predictive model in the system by the expected number of users; and conditionally enabling the predictive model in the system based on the combined sustainability costs associated with enabling the predictive model in the system and the combined gains associated with enabling the predictive model in the system.

[0018] In some, but not all, examples, the enabling of a predictive model in a system is conditionally controlled to be dynamic and depends on dynamically determined sustainability costs associated with the enabling of the predictive model in the system, and dynamically determined gains associated with the enabling of the predictive model in the system, wherein the enabling of the predictive model in the system responds to changes in the sustainability costs associated with the enabling of the predictive model in the system, and responds to changes in the gains associated with the enabling of the predictive model in the system.

[0019] In some, but not necessarily all, examples show that the determined gain associated with enabling the predictive model in the system is based on the usefulness of the predictive model.

[0020] In some, but not all, examples, the determined gain associated with enabling the prediction model in the system is based on the utilization level of the prediction model.

[0021] In some, but not all, examples, the system is a hierarchical distributed system comprising: a higher level of a higher rank in a hierarchy, and at least a lower level of a lower rank in a hierarchy comprising multiple domains, wherein the sustainability cost is an estimated energy cost at the higher level, at the lower level, or at both the higher and lower levels.

[0022] In some, but not necessarily all, examples, the device is configured to conditionally control the enabling of predictive models across the system, where sustainability costs are estimated energy costs for a higher level.

[0023] In some, but not necessarily all, examples, the device is configured to validate the conditional enablement of the prediction model at a lower-level domain of the system.

[0024] In some, but not necessarily all, examples, the device is configured to select a predictive model for a specific domain of the system for conditional deployment, wherein the predictive model depends on the specific domain, and wherein the sustainability cost is an estimated energy cost for the specific domain.

[0025] In some, but not all, examples, the estimated energy cost for a particular domain is based on user-related parameters specific to that domain.

[0026] In some, but not all, examples, the apparatus includes components for activating or deactivating predictive models deployed in different domains and for balancing overall performance gains with sustainability costs.

[0027] In some, but not necessarily all, examples, the device is configured to conditionally control the enabling of the predictive model at a lower-level domain of the system, where the sustainability cost is the estimated energy cost at a lower-level domain of the system.

[0028] In some, but not necessarily all, examples, the device is configured to obtain validation from organizational functions at a higher level before enabling predictive models in lower-level domains of the system.

[0029] In some, but not necessarily all, examples

[0030] Examples claimed in the appended claims are provided according to various, but not necessarily all, embodiments.

[0031] Although the examples and optional features of this disclosure have been described separately, it should be understood that their availability in all possible combinations and permutations is included within this disclosure. It should be understood that various examples of this disclosure may include any or all of the features described with respect to other examples of this disclosure, and vice versa. Furthermore, it should be understood that any one or more features in any combination may be implemented, included in, or executable by means of, apparatus, method, and / or computer program instructions as needed and suitably. Additionally, the description of the function should be considered as also disclosing any components suitable for performing that function. Attached Figure Description

[0032] Some examples will now be described with reference to the accompanying drawings, in which:

[0033] Figure 1 An example of predictive model creation and deployment is shown;

[0034] Figure 2 An example of conditional control enabling a predictive model in a system is shown, where conditional control is based on sustainability costs and gains.

[0035] Figure 3 An example of conditional control enabling a predictive model in a system is shown, where conditional control is based on energy cost and gain;

[0036] Figure 4AAn example of conditional control for the deployment of a predictive model in a system is shown, where conditional control is based on energy cost and gain;

[0037] Figure 4B An example of conditional control over the activation of a predictive model in a system is shown, where the conditional control is based on energy cost and gain;

[0038] Figure 4C An example of conditional control for deactivation of the predictive model in a system is shown, where the conditional control is based on energy cost and gain;

[0039] Figure 5 An example of a multi-level system including domains is shown;

[0040] Figure 6 An example of a multi-level system comprising domains configured as a cellular radio network is shown;

[0041] Figure 7A , Figure 7B , Figure 7C The illustration shows an example of higher-level conditional control enabling the predictive model in the system.

[0042] Figure 8A , Figure 8B The illustration shows an example of lower-level conditional control enabling the predictive model in the system.

[0043] Figure 9 The diagram illustrates a controller configured to enable conditional control for a predictive model in the system.

[0044] Figure 10 The diagram illustrates a computer program configured to enable conditional control for a predictive model in the system.

[0045] The accompanying drawings are not necessarily drawn to scale. For clarity and simplicity, some features and views in the drawings may be shown schematically or exaggerated in scale. For example, the dimensions of some elements in the drawings may be exaggerated relative to other elements to aid illustration. Similar reference numerals are used in the drawings to indicate similar features. For clarity, not all reference numerals may appear in all drawings. Detailed Implementation

[0046] The accompanying drawings disclose an example of device 2, which includes components for: determining a sustainability cost 42 associated with enabling a predictive model 24 in system 100; determining a gain 44 associated with enabling a predictive model 24 in system 100; and conditionally enabling a predictive model 24 in system 100 based on the determined sustainability cost 42 associated with enabling a predictive model 24 in system 100 and the determined gain 44 associated with enabling a predictive model 24 in system 100.

[0047] Figure 1 The diagram illustrates the system and process used to create and use the predictive model 24. The data collection module 24 provides training data 12 to the model training module 14. The model training module 14 uses the training data 12 to train the model to create the predictive model 24. The predictive model 24 can then be deployed 20 for use. The data collection module 24 provides inference data 22 to the inference module 23, which uses the deployed predictive model 24 to produce output 30.

[0048] The prediction model 24 can be, for example, a trained machine learning model or an artificial intelligence model. The prediction model 24 can be, for example, a trained artificial neural network (ANN).

[0049] Inference module 23 can be used, for example, in system 100. As will be described later, in some, but not necessarily all, examples, system 100 can be a hierarchical system (…). Figure 5 In some, but not necessarily all, examples, system 100 may be a cellular radio network. Figure 6 ).

[0050] Figure 2 An example of device 2 is illustrated, which includes components for determining the relationship between system 100 and the device. Figure 2 The system 100 determines the sustainability costs 42 associated with enabling the prediction model 24 (not shown in the figure); determines the gains 44 associated with enabling the prediction model 24 in the system 100; and conditionally enables the prediction model 24 in the system 100 based on the determined sustainability costs 42 associated with enabling the prediction model 24 in the system 100 and the determined gains 44 associated with enabling the prediction model 24 in the system 100.

[0051] In some examples, sustainability cost 42 includes one or more of the following: environmental costs, social costs, and economic costs. In some examples, sustainability cost 42 is environmental sustainability cost.

[0052] Figure 3 An example is illustrated, where sustainability cost 42 is energy cost, for example, estimated energy cost.

[0053] Figure 3 An example of device 2 is illustrated, which includes components for determining the relationship between system 100 and the device. Figure 2 Energy cost 42 associated with enabling the prediction model 24 in the system 100 (not shown); and gain 44 associated with enabling the prediction model 24 in the system 100. Based on the determined energy cost 42 associated with enabling the prediction model 24 in system 100 and the determined gain 44 associated with enabling the prediction model 24 in system 100, the enabling 46 of the prediction model 24 in system 100 is conditionally controlled.

[0054] In some examples, energy cost 42 includes: the energy cost of data movement within system 100 (e.g., inference data movement and / or training data movement), and the energy cost of movement of prediction model 24 within system 100.

[0055] like Figure 4A , Figure 4B , Figure 4C As shown, in at least some examples, conditionally controlling enable 46 includes: with respect to prediction model 24, performing one of the following: conditionally deploying prediction model 24 ( Figure 4A Conditionally activate prediction model 24 ( Figure 4B ); conditionally deactivate prediction model 24 ( Figure 4C ).

[0056] Figure 4A An example of device 2 is illustrated, which includes components for: determining a sustainability cost (energy cost) 42 associated with enabling (deploying) a predictive model 24 in system 100; determining a gain 44 associated with the deployment of the predictive model 24 in system 100; and conditionally controlling the deployment 46 of the predictive model 24 in system 100 based on the determined sustainability cost (energy cost) 42 associated with the deployment of the predictive model 24 in system 100 and the determined gain 44 associated with the deployment of the predictive model 24 in system 100.

[0057] In some examples, one or more control parameters 48, such as deployment conditions based on the estimated sustainability costs 42 determined in association with the use of the predictive model 24 in system 100 and the estimated gains 44 determined in association with the use of the predictive model 24 in system 100, are used to enable the deployment of the predictive model 24.

[0058] Figure 4BAn example of device 2 is illustrated, which includes components for: determining a sustainability cost (energy cost) 42 associated with enabling (activating) a prediction model 24 in system 100; determining a gain 44 associated with activation of the prediction model 24 in system 100; and conditionally controlling the activation 46 of the prediction model 24 in system 100 based on the determined sustainability cost (energy cost) 42 associated with activation of the prediction model 24 in system 100 and the determined gain 44 associated with activation of the prediction model 24 in system 100.

[0059] In some examples, one or more control parameters 48, such as activation conditions based on the estimated sustainability costs 42 determined in association with the use of the predictive model 24 in system 100 and the estimated gains 44 determined in association with the use of the predictive model 24 in system 100, are used to enable the activation of the deployed predictive model 24.

[0060] Figure 4C An example of device 2 is illustrated, which includes components for: determining a sustainability cost (energy cost) 42 associated with enabling (deactivating) a predictive model 24 in system 100; determining a gain 44 associated with deactivating a predictive model 24 in system 100; and conditionally deactivating a predictive model 24 in system 100 based on the determined sustainability cost (energy cost) 42 associated with deactivating a predictive model 24 in system 100 and the determined gain 44 associated with deactivating a predictive model 24 in system 100.

[0061] In some examples, one or more control parameters 48, such as the determined estimated sustainability cost 42 associated with the use of the predictive model 24 in system 100 and the determined estimated gain 44 associated with the use of the predictive model 24 in system 100, are used to enable the deactivation of the activated, deployed predictive model 24.

[0062] In some examples, the deployment conditions, activation conditions, and deactivation conditions are the same. That is, the same conditions are used for deployment and activation if they are met (or not met), and for deactivation if they are not met (or met).

[0063] In some examples, the deployment, activation, and deactivation conditions are not the same, but different. That is, the same condition, if met (or not met), is not used for deployment and activation, and if not met (or met), it is not used for deactivation. For example, the condition can be stateful, depending on the current state (deployed or not deployed; activated or not activated). In some examples, there is a hysteresis between the activation and deactivation conditions to prevent frequent switching.

[0064] In some examples, the enable 46 of the conditionally controlled prediction model 24 can be reconfigured to include one of the following: conditionally deploying the prediction model 24; conditionally activating the prediction model 24; or conditionally deactivating the prediction model 24.

[0065] In some examples, the prediction model 24 can be reconfigured via control parameter 48, which controls whether the enable is being deployed, activated, or deactivated.

[0066] In some examples, the enabling 46 of the predictive model 24 is conditionally controlled based on a reconfigurable comparison threshold T, which is used to compare the determined sustainability costs 42 (C) associated with the enabling of the predictive model 24 in system 100 with the determined gains 44 (G) associated with the enabling of the predictive model 24 in system 100.

[0067] Therefore, if C is the determined sustainability cost associated with enabling the predictive model 24 in system 100, and G is the determined gain 44 associated with enabling the predictive model 24 in system 100, then the condition is satisfied if GC > T1, where T1 is the comparison threshold; or if G / C > T2, where T2 is the comparison threshold. In this example, T2 = T1 / C + 1. In some examples, the reconfigurable comparison threshold can be reconfigured via control parameter 48 that controls the comparison threshold.

[0068] In some examples, cost 42 is integrated through anticipated users and / or gain 44 is integrated through anticipated users. Device 2 includes components for: integrating sustainability costs 42 associated with enabling predictive model 24 in system 100 through anticipated users, and integrating gains 44 associated with enabling predictive model 24 in system 100 through anticipated users, and conditionally enabling predictive model 24 in system 100 based on the integrated sustainability costs 42 associated with enabling predictive model 24 in system 100 and the integrated gains 44 associated with enabling predictive model 24 in system 100.

[0069] In some examples, cost 42 is based on the cost per user and the number of users being integrated with the expected number of users, and / or gain 44 is based on the gain per user and the number of users being integrated with the expected number of users. Apparatus 2 includes components for: obtaining combined sustainability cost 42, including multiplying the per-user sustainability cost associated with enabling prediction model 24 in system 100 by the expected number of users; and obtaining combined gain 44, including multiplying the per-user gain associated with enabling prediction model 24 in system 100 by the expected number of users; and conditionally enabling prediction model 24 in system 100 based on the combined sustainability cost 42 associated with enabling prediction model 24 in system 100 and the combined gain 44 associated with enabling prediction model 24 in system 100.

[0070] In some examples, the enabling 46 of the predictive model 24 in the system 100 is conditionally controlled to be dynamic and depends on the dynamically determined sustainability cost 42 associated with the enabling of the predictive model 24 in the system 100 and the dynamically determined gain 44 associated with the enabling of the predictive model 24 in the system 100, wherein the enabling 46 of the predictive model 24 in the system 100 responds to the time-varying sustainability cost 42 associated with the enabling of the predictive model 24 in the system 100 and the time-varying gain 44 associated with the enabling of the predictive model 24 in the system 100.

[0071] In some examples, the enable 46 of the predictive model 24 in the system 100 is conditionally controlled in real time and depends on the real-time determined sustainability cost 42 and the real-time determined gain 44 associated with the enable of the predictive model 24 in the system 100, wherein the control enable 46 of the predictive model 24 in the system 100 responds to real-time changes in the sustainability cost 42 associated with the enable of the predictive model 24 in the system 100 and in the real-time changes in the gain 44 associated with the enable of the predictive model 24 in the system 100.

[0072] For example, device 2 can be configured to monitor changes in system 100 and predict the performance of model 24, such as actual energy-related performance (energy cost 42) and actual model performance (gain 44).

[0073] In at least some examples, the determined gain 44 associated with enabling the prediction model 24 in system 100 is based on the usefulness of the prediction model 24.

[0074] In at least some examples, the determined gain 44 associated with enabling the prediction model 24 in system 100 is based on the utilization level of the prediction model 24.

[0075] In at least some examples, energy cost 42 is the energy cost for end-to-end energy consumption across system 100.

[0076] Figure 5 An example of system 100 is illustrated. In this example, system 100 is a hierarchical distributed system comprising: a higher level 102 with a higher rank in the hierarchy, and at least a lower level 104 with a lower rank in the hierarchy. The lower level 104 comprises multiple domains 106.

[0077] Predictive model 24 can be selectively deployed at a higher level 102. Predictive model 24 can also be selectively deployed at a lower level 104 in domain 106. The predictive model 24 deployed at a lower level 104 in domain 106 can be selectively the same as or different from the predictive model 24 deployed at a higher level 102. Predictive model 24 can be deployed in multiple domains 106. The predictive model 24 deployed in domain 106 can be selectively the same or different.

[0078] In at least some examples, sustainability cost 42 is an estimated energy cost at a higher level, at a lower level, or at both a higher and lower level. For example, sustainability cost 42 may include estimated energy costs for multiple domains 106.

[0079] In at least some examples, the determined gain 44 associated with enabling the prediction model 24 in system 100 is based on the usefulness of the prediction model 24.

[0080] In some examples, device 2 (end-to-end orchestration function) is configured to conditionally control the enabling of predictive model 24 across system 100, where sustainability cost 42 is an estimated energy cost for higher level 102. In some examples, device 2 is configured to validate the conditional enabling of predictive model 24 at domain 106 of lower level 104 of system 100. This provides end-to-end sustainability model deployment decisions.

[0081] In some examples, device 2 is configured to select a prediction model 24 for conditional deployment to a specific domain 106 of system 100, wherein prediction model 24 depends on the specific domain 106, and wherein sustainability cost 42 is an estimated energy cost for the specific domain 106. In some examples, the estimated energy cost for the specific domain 106 is based on user-related parameters of the specific domain 106, such as the expected number of users for each domain 106. The device may include components for activating or deactivating prediction models 24 deployed in different domains 106 and balancing overall gain 44 with sustainability cost 42.

[0082] In some examples, device 2 (domain-based) is configured to conditionally control the enabling of predictive model 24 at domain 106 of lower level 104 of system 100, where sustainability cost 42 is an estimated energy cost at domain 106 of lower level 104 of system 100. This provides domain-based sustainability model deployment decisions. In some examples, device 2 is configured to obtain validation from organizational functions at higher level 102 before enabling predictive model 24 within domain 106 of lower level 104 of system 100.

[0083] Figure 6 An example of system 100 is illustrated. In this example, system 100 is a cellular radio network 120. In this example, cellular radio network 120 includes: a core system 110, multiple radio access network (RAN) systems 112, and multiple user equipment (UE) 114 for each radio access network (RAN) system 112. At least some of the UEs 114 may be mobile.

[0084] In this example, cellular radio network 120 is as referenced Figure 5 The described hierarchical distributed system 100 includes: a higher level 102 with a higher rank in the hierarchical structure (core network 110), and at least a lower level 104 with a lower rank in the hierarchical structure (RAN). The lower level 104 includes multiple domains 106 (RAN system 112).

[0085] Predictive model 24 can be deployed at core system 110. Predictive model 24 can also be deployed at RAN system 112 (domain 106) at a lower level 104. The predictive model 24 deployed at RAN system 112 (domain 106) at a lower level 104 can be the same as or different from the predictive model 24 deployed at core system 110 (higher level 102). Predictive model 24 can be deployed at multiple RAN systems 112 (domain 106). The predictive model 24 deployed at RAN systems 112 (domain 106) can be the same or different.

[0086] refer to Figure 5 The described process also applies to Figure 6 Cellular radio network 120.

[0087] In at least some examples, energy cost 42 is the energy cost for end-to-end energy consumption across radio network 120, which includes core system 110, RAN system 112, and user equipment 114.

[0088] The integration of machine learning (ML) into cellular radio networks enables intelligent and adaptive communications. The ML model 24 itself requires significant computational power, which can lead to substantial energy consumption, impacting not only the network but also the user equipment (UE). ML model deployment encompasses all aspects of model selection and delivery. Orchestration, model management, and data management need to be controlled. In current 3GPP (3rd Generation Partnership Project) standards, model deployment / update signaling is used to initially deploy trained, validated, and tested AI / ML models (predictive models 24) to the model inference function or to deliver updated models (predictive models 24) to the model inference function. It describes AI / ML model training in OAM (Operations, Administration, and Maintenance) and AI / ML model inference in RAN (Radio Access Network) nodes. For example, ML model 24 is used by RAN nodes to control UE mobility optimization / handover.

[0089] The solution described above uses sustainability as a deployment criterion (in conjunction with gain) and avoids unsustainable distribution of ML models 24 in system 100. An example of a model sustainability requirement during ML model deployment is deploying multiple models 24 (e.g., at RAN nodes) only after considering the gain 44 of the expected rate based on incoming inference requests (e.g., de-prioritizing deployment to rural areas where only a small number of UEs are connected at specific locations and times of day).

[0090] Sustainability can be determined either at the domain level or at the end-to-end AI / ML orchestrator.

[0091] In some examples, cellular radio network orchestration, model management, and data management entities are used to deploy and dynamically enable / disable AI / ML models 24 in the network, taking into account some or all of the following gain 44 parameters: AI / ML model performance requirements (e.g., inference accuracy level); AI / ML model inference utilization (e.g., the minimum number of predictions requested per unit of inference time). Use case related gains 44 (e.g., reduced e2e energy consumption, socioeconomic benefits, such as ubiquitous coverage for the entire population). Network / UE platform capabilities (computer, memory, networking, etc.). And the following 42 sustainability parameters: The energy cost of the AI / ML chain is 42 (including the data, model, and executor phases, as well as the x / r applications involved). And the following sustainability / cost system parameters: Exposure (including observability) of data sources and network entities to be considered for AI / ML model deployment; The exposure and selection of different available profiles for network entities with different ML capabilities, which imply different energy consumption and performance capabilities; Security and privacy restrictions.

[0092] This approach can be reconfigurable in the sense of considering the different KPIs / KVIs of the (multiple) use cases supported by the AI / ML model 24. Dynamic enabling / disabling of the deployed ML model 24 can be based on network dynamics. Dynamic activation and deactivation of AI / ML management capabilities can be used to reduce the energy costs associated with management.

[0093] End-to-end (e2e) sustainable deployment can be handled by an e2e orchestrator working in conjunction with ML model management and data management capabilities, using a complete end-to-end system view, including consumer needs and sustainability constraints, as well as system status and capabilities.

[0094] Domain-specific sustainable deployment involves considering a local, domain-specific view for sustainable deployment and ML model activation / deactivation decisions. This decision is validated by an e2e orchestrator to ensure consistency at the end-to-end level.

[0095] Both approaches (e2e orchestration and domain orchestration) are used in Sustainability 42, and in... Figure 7A , Figure 7B , Figure 7C and Figure 8A , Figure 8B It is described in the text.

[0096] Figure 7A , Figure 7B , Figure 7C An example of an end-to-end (e2e) sustainable model enabling decision is illustrated. Higher-level functionality, e2e orchestrator 204 or model management 206, plays a central role in the decision 210 for enabling an e2e sustainable ML model.

[0097] Figure 7A The illustration shows the deployment of Sustainable ML Model 24. Figure 7B and Figure 7C The diagram illustrates the activation / deactivation of the deployed ML model 24. After ML model 24 is deployed, it can be dynamically activated and deactivated based on sustainability cost 42 and gain 44 criteria.

[0098] refer to Figure 7AThe orchestrator 204 determines the sustainable ML model deployment 210 based on the following inputs: service requirements 201 represented by service consumers 202, such as required service KPIs and energy consumption / efficiency targets; security and privacy requirements of network entities and UEs; and policies and requests 203 from service consumers 202 (e.g., operators). This could be, for example, in the following aspects: Due to the minimum network performance required by ML model 24 The desired performance of the target ML model, for example, in terms of achievable accuracy, energy efficiency, and net energy. The maximum resource usage required when using ML. The use of different energy sources is represented by the minimum amount of green energy required to power the ML model 24 (training, inference). Data management function 208 provides information 205 about the current and predicted network state, including but not limited to: For each given network load at different granularities, such as: Geographical region Specific network functions Specific residential communities or community groups, etc. The expected number of users given at each different granularity, such as: Geographical region Specific residential communities or community groups, etc. Network KPIs include performance, fault and configuration information, MDT measurements, etc. UE type (e.g., FWA or fixed / mobile UE). UE environment (e.g., a standalone UE or a UE that acts as a relay for other UEs or devices).

[0099] Model management function 206 provides information 207 from the ML model management entity about the available ML model 24, including: information about the use cases that the ML model 24 is processing; expected performance of the model 24; expected energy footprint of the model 24; and expected signaling burden due to the use of the model 24, such as data transmission as input to model inference or as input to model (re)training.

[0100] Based on input information 201, 203, 205, and 207, the orchestrator determines the deployment options that match 210. This can include: the type of ML model to be deployed (e.g., indicated by the use case name) or the ML model ID; the location where ML model 24 needs to be deployed is indicated by: the geographic region; and the network function ID to which ML model 24 needs to be deployed.

[0101] The orchestrator 204 is a device 2 that includes components for: determining a sustainability cost 42 associated with enabling the predictive model 24 in the system 100; determining a gain 44 associated with enabling the predictive model 24 in the system 100; and conditionally enabling the predictive model 24 in the system 100 based on the determined sustainability cost 42 associated with enabling the predictive model 24 in the system 100 and the determined gain 44 associated with enabling the predictive model 24 in the system 100.

[0102] Gain 44 can be based, for example, on service / network performance resulting from the use of (multiple) ML models 24, which can include the performance of ML models 24, which can be based on the number of users given at each different granularity.

[0103] Sustainability cost 42 can, for example, be based on energy consumption associated with gain 44, which includes resource usage when using ML model 44. It can include the network load given at each different granularity. It can include the use of different energy resources.

[0104] The deployment / activation environment can also be considered.

[0105] refer to Figure 7B After ML model 24 is deployed based on orchestrator 204 decision 210, the performance of the ML model is monitored by ML model management 206. Together with information 211 from ML data management function 208 regarding the monitoring of system / network performance due to the adoption of ML model(s) 24, ML model management 206 decides 210 to deactivate ML model 24, and activates ML model 24 based on this condition, and provides notification 213 to orchestrator 204.

[0106] Model management 206 is an apparatus 2 that includes components for: determining a sustainability cost 42 associated with enabling a predictive model 24 in system 100; determining a gain 44 associated with enabling a predictive model 24 in system 100; and conditionally enabling a predictive model 24 in system 100 based on the determined sustainability cost 42 associated with enabling a predictive model 24 in system 100 and the determined gain 44 associated with enabling a predictive model 24 in system 100.

[0107] refer to Figure 7C The orchestrator 204 uses information 207 about the performance of model 24 along with network / UE performance 211, and, if available, uses notifications 213 about activation / deactivation from ML model management 206, and derives a decision 210 about activating, deactivating, or redeploying model 24.

[0108] Based on consumer requirements and strategies, feedback can be provided to consumers regarding successful or unsuccessful ML deployments.

[0109] The orchestrator 204 is a device 2 that includes components for: determining a sustainability cost 42 associated with enabling the predictive model 24 in the system 100; determining a gain 44 associated with enabling the predictive model 24 in the system 100; and conditionally enabling the predictive model 24 in the system 100 based on the determined sustainability cost 42 associated with enabling the predictive model 24 in the system 100 and the determined gain 44 associated with enabling the predictive model 24 in the system 100.

[0110] When making decision 210, the following inputs can be considered: The performance of ML Model 24 includes energy-related performance, which includes: Observed model performance, such as model accuracy; The observed energy footprint of Model 24, such as model energy consumption, the proportion of green energy used for inference or training, etc. The observed increase in signaling load is due to the use of Model 24, for example, data transmission as input to model inference or as input to model (re)training; The observed network performance includes a comparison with the network performance without using ML model 24; Network KPIs that indicate the current state of the network, such as performance, number of users, network congestion, channel conditions and signal propagation conditions in a given area; Changes, for example, increases or decreases caused by ML model 24 operations; Network analysis and prediction, such as predictions and patterns of UE mobility, network load prediction, etc.; UE performance. Based on this input information, the ML model management 206 can decide to deactivate ( Figure 7B Or orchestrator 204 can decide to activate / deactivate, or redeploy already deployed model 24. Figure 7CFor example, if there is a guaranteed line-of-sight based on network and channel KPIs in a given area and ML model 24 is not needed for UE location prediction, then enabling ML model 24 has relatively lower gain and higher cost than disabling it, and ML model 24 is either not deployed, not activated, or deactivated. For example, if based on the measurement there are few UEs in a given area, and therefore adopting ML model 24 would be more energy-efficient than its effective factor cost (i.e., model 24 would only serve a very small number of users), then enabling ML model 24 has relatively lower gain and higher cost than disabling it, and ML model 24 is either not deployed, not activated, or deactivated.

[0111] After network / system conditions change, ML model management 206 may decide to reactivate the previously deactivated ML model 24, or orchestrator 204 may decide to redeploy ML model 24 for better reasons, such as due to a large increase in the number of UEs, changes in channel conditions, or the use of critical services, for which ML model prediction is needed.

[0112] Both approaches (e2e orchestration and domain orchestration) are used in Sustainability 42, and in... Figure 7A , Figure 7B , Figure 7C and Figure 8A , Figure 8B It is described in the text.

[0113] Figure 8A , Figure 8B The illustration shows an example of a domain sustainable model enabling decision. Lower-level domain components, domain node 200_1 (RAN domain) or domain orchestrator 304 (core network domain), play a central role in decision 210 for enabling the sustainable ML model.

[0114] Figure 8A and Figure 8B The illustration shows an example of a deployment / activation / deactivation decision 210 being taken at the domain level. Figure 8A In this process, decision 210 regarding deployment / activation / deactivation is taken at RAN node 200_1. Figure 8B In this process, the decision 210 regarding deployment / activation / deactivation is taken in each domain orchestrator 304.

[0115] exist Figure 8A In this process, adjacent RAN nodes 600_1 and 600_2 exchange information 221_1 and 223_1 that reveals the expected AI / ML model utilization coefficients, and reach an agreement on how to activate / deactivate or even transfer the AI / ML model 24 between them.

[0116] The proposed method can utilize standardized Xn messages (e.g., the predicted number of active UEs in a data collection request to assess the size of a potential data source set).

[0117] NG-RAN Node1 200_1 requests 221 information from neighboring node 200_2 (e.g., NG-RAN node 2) regarding the predicted number of active UEs, as well as further information indicating the number of users and the potential usefulness of the ML model 24 in the active RAN. This information 223 is exchanged via Xn, such as predicted RRC connections, average UE throughput UL / DL, measured UE trajectories, etc.

[0118] NG-RAN Node2 200_2 provides the necessary information 223 to NG-RAN Node1 200_1.

[0119] NG-RAN Node 1 200_1 can determine how useful enabling (e.g., activating) ML model 24 will be. For example, if only a limited number of UEs are expected to exist over a longer period, activating ML model 24 to provide inference for such a limited number of UEs may be unsustainable. Otherwise, in the case of a large number of UEs, high traffic, etc., NG-RAN Node 1 can decide to activate ML model 24 with specific capabilities (e.g., for load balancing).

[0120] The activation decision is sent to the adjacent RAN node 200_2 (225) and to the orchestrator (227). In this example, the orchestrator is the e2e orchestrator 204; however, in other examples, it could be the domain orchestrator 304. Orchestrators 204 and 304 can verify (230) whether the decision is based on sustainability criteria and confirm or reject (229, 231) the activation decision made at RAN node 1 200_1.

[0121] RAN node 200_1 is device 2, which includes components for: determining a sustainability cost 42 associated with enabling the prediction model 24 in system 100; determining a gain 44 associated with enabling the prediction model 24 in system 100; and conditionally enabling the prediction model 24 in system 100 based on the determined sustainability cost 42 associated with enabling the prediction model 24 in system 100 and the determined gain 44 associated with enabling the prediction model 24 in system 100.

[0122] Sustainability costs 42 and gains 44 are determined based on communication from node 200_2 at the same level.

[0123] exist Figure 8B In the middle, decision 210 was adopted at domain-level orchestrator 304, instead of as Figure 8A RAN node 200 was taken.

[0124] Domain orchestrator 304 requests information from neighboring nodes 200_1 and 200_2 (e.g., NG-RAN Node1 & NG-RAN Node2) regarding the predicted number of active UEs, as well as further information indicating the number of users and the potential usefulness of the ML model 24 in the active RAN. This information is exchanged via Xn, including predicted RRC connections, average UE throughput UL / DL, measured UE trajectories, etc.

[0125] NG-RAN node 1 200_1 provides the required information 223_1 to domain orchestrator 304.

[0126] NG-RAN node 2 200_2 provides the required information 223_2 to domain orchestrator 304.

[0127] Domain orchestrator 304 determines how useful the activation of ML model 24 will be. For example, if only a limited number of UEs are expected to exist over a longer period of time, activating ML model 24 to provide inference for such a limited number of UEs may be unsustainable. Otherwise, in the case of a large number of UEs, high traffic, etc., domain orchestrator 304 may decide to enable (e.g., activate) ML model 24 with specific capabilities (e.g., for load balancing).

[0128] Activation decisions are sent to adjacent RAN nodes 200_1 and 200_2 via 253 and 257, respectively, and to e2e orchestrator 204 via 249.

[0129] The e2e orchestrator 204 can verify whether the decision 230 (made at the domain orchestrator 304) is in accordance with sustainability criteria, and confirm or reject the activation decision 251.

[0130] Domain orchestrator 304 is device 2, which includes components for: determining a sustainability cost 42 associated with enabling the prediction model 24 in system 100; determining a gain 44 associated with enabling the prediction model 24 in system 100; and conditionally enabling the prediction model 24 in system 100 based on the determined sustainability cost 42 associated with enabling the prediction model 24 in system 100 and the determined gain 44 associated with enabling the prediction model 24 in system 100.

[0131] Sustainability costs 42 and gains 44 are determined based on communication from node 200_2 at the same level in the system.

[0132] The above method can be extended to ML model enabling (deployment / activation / deactivation) at the core network, for example, at an NWDAF (Network Data Analysis Function) instance or LMF (Location Management Function), where core network functions act as potential AI / ML model hosts (e.g., LMFs for UE location estimation). Considering an AI / ML-based UE location estimation task (e.g., ML model 24 deployed at an LMF), such a decision is made to activate ML model 24 for this task in a given cell when the relationship between gain 44 and cost 42 supports enabling ML model 24 (e.g., for most UEs connected to the cell, line-of-sight (LoS) paths are expected to be nonexistent).

[0133] In this example, the core network provides a device 2 that includes components for: determining a sustainability cost 42 associated with enabling the prediction model 24 in system 100; determining a gain 44 associated with enabling the prediction model 24 in system 100; and conditionally enabling the prediction model 24 in system 100 based on the determined sustainability cost 42 associated with enabling the prediction model 24 in system 100 and the determined gain 44 associated with enabling the prediction model 24 in system 100.

[0134] Orchestrators 204 and 304 can perform verification 230 and reject domain (RAN / CN) sustainable AI / ML model actions (or overall policies), and alternatively, recommend another more sustainable enable (deployment / activation / deactivation) option at the level of orchestrators 204 and 404. e2e orchestrator 204 can also define a set of e2e sustainable ML model deployment policies to be selected by the RAN node (or CN NF). For example, AI / ML model deployment for UE location estimation at the LMF rather than a gNB.

[0135] Based on AI / ML orchestration requirements 201, 203 (e.g., energy consumption targets / thresholds), information about the current and predicted network state 211, and collected data 207 about the energy consumption of management functions, the e2e orchestrator 204 can make decisions 210 about enabling (deploying / activating / deactivating) management capabilities in different domains 106 and balancing performance gains 44 with energy costs 42.

[0136] Figure 7A(Deployment): The AI / ML orchestrator 204 collects deployment requirements and strategies from the consumer 202, such as KP, energy requirements, gain 44, performance, resource usage, energy source, and security. The AI / ML orchestrator 204 collects network status from the data management 208, such as network load, expected number of users, KPIs, UE type, and UE environment. The AI / ML orchestrator 204 collects ML model information from the ML model management 206, such as model 24 use cases, model 24 performance, energy expectations, and signaling expectations.

[0137] AI / ML orchestrator 204 determines, based on the collected information, the sustainable deployment of ML models 210 with ML model types or model IDs, and the locations where ML models 24 must be deployed.

[0138] Figure 7B After ML model 24 is deployed, ML model management 206 monitors the performance of ML model 24 and collects updated network status, and decides to activate or deactivate ML model 24, and this decision is communicated to the AI / ML orchestrator.

[0139] Figure 7C Alternatively, after ML model 24 is deployed, AI / ML orchestrator 204 monitors the performance of ML model 24 and collects updated network status, and decides to activate, deactivate, or redeploy ML model 24. AI / ML orchestrator 204 sends ML model feedback (such as success or failure) to consumer 202.

[0140] Network status (such as network load and expected number of users) is collected at different granularities (such as geographic region, specific network function, specific cell, or cell group). The location where ML Model 24 needs to be deployed is provided as the geographic region and / or NF ID where ML Model 24 needs to be deployed. When making a decision about activating / deactivating ML Model 24, the following are considered, such as: observed Model 24 performance, observed energy consumption of ML Model 24, observed increase in signaling load due to the use of ML Model 24, and network performance observed before and after the deployment of ML Model 24, such as network KPIs indicating the current state of the network, changes caused by the operation of ML Model 24, network analysis and prediction (such as UE mobility and network load), and UE performance.

[0141] Figure 8AIn this context, domain 106 is the RAN, and decision 210 is made at RAN node 200. RAN node 1 sends a data collection request to RAN node 2 and receives a data collection response from RAN node 2. RAN node 1 (200_1) determines the sustainable activation / deactivation of ML model 24 (210) based on the data collection response. RAN node 1 sends the activation decision to neighboring RAN nodes 200 and orchestrators 204 and 304. The orchestrators (230) verify whether the decision is based on sustainability criteria and either confirm or reject the activation decision.

[0142] Figure 8B Domain 106 is the RAN, and decision 210 is made at the domain-level (RAN) orchestrator 304. RAN orchestrator 304 sends a data collection request to RAN node 200 and receives a data collection response from RAN node 200. RAN orchestrator 304 determines the sustainable activation / deactivation of ML model 24 based on the data collection response from RAN node 200. RAN orchestrator 304 sends the activation decision to e2e orchestrator 204. e2e orchestrator 204 verifies 230 whether the decision is based on sustainability criteria and confirms or rejects the activation decision, which is then sent to RAN orchestrator 304. RAN orchestrator 304 sends the decision received from e2e AI / ML orchestrator 204 to RAN node 200.

[0143] Data collection requests include the predicted number of active UEs, the usefulness of ML model 24 activated in RAN node 1, predicted RRC connections, average UE throughput UL / DL, and measured UE trajectories. The decision to activate ML model 24 can be made when there are a large number of UEs and high traffic, while ML model 24 may not be activated during times with a limited number of UEs and less traffic. Activation decisions include ML model ID, inference type, etc. ML model deployment can be extended to the core network (CN), where CN entities act as potential model hosts. OAM can act as the management entity for validating activation decisions in accordance with sustainability criteria.

[0144] Figure 7C The e2e orchestrator 204 collects orchestration requirements from the consumer 202, such as service KPIs and energy requirements / consumption. The e2e orchestrator 204 collects network status information from the data management 208, such as network load, expected number of users, KPIs, UE type, and UE environment. Based on the orchestration requirements and network status, the e2e orchestrator 204 determines whether to activate or deactivate AI / ML management capabilities within different domains 106. The e2e orchestrator 204 sends the decision to the domain-level orchestrator 304 and receives a response from the domain-level orchestrator 304. The e2e orchestrator 204 sends decision feedback to the consumer 202.

[0145] The described example demonstrates how coordinated efforts across AI / ML orchestration, model management, and data management can provide sustainable AI / ML enablement (deployment / activation / deactivation).

[0146] The example uses an end-to-end network view, taking into account consumer demand, sustainability constraints, and network capabilities to optimize model 24 deployment and activation decisions.

[0147] The example provides domain-specific enablement that allows for localized decisions tailored to specific conditions, and reduces energy costs through the dynamic activation and deactivation of AI / ML management capabilities.42

[0148] End-to-end sustainable deployment helps leverage a complete e2e network view, including consumer demand and sustainability constraints, as well as network status and capabilities. Domain-specific sustainable deployment helps provide a local, domain-specific view that takes into account sustainable deployment and ML model activation decisions. Dynamic activation and deactivation of AI / ML management capabilities reduces the energy costs associated with management.

[0149] It should be understood that the foregoing example describes a method comprising: determining a sustainability cost 42 associated with enabling a predictive model 24 in system 100; determining a gain 44 associated with enabling a predictive model 24 in system 100; and conditionally enabling a predictive model 24 in system 100 based on the determined sustainability cost 42 associated with enabling a predictive model 24 in system 100 and the determined gain 44 associated with enabling a predictive model 24 in system 100.

[0150] Figure 9 An example of a controller 400 is illustrated, which is configured to conditionally enable the predictive model 24 in system 100 based on a determined sustainability cost 42 associated with enabling the predictive model 24 in system 100 and a determined gain 44 associated with enabling the predictive model 24 in system 100. The controller 400 is suitable for use in device 2. The controller 400 can be implemented as a controller circuit system. The controller 400 can be implemented separately in hardware, have certain aspects of software including separate firmware, or can be a combination of hardware and software (including firmware).

[0151] like Figure 9 As shown, the controller 400 can be implemented using instructions that enable hardware functions, for example, by using executable instructions 406 in a general-purpose or special-purpose processor 402, which can be stored on a machine-readable storage medium (disk, memory, etc.) to be executed by such processor 402.

[0152] Processor 402 is configured to read from and write to memory 404. Processor 402 may also include: an output interface through which data and / or commands are output by processor 402, and an input interface through which data and / or commands are input to processor 402.

[0153] Memory 404 stores instructions, programs, or code 406 that control the operation of device 2 when loaded onto processor 402. The computer program instructions, programs, or code 406 provide the logic and routines that enable device 2 to perform the methods shown in the figures. By reading memory 404, processor 402 is configured to load and execute instructions, programs, or code 406.

[0154] The device 2 includes: at least one processor 402; and at least one memory 404 storing instructions that, when executed by the at least one processor 402, cause the device to conditionally enable the predictive model 24 in the system 100 based on a determined sustainability cost 42 associated with enabling the predictive model 24 in the system 100 and a determined gain 44 associated with enabling the predictive model 24 in the system 100.

[0155] like Figure 10 As shown, instructions, programs, or code 406 can reach device 2 via any suitable delivery mechanism 408. Delivery mechanism 408 can be, for example, a machine-readable medium, a computer-readable medium, a non-transitory computer-readable storage medium, a computer program product, a memory device, a recording medium (such as an optical disc read-only memory (CD-ROM), or a digital versatile optical disc (DVD), or a solid-state memory), including or tangibly embodying the computer program 406. The delivery mechanism can be a signal configured to reliably transmit the computer program 406. Device 2 can propagate or transmit the computer program 406 as a computer data signal.

[0156] As used herein, the term “non-transient” refers to a limitation on the medium itself (i.e., tangible, not tactile) rather than a limitation on the persistence of data storage (e.g., RAM vs. ROM).

[0157] Computer program instructions for causing the device to perform at least the following, or for performing at least the following: conditionally enabling the prediction model 24 in system 100 based on a determined sustainability cost 42 associated with enabling the prediction model 24 in system 100 and a determined gain 44 associated with enabling the prediction model 24 in system 100.

[0158] Computer program instructions can be included in a computer program, a non-transitory computer-readable medium, a computer program product, or a machine-readable medium. In some, but not necessarily all, examples, computer program instructions can be distributed across more than one computer program.

[0159] Although memory 404 is illustrated as a single component / circuit system, it can be implemented as one or more separate components / circuit systems, some or all of which may be integrated / removable and / or provide permanent / semi-permanent / dynamic / cached storage devices.

[0160] Although processor 402 is illustrated as a single component / circuit system, it can be implemented as one or more separate component / circuit systems, some or all of which may be integrated / removable. Processor 402 may be a single-core or multi-core processor.

[0161] References to “computer-readable storage medium,” “computer program product,” “tangibly embodying a computer program,” or “controller,” “computer,” “processor,” etc., should be understood to include not only computers with different architectures (such as single / multiple processor architectures and sequential (von Neumann) / parallel architectures), but also specialized circuits (such as field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), signal processing devices, and other processing circuitry systems). References to computer programs, instructions, code, etc., should be understood to include software for programmable processors or firmware, such as, for example, the programmable content of hardware devices, whether instructions for processors or configuration settings for fixed-function devices, gate arrays, or programmable logic devices, etc.

[0162] As used in this application, the term "circuit system" may refer to one or more or all of the following: (a) Implementation only in hardware circuit systems (such as implementation only in analog and / or digital circuit systems); and (b) A combination of hardware circuitry and software, such as (if applicable): i. A combination of (multiple) analog and / or digital hardware circuits with software / firmware, and ii. Any part of the (multiple) hardware processor(s) having software (including (multiple) digital signal processors), software, and (multiple) memories(s), working together to enable a device (such as a mobile phone or server) to perform various functions; and (c) (Multiple) hardware circuits and / or (multiple) processors, such as (multiple) microprocessors or portions of (multiple) microprocessors, which require software (e.g., firmware) to operate, but the software may not exist when operation is not required.

[0163] This definition of "circuit system" applies to all uses of the term in this application (including in any claim). As another example, as used in this application, the term "circuit system" also covers only the implementation of hardware circuitry or processors and their accompanying software and / or firmware. For example, and if applicable to elements of the claims, the term "circuit system" also covers baseband integrated circuits for mobile devices, or similar integrated circuits in servers, cellular network devices, or other computing or networking devices.

[0164] The boxes illustrated in the accompanying drawings may represent steps in the method and / or code segments in computer program 406. The specific order of the illustrated boxes does not necessarily imply a required or preferred order for the boxes, and the order and arrangement of the boxes can be changed. Furthermore, it may be possible to omit some boxes.

[0165] As used herein, "module" refers to a unit or device that excludes certain parts / components that will be added by the terminal manufacturer or user. Device 2 can be, for example, a module. The controller 400 of device 2 can be, for example, a module.

[0166] Where a structural feature has been described, the structural feature may be replaced by a component that performs one or more functions of the function that performs the structural feature, whether the function or these functions are explicitly or implicitly described.

[0167] Systems, apparatuses, methods, and computer programs can use machine learning, which may include statistical learning. Machine learning is a field of computer science that gives computers the ability to learn without being explicitly programmed. If a computer's performance on a task in T, as measured by P, improves with experience E, then the computer learns from experience E with respect to a class of tasks T and performance measurement P. Computers can typically learn from previous training data to predict future data. Machine learning includes fully or partially supervised learning and fully or partially unsupervised learning. It can enable discrete outputs (e.g., classification, clustering) and continuous outputs (e.g., regression). Machine learning can be implemented, for example, using different methods such as cost function minimization, artificial neural networks, support vector machines, and Bayesian networks. Cost function minimization can be used, for example, for linear and multinomial regression and K-means clustering. Artificial neural networks (e.g., with one or more hidden layers) model complex relationships between input and output vectors. Support vector machines can be used for supervised learning. A Bayesian network is a directed acyclic graph representing the conditional independence of multiple random variables.

[0168] Machine learning models24 can find applications in many technological fields. For example: - Classification of digital images, videos, audio, or speech signals based on low-level features (e.g., edge or pixel attributes of an image). - Control technology systems or processes, such as computer-controlled classification systems or industrial processes. - Determine the suitability for industrial processes based on measurements; - Digital audio, image, or video enhancement or analysis - Separation of source signals in speech signals; speech recognition, - Encoded data for reliable and / or efficient transmission or storage (and corresponding decoding); compression of audio, image, video, or sensor data; - Encrypt / decrypt or sign electronic communications; - Technical parameters (e.g., energy consumption, core temperature) are determined by processing data acquired from sensors. - Provides a reliability estimate for technical information (e.g., genotype); - Provide medical diagnostics through automated systems that process physiological measurements. - Derive or predict the physical state of an existing real object from the measurement of its physical properties. - The sensor data provided as input to ML model 24 is causally linked to the control command output for the control device provided as output to ML model 24.

[0169] The examples above find applications as enabling components for: automotive systems; telecommunications systems; electronic systems including consumer electronics; distributed computing systems; media systems for generating or presenting media content, including audio, visual, and audiovisual content, as well as mixed, media, virtual, and / or augmented reality; personal systems including personal health systems or personal fitness systems; navigation systems; user interfaces also known as human-machine interfaces; networks including cellular, non-cellular, and optical networks; ad-hoc networks; the Internet of Things; the Internet of Things; virtualized networks; and related software and services.

[0170] According to the examples of this disclosure, the device can be provided in an electronic device (e.g., a mobile terminal). However, it should be understood that the mobile terminal is merely an exemplary illustration of an electronic device that can benefit from examples of implementations of this disclosure, and therefore should not be used to limit the scope of the invention to the same scope. Although in some implementation examples, the device can be provided in a mobile terminal, other types of electronic devices, such as, but not limited to: mobile communication devices, handheld portable electronic devices, wearable computing devices, portable digital assistants (PDAs), pagers, mobile computers, desktop computers, televisions, gaming devices, laptop computers, cameras, video recorders, GPS devices, and other types of electronic systems that can readily employ the examples of this disclosure. Furthermore, the device can readily employ the examples of this disclosure regardless of whether it is intended to provide mobility.

[0171] The term "includes" is used in this document and has an inclusive rather than exclusive meaning. That is, any reference to X that includes Y indicates that X may include only one Y or may include more than one Y. If it is intended to use "includes" with an exclusive meaning, it will become clear in the context by referring to "only one..." or by using "comprises".

[0172] In this specification, the terms “connection,” “coupling,” and “communication,” and their derivatives, mean operatively connecting / coupling / communicating. It should be understood that any number or combination of intermediate components (including no intermediate components) may be present, i.e., providing direct or indirect connection / coupling / communication. Any such intermediate component may include hardware and / or software components.

[0173] As used herein, the term "determine" (and its grammatical variations) can include, but is not limited to: calculation, operation, processing, derivation, measurement, investigation, identification, lookup (e.g., searching in a table, database, or other data structure), ascertainment, etc. Furthermore, "determine" can include receiving (e.g., receiving information), accessing (e.g., accessing data in memory), obtaining, etc. Additionally, "determine" can include parsing, selecting, picking, building, etc.

[0174] Various examples are referenced in this specification. Descriptions of features or functions associated with an example indicate which features or functions exist in that example. Whether explicitly stated or not, the use of the terms "example," "for example," "able," or "may" in the text indicates that, whether described as an example or not, such features or functions exist at least in the described example, and they may, but not necessarily, exist in some or all other examples. Therefore, "example," "for example," "able," or "may" refers to a specific instance of a particular class of examples. A characteristic of an instance may be a characteristic of only that instance, a characteristic of the class, or a characteristic of a subclass of a class that includes some, but not all, instances of that class. Thus, a feature described with reference to one example but not to another is implicitly disclosed, and where possible, may be used as part of a working composition in other examples, but is not necessarily required to be used in other examples.

[0175] As used herein, “at least one of the following:” and “at least one of the following” and similar wording, where a list of two or more elements is connected by “and” or “or”, means at least any one of the elements, or at least any two or more of the elements, or at least all of the elements.

[0176] Although examples have been described with reference to various examples in the preceding paragraphs, it should be understood that modifications to the given examples may be made without departing from the scope of the claims. Features described in the preceding specification may be used in combinations other than those expressly described above. Although a function has been described with reference to certain features, that function may be performed by other features, whether or not it is described.

[0177] Additionally, a description of a feature (such as a device or a component of a device) configured to perform a function or for performing a function should be considered as disclosing a method for performing that function. For example, a description of a device configured to perform one or more actions, or for performing one or more actions, should be considered as disclosing a method for performing those one or more actions with or without using the device.

[0178] Although a feature has been described with reference to some examples, that feature may exist in other examples, whether or not it has been described.

[0179] The terms “a,” “an,” or “the” are used in this document and have an inclusive rather than exclusive meaning. That is, unless the context clearly indicates otherwise, any reference to X that includes “a,” “an,” or “the” Y indicates that X may include only one Y or may include more than one Y. If it is intended to use “a,” “an,” or “the” with an exclusive meaning, it will be clearly stated in the context. In some cases, the use of “at least one” or “one or more” may be used to emphasize an inclusive meaning, but the omission of these terms should not be used to infer any exclusive meaning.

[0180] The presence of a feature (or combination of features) in a claim is a reference to that feature or combination of features itself, as well as to a feature that achieves the same technical effect (an equivalent feature). An equivalent feature includes, for example, a feature that is a variation and achieves the same result in the same manner. An equivalent feature includes, for example, a feature that performs the same function in the same manner to achieve the same result.

[0181] In this specification, reference has been made to various examples in which adjectives or adjective phrases are used to describe the characteristics of the examples. Such descriptions of characteristics associated with examples indicate that the characteristic exists exactly as described in some examples, and substantially as described in others.

[0182] The foregoing description describes some examples of this disclosure; however, those skilled in the art will recognize possible alternative structures and methodological features that provide functionality equivalent to specific examples of such structures and features described above, and for the sake of brevity and clarity, such structures and features are omitted from the foregoing description. Nevertheless, the foregoing description should be construed as implicitly including references to such alternative structures and methodological features that provide equivalent functionality, unless such alternative structures or methodological features are expressly excluded in the foregoing description of examples of this disclosure.

[0183] Although efforts have been made in the foregoing specification to focus attention on those features deemed important, the applicant may seek protection by means of the claims for any patentable features or combinations thereof mentioned above and / or shown in the figures, regardless of whether they are particularly emphasized.

Claims

1. A communication apparatus comprising components for: Determine the sustainability costs associated with enabling the predictive models in the system; Determine the gain associated with enabling the prediction model in the system; The enabling of the prediction model in the system is conditionally controlled based on the determined sustainability cost associated with enabling the prediction model in the system and the determined gain associated with enabling the prediction model in the system.

2. The apparatus of claim 1, wherein sustainability cost is an estimated energy cost, wherein the energy cost includes: The energy cost of data movement within the system, and the energy cost of moving the prediction model within the system.

3. The apparatus of claim 1 or 2, wherein the system is a cellular radio network, and wherein conditional control enable includes: Regarding the prediction model, perform one of the following: The prediction model may be deployed conditionally. The prediction model is conditionally activated; The prediction model is activated conditionally.

4. The apparatus of claim 1 or 3, wherein deployment conditions based on the estimated sustainability cost determined in association with the use of the prediction model in the system and the estimated gain determined in association with the use of the prediction model in the system are used to enable the deployment of the prediction model. Activation conditions based on the estimated sustainability cost determined in association with the use of the prediction model in the system and the estimated gain determined in association with the use of the prediction model in the system are used to enable the activation of the deployed prediction model. The deactivation conditions, based on the estimated sustainability cost determined in association with the use of the predictive model in the system and the estimated gain determined in association with the use of the predictive model in the system, are used to enable the deactivation of the activated, deployed predictive model, wherein the deployment conditions, the activation conditions, and the deactivation conditions are different.

5. The apparatus according to any one of claims 1 to 4, wherein the conditional control of the enabling of the prediction model is reconfigurable to include one of the following: The prediction model may be deployed conditionally. The prediction model is conditionally activated; The prediction model is activated conditionally.

6. The apparatus according to any one of claims 1 to 5, characterized by one or more of the following: The enabling of the prediction model is conditionally controlled based on a reconfigurable comparison threshold, which is used to compare the determined sustainability cost associated with enabling the prediction model in the system with the determined gain associated with enabling the prediction model in the system. Includes components for: the sustainability costs associated with enabling the predictive model in the system through anticipated user integration, and the gains associated with enabling the predictive model in the system through anticipated user integration. The prediction model in the system is conditionally enabled based on the integrated sustainability costs associated with enabling the prediction model in the system, and the integrated gains associated with enabling the prediction model in the system. Components for: obtaining combined sustainability costs, including: Multiplying the per-user sustainability cost associated with enabling the predictive model in the system by the expected number of users, and obtaining the combined gain, includes: multiplying the per-user gain associated with enabling the predictive model in the system by the expected number of users. The predictive model in the system is conditionally enabled based on the combined sustainability cost associated with enabling the predictive model in the system, and the combined gain associated with enabling the predictive model in the system; or The enabling of the predictive model in the system is conditionally controlled to be dynamic and depends on the dynamically determined sustainability cost associated with the enabling of the predictive model in the system and the dynamically determined gain associated with the enabling of the predictive model in the system, wherein the enabling of the predictive model in the system is controlled in response to changes in the sustainability cost associated with the enabling of the predictive model in the system and changes in the gain associated with the enabling of the predictive model in the system.

7. The apparatus according to any one of claims 1 to 6, wherein one or more of the following: The determined gain, associated with enabling the prediction model in the system, is based on the usefulness of the prediction model; or The determined gain, which is associated with enabling the prediction model in the system, is based on the utilization level of the prediction model.

8. The apparatus according to any one of claims 1 to 7, wherein the system is a hierarchical distributed system, the hierarchical distributed system comprising: The higher level of a higher ranking in the hierarchy, and at least the lower level of a lower ranking that includes multiple domains in the hierarchy, wherein sustainability cost is an estimated energy cost at the higher level, at the lower level, or at both the higher and lower levels.

9. The apparatus according to claim 8, characterized by one or more of the following: The device is configured to conditionally control the enabling of the predictive model across the system, wherein sustainability cost is the estimated energy cost for the higher level. The device is configured to verify the conditional enabling of the prediction model at a lower-level domain of the system. The device is configured to select a predictive model for conditional deployment to a specific domain of the system, wherein the predictive model depends on the specific domain, and wherein the sustainability cost is an estimated energy cost for the specific domain. The estimated energy cost for the specific domain is based on user-related parameters for that specific domain. The device includes: A component used to activate or deactivate predictive models deployed in different domains, balancing overall performance gains with sustainability costs. The device is configured to conditionally control the enabling of the predictive model at the lower-level domain of the system, wherein the sustainability cost is an estimated energy cost at the lower-level domain of the system, or The device is configured to obtain validation from organizational functions at a higher level before enabling the prediction model within the domain at a lower level of the system.

10. A method for communication, comprising: Determine the sustainability costs associated with enabling the predictive models in the system; Determine the gain associated with enabling the prediction model in the system; as well as The enabling of the prediction model in the system is conditionally controlled based on the determined sustainability cost associated with enabling the prediction model in the system and the determined gain associated with enabling the prediction model in the system.

Images