The project leading to this application has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No 101015956.

Technical Field

The present invention relates to various methods and devices for handling the transfer of a message in a wireless communications network, wherein the message is related to artificial intelligence and the handling is based on characteristics of the artificial intelligence.

Background

Fig. 1 shows a 5G NR architecture with service based interfaces. The 5G core network part comprises a Network Slice Selection Function, NSSF 10, a Network Exposure Function, NEF, 1 1 , a Network Repository Function, NRF, 12, a Policy Control Function, PCF, 13, a Unified Data Management, UDM, 14, an Application Function, AF, 15, an Authentication Server Function, AUSF, 16, an Access and Mobility Management Function, AMF, 17, and a Session Management Function, SMF, 18. A User Equipment, UE, 1 , is connected to the Radio Access Network, RAN, 19, wherein a User Plane Function, UPF, 20 is provided to connect the UE 1 to a data network, DN, 21 .

Having service based interfaces in the 5G Core Control Plane (CPF) implies that the Network Functions, NFs, in the 5G Core CPF provide services that are consumed by other NFs in the 5G Core.

The roles of these entities and the interfaces have been defined in the 3GPP TS 23.181 and the procedures have been described in TS 23.182.

In future evolution of the 5G network, such as 6G, usage of artificial intelligence, Al, is expected in various layers of the radio stack, from low layers such as PHY to higher layers such as RRC, as well as in various of the entities described above.

This imply that signaling, and in particular over-the-air signaling, will need to efficiently support transfer of Al-related information such as, but not limited to, Al models, data, metadata, model control parameters, etc. In current wireless communications networks, the UPF exploits the so-called Quality of Service, QoS, framework, where traffic is associated to a certain QoS, for instance 5QI in 5G, and rules are defined to understand which traffic should be mapped to which QoS. More details on this can be found, for instance, in 3GPP TS 23.501 , TS 23.502.

Information about which QoS to apply for a certain traffic are given by the policy framework, usually based on agreement between the mobile operator and the customer/subscriber. The policy framework then generates QoS rules which are delivered to the different entities of the wireless communications network, for instance RAN, AMF, SMF, UPF, where such rules indicate how traffic should be filtered and which 5QI should be associated. Known wireless communications networks also provide the possibility to dynamically change the QoS configuration based on inputs/requests from the end-application. However, in general, once a QoS policy is agreed then all the traffic associated to it would be exploiting the same QoS.

From a UPF point of view, the framework allows the wireless communications network to identify a specific flow of packets to consequently apply a certain QoS. However, within the same flow it is not possible to differentiate the treatment of specific packets according to the 3GPP QoS framework.

Per-packet prioritization has been analyzed in literature, for instance in prior art documents US 2008/0123660 A1 or US 2009/0073971 A1 , but overall all solutions for per-packet traffic handling mechanisms rely on having QoS-related information in fields of the upper layers such as IPv4 type of service, TOS, differentiated service code point DSCP field in an IP packet, etc.

That is, per-packet prioritization for UPF traffic is based on the assumption that there are QoS- related information from upper-layers such as TCP/IP, so that the wireless communications network stack can check such packet header fields and consequently map the packet to the relevant QoS. Whereas having such upper-layers is feasible for user-plane traffic generated by end-applications, such layers might not be present in a packet generated by an in-network functionality. In fact, 3GPP stack has no IP layer between the UE and the base station, as well as no IP layer between the UE and the AMF for NAS signaling, and consequently per-packet prioritization based on solutions relying on having layers such as TCP/IP are not applicable in Al traffic for in-network Al functionalities.

From a CPF point of view, there is a difference between radio interface and within the core network. For the radio interface, CPF, also including messages from NAS, is transmitted via RRC, as described in 3GPP TS 38.331 , which associates different types of message to different signaling radio bearers, SRBs. The SRBs can be configured in a different way to guarantee different signaling prioritization and treatment, but the mapping is fixed and once the configuration of a certain SRB is done then such configuration applies to all traffic mapped over that SRB.

On the other hand, within the core network, there is no mentioning or concept of different treatment for different types of messages. If messages from core network functions have to be delivered to a UE, then NAS, from AMF, should be used, and the messages should be mapped to either SRB1 or SRB2.

In future evolutions of wireless communications networks, the networks will need efficient ways of transferring Al-related traffic used by in-network Al functionalities. This aspect is of particular importance especially for transfer over the radio interface given the scarcity of radio resources and multiplexing needs considering other traffic.

When transferring Al-related traffic, such as Al-models or configuration parameters, how the traffic should be treated, for instance in terms of priority, reliability, etc., depends on several aspects, including which Al solution is used, training set, training procedure, etc.

For instance, when transferring an Al model, the exchanged traffic might be treated differently if two different models are transferred, although the signaling procedure is basically the same.

As an example, different Al models differ in their robustness toward reception error. Consequently, reception errors have an impact on the accuracy of the model which depends on the specific model under consideration. As an example, a model A can be trained on dataset D1 , whereas a model B can be trained on dataset D2. If D1 is independent and identically distributed, compared to D2, and it expresses richer scenarios, then it is expected that model A has more to learn than model B. In this case it might be that errors in receiving weights is more critical to its performance. As another example, the depth and wideness of a model C can be larger than for a model D, where both are trained on the same dataset. Such difference in the size, might enable model C to be acceptable to more loss on weights.

Another example is the robustness of the Al related information to communication delay. For example, Al data related to timewise fast scale quantities, for instance predicted CSI, are not delay tolerant compared to slow scale quantities, for instance predicted L3 filtered serving cell RSRP.

Current 3GPP mechanism is based on a static mapping and static SRB configuration, where the granularity of the mapping depends on which message is transferred and not on the content of the message itself.

Therefore, the strict legacy prioritization approach with static mapping of the UPF and/or CPF might involve several limitations when transferring Al related information.

There is therefore a need to provide a signalling of Al-related information which overcomes the issues described above. This need is met by the features of the independent claims. Further aspects are described in the dependent claims.

An embodiment can relate to a method for controlling the transfer of messages related to artificial intelligence in a wireless communications network. The wireless communications network can comprise at least a service provider, implementing an artificial intelligence model for providing services to at least one service consumer. The method can comprise a step of generating, by the service provider, or by the at least one service consumer, a message related to the artificial intelligence model. The method can further comprise a step of generating, by a generating entity, assistance information related to characteristics of the artificial intelligence model and/or characteristics of the message. The method can further comprise a step of transmitting, by the generating entity to a mapping entity, the assistance information. The method can further comprise a step of mapping, by the mapping entity, the message to resources of the wireless communications network, based on the assistance information. The method can further comprise a step of controlling transmission of the message using the mapping resulting from the mapping step.

In some embodiments, the message can comprise any of the artificial intelligence model, control information for the artificial intelligence model, input for the artificial intelligence model, or output of the artificial intelligence model.

In some embodiments, the assistance information can comprise any of an importance indicator, for indicating a contribution of the message to operation of the artificial intelligence model, a robustness indicator, for indicating robustness of the artificial intelligence model against radio losses in the message, an experience indicator, for indicating a level of training of the artificial intelligence model, or a generality indicator, for indicating a capability of the artificial intelligence model to perform with high robustness against different sets of inputs.

In some embodiments, the level of training of the artificial intelligence model comprises any of a number of users having contributed to training of the artificial intelligence model, a number of iterations performed for training of the artificial intelligence model, or statistics of a training set used when training the artificial intelligence model.

In some embodiments, the importance indicator can have a lower value if the contribution is low, or a higher value if the contribution is high. Alternatively, or in addition the robustness indicator can have a lower value if the robustness is low, or a higher value if the robustness is high. Alternatively, or in addition, the experience indicator can have a lower value if the number of users is low, or a higher value if the number of users is high. Alternatively, or in addition, the generality indicator can have a lower value if the capability if low, or a higher value if the capability is high.

In some embodiments, the mapping step can comprise computing a priority value for the message based on a function having as input values any of the importance indicator, the robustness indicator, the experience indicator, or the generality indicator.

In some embodiments, the function can be a monotonic function configured to increase if at least one of the input values increases, and/or configured to decrease if at least one of the input values decreases.

In some embodiments, in the mapping step, resources of the wireless communications network can be allocated to the message as a function of the priority value.

In some embodiments, resources of the wireless communications network can be allocated so that a priority value results in a higher priority of use of resources by the message.

In some embodiments, the mapping step can be configured to increase a logical channel priority (LCP) of the message, preferably to a level above “MAC CE for BSR” and/or below “C- RNTI of CCCH”, if the importance indicator has a value higher than a predetermined threshold. Alternatively, or in addition, the mapping step can be configured to increase MAC Bucket Size, preferably via Prioritized Bit Rate and/or Bucket Size Duration, if the importance indicator and/or the experience indicator has a value higher than a predetermined threshold. Alternatively, or in addition, the mapping step can be configured to increase CSI-RS occasions for multiple UEs or for specific group of UEs, if the robustness indicator and/or the experience indicator and/or the generality indicator has a value higher than a predetermined threshold. Alternatively, or in addition, the mapping step can be configured to allocate broadcast channel to a group of UEs to enable generality via federated learning, wherein the artificial intelligence model can be broadcasted to the group of UEs, if the experience indicator and/or the generality indicator has a value higher than a predetermined threshold. Alternatively, or in addition, the mapping step can be configured to decrease a BLER target of a Link adaptation, or MCS selection, if the importance indicator and/or the robustness indicator has a value higher than a predetermined threshold.

In some embodiments, the artificial intelligence model can be configured to exchange RAN-AI messages with one or more UE(s), for exchanging data between the artificial intelligence model and the one or more UE(s). Moreover, the artificial intelligence model can be configured to output RAN-AI actuation messages, for controlling Radio Resource Management function, preferably for MAC and PHY layers.

In some embodiments, the method can further comprise the step of obtaining, by the mapping entity from the service consumer, a mapping configuration. The mapping step can then be based on the assistance information and on the mapping configuration.

In some embodiments, the method can further comprise the step of obtaining, by the mapping entity from the service consumer, a correction value, and a step of changing, by the mapping entity, any of the importance indicator, the robustness indicator, the experience indicator or the generality indicator by an amount corresponding to the correction value.

In some embodiments, the method can further comprise the step of obtaining, by the generating entity, a generating policy for configuring the way in which the generating entity performs the generating step.

In some embodiments, the method can further comprise the step of obtaining, by the mapping entity, a mapping policy for configuring the way in which the mapping entity performs the mapping step.

A further embodiment can relate to a method for operating a generating entity in a wireless communications network. The wireless communications network can comprise a service provider implementing an artificial intelligence model for providing services to at least one service consumer. The method can comprise a step of obtaining characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model. The method can further comprise a step of generating assistance information related to characteristics of the artificial intelligence model and/or characteristics of the message. The method can further comprise a step of transmitting, to a mapping entity, the assistance information.

A further embodiment can relate to a method for operating a mapping entity in a wireless communications network. The wireless communications network can comprise a service provider implementing an artificial intelligence model for providing services to at least one service consumer. The method can comprise a step of obtaining assistance information related to characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model. The method can further comprise a step of mapping the message to resources of the wireless communications network, based on the assistance information.

A further embodiment can relate to a generating entity for use in a wireless communications network. The wireless communications network can comprise a service provider implementing an artificial intelligence model for providing services to at least one service consumer. The generating entity can comprise a processor and a memory, the memory comprising instructions configured to, when executed by the processor, cause the processor to carry out method steps. The method can in particular comprise a step of obtaining characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model. The method can further comprise a step of generating assistance information related to characteristics of the artificial intelligence model and/or characteristics of the message. The method can further comprise a step of transmitting to a mapping entity (240), the assistance information.

A further embodiment can relate to a mapping entity for use in a wireless communications network. The wireless communications network can comprise a service provider implementing an artificial intelligence model for providing services to at least one service consumer. The mapping entity can comprise a processor and a memory, the memory comprising instructions configured to, when executed by the processor, cause the processor to carry out method steps. The method can in particular comprise a step of obtaining assistance information related to characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model. The method can further comprise a step of mapping the message to resources of the wireless communications network, based on the assistance information.

Brief Description of Drawings

Various features of embodiments will become more apparent when read in conjunction with the accompanying drawings. In these drawings:

Figure 1 schematically illustrates the 5G NR reference architecture as defined by 3GPP;

Figure 2 schematically illustrates a possible configuration of elements of a wireless communications network;

Figure 3 schematically illustrates a method for controlling the transfer of messages related to artificial intelligence in a wireless communications network;

Figures 3A, 3B and 3C schematically illustrate further possible steps of the method of figure 3;

Figure 4 schematically illustrates a method for operating a generating entity in a wireless communications network;

Figure 5 schematically illustrates a method for operating a mapping entity in a wireless communications network;

Figure 6 schematically illustrates a generating entity for use in a wireless communications network;

Figure 7 schematically illustrates a mapping entity for use in a wireless communications network.

Detailed

In the following, embodiments of the invention will be described in detail with reference to the accompanying drawings. It is to be understood that the following description of embodiments is not to be taken in a limiting sense. The scope of the invention is not intended to be limited by the embodiments described hereinafter or by the drawings, which are to be illustrative only. The drawings are to be regarded as being schematic representations, and elements illustrated in the drawings are not necessarily shown to scale. Rather, the various elements are represented such that their function and general purpose becomes apparent to a person skilled in the art. Any connection or coupling between functional blocks, devices, components of physical or functional units shown in the drawings and described hereinafter may also be implemented by an indirect connection or coupling. A coupling between components may be established over a wired or wireless connection. Functional blocks may be implemented in hardware, software, firmware, or a combination thereof.

Figure 2 schematically illustrates a possible configuration of elements of a wireless communications network 200. It will be clear to those skilled in the art that the elements of figure 2 are to be understood as functional elements, and can be implemented by any of the entities, or functions, of a 5G network, as illustrated in figure 1 . Alternatively, or in addition, one or more new entity, or function, can be introduced to a 5G network in order to implement one or more of the elements of figure 2. That is, the network 200 could be built on the basis of an existing 5G network, and in the following references will also be made to specific 5G functions. Nevertheless, it is clear that the invention can also be implemented in other context, for instance on the basis of a 3G or 4G network, or on the basis of yet other networks.

Wireless communications network 200 therefore comprises at least a service provider 210 and at least one service consumer 220.

In particular, the service provider 210 provides services which can be used by the service consumer 220. Various manners for providing and consuming services are, per se, known. In the invention, the service provider is particularly characterized in that it implements an artificial intelligence model for providing the services to the service consumer. It will be clear to those skilled in the art that a plurality of artificial intelligence models can be implemented, depending on the needs of the specific service. Moreover, it will be clear that a plurality of services can benefit from using an artificial intelligence model. Therefore, although specific embodiments can be described with reference to specific services and/or specific artificial intelligence models, the present invention is not to be understood as being limited to those embodiments.

Wireless communications network 200 further comprises a transmission channel 260 and a resources manager 250. The transmission channel can be implemented by any physical or logical channel through which messages can be exchanged between the service provider 210 and the service consumer 220. The resource manager 250 can be any function of the network 200 which has the capability of configuring the use of the transmission channel 260, and/or controlling how messages exchanged between the service provider 210 and the service consumer 220 are transmitted through the transmission channel 260.

For instance, in an implementation leveraging a 5G architecture, the service consumer 220 could be a User Equipment, and the service provider could be any of the function of the core network. In this exemplary configuration, the transmission channel 260 could be the physical, wireless, channel between the User Equipment and the core network, or any of the logical channels defined in this physical channel. Still in the same exemplary configuration, the resource manager 250 could be any entity, or function, of the 5G network capable of affecting how the messages are exchanged on the transmission channel 260. One example of the resource manager 250 could be, for instance, the UPF, or the SMF, as it will be clear to those skilled in the art how those elements can configure the use of resources of the network for the transmission of messages.

Wireless communications network 200 further comprises at least a generating entity 230 and a mapping entity 240. Those two entities are illustrated separately to independently describe their function. It will however be clear that they can be combined, at a logical and/or physical level, in a single entity.

Figure 3 schematically illustrates a method 300 for controlling the transfer of messages related to artificial intelligence in wireless communications network 200.

In particular, it has been recognized by the inventors that messages related to artificial intelligence, for instance related to the artificial intelligence model implemented by the service provider 210, can be advantageously mapped to resources of the transmission channel 260 depending on the characteristics of the artificial intelligence model and/or of the message itself.

That is, in the prior art, messages between a service provider and a service consumer are mapped to specific resources of the transmission channel 260 depending on various parameters. However, when the service provider implements an artificial intelligence model, in addition or alternatively to the current used parameters, it is advantageous to maps the messages to specific resources also based on characteristics of the artificial intelligence model and/or of the message.

In particular, as visible in figure 3, the method 300 comprises a step S31 of generating, by the service provider 210, or by the at least one service consumer 220, a message related to the artificial intelligence model. In preferred embodiments, the message can comprise any of the artificial intelligence model, control information for the artificial intelligence model, input for the artificial intelligence model, or output of the artificial intelligence model. In general, all of those messages relate to the artificial intelligence model in that their include data to and/or from the artificial intelligence model, or the artificial intelligence model itself.

As schematically illustrated in figure 3, the step S31 can be implemented at the service provider 210 and/or at the service consumer 220. For instance, a message comprising an output from the artificial intelligence model can be generated at the service provider 210 for transmission to the service consumer 220. Conversely, a message comprising an input from the artificial intelligence model can be generated at the service consumer 220 for transmission to the service provide 210.

The method 300 can further comprise a step S32 is transmitting characteristics of the artificial intelligence model and/or characteristics of the message to the generating entity 230. As for step S31 , step S32 can be implemented at the service provider 210 and/or at the service consumer 220.

In the illustrated embodiment, the step S32 is illustrated as being implemented by the same entity implementing step S31. However, the present invention is not limited thereto. For instance, the service consumer 220 could generate the message, while the characteristics of the artificial intelligence model, could be known at the service provider. In this configuration, step S32 could for instance be implemented by both the service provider 210, in order to transmit to the generating entity 230 information on the artificial intelligence model, and by the service consumer, 220, in order to transmit to the generating entity 230 information on the message. It will be therefore clear that the service provider 210 and/or the service consumer 220 can generate the message at step S31 and the service provider 210 and/or the service consumer 220 can transmit the characteristics at step S32, independently on which of the two entities carried out step S31. In preferred embodiments of the invention, the service provider 210 can be configured to implement step S32, in order to transmit to the generating entity 230 information on the artificial intelligence model, and the service consumer 220 and/or the service provider 210 can be configured to implement step S32, in order to transmit to the generating entity 230 information on the message.

Examples for the characteristics of the artificial intelligence model and/or of the message will be described in the following description. Nevertheless, it will be clear to those skilled in the art that generally any type of characteristic, which has an impact on how transmission of the message can be affected by the mapping to different resources of the transmission channel 260, can be transmitted.

The method 300 further comprises a step S33 of generating, by the generating entity 230, assistance information related to characteristics of the artificial intelligence model and/or characteristics of the message.

Examples of the assistance information will be described more in details in the following. It will however be clear that the assistance information can comprise one or more parameters which allow various types of characteristics of the artificial intelligence model and/or characteristics of the message to be measured with predetermined values.

That is, for instance with respect to the characteristics of the artificial intelligence model, it is clear that different artificial intelligence models A and B can comprise different sets of characteristics, so that it can be difficult to define which resources of the transmission channel are to be associated to the messages related to artificial intelligence model A and which to the messages related to artificial intelligence model B. The generating entity 230 can thus receive as input the characteristics of the artificial intelligence model, and/or of the message, and output the assistance information in a manner which allows a direct comparison of characteristics of the artificial intelligence model A and B. In some embodiments, this can be implemented by configuring the generating entity 230 so that for any given characteristic of any artificial intelligence model, and/or message, the generating entity 230 can associate a numerical value and then output one or more numerical values deriving from this association.

The method 300 further comprises a step S34 of transmitting the assistance information from the generating entity 230 to a mapping entity 240. As already described, the generating entity 230 and the mapping entity 240 could be implemented in a single logical, or physical unit, so that this transmission could take place inside such single unit. In general, if the generating entity 230 and the mapping entity 240 are implemented as network functions of the network, the transmission can be operated in any known manner for transmission of data within the network. While both the generating entity 230 and the mapping entity 240 can be implemented as part of any known function of the network, including the user equipment, in preferred embodiments they are implemented in the core network.

The method 300 further comprises a step S35 of mapping, at the mapping entity 240, the message to resources of the wireless communications network 200, based on the assistance information. In particular, the message can be mapped to resources of the transmission channel 260.

Mapping of a message to resources can be understood as deciding which and/or in what manner, of the available resources, are to be used for the transmission of the message. Alternatively, or in addition, in case a resource has a plurality of available configurations, mapping of the message to resources can be understood as deciding which, of the available configurations, are to be used for the transmission of the message. For instance, there may be multiple options to use a single resource. For example, available spectrum can be understood as being a single resource, however but the subcarrier spacing within the spectrum results in different performance trade-offs.

In general, the mapping step therefore allows the network 200 to decide how the message can be transferred. Thanks to the encoding of the characteristics of the artificial intelligence model, and/or of the message into the assistance information, which can preferably be encoded by one or more numerical values, the mapping entity 240 can proceed to the mapping based on those numerical values, without requiring any knowledge or understanding of the characteristics of the artificial intelligence model, and/or of the message.

The method 300 can further comprise a step S37 of controlling transmission of the message using the mapping resulting from the mapping step S35. That is, once the message is mapped to one or more resources, the network can control the transmission of the message by using those resources. In the illustrated embodiments, this is performed by the resources manager 250. It will be clear to those skilled in the art, that several existing functions in the 5G network, such as for instance the SMF or the UPF could be used to implement at least part of the functionality of the resources manager 250.

In some embodiments, the resource manager 250 can be implemented in a single logical and/or physical unit with the mapping entity 240. In this case, the mapping resulting from the mapping step S35 can be made immediately available to the resource manager 250. In alternative embodiments, the resource manager 250 and the mapping entity 240 can be implemented independently from each other, as illustrated in the exemplary embodiment of figure 3. In this case it is possible for the method 300 to comprise a further step S36 of transferring the mapping from the mapping entity 240 to the resource manager 250.

In the description above, the assistance information have been described in general terms. In the following, more specific examples will be provided. In some specific implementations, the assistance information can comprise an importance indicator, I, for indicating a contribution of the message to operation of the artificial intelligence model. In this case, the assistance information can be based both on the characteristics of the artificial intelligence model and on the characteristics of the message. It is in fact clear that, given a type of artificial intelligence model, some messages will be more important for its correct operation, while some messages will be less important. Which messages are more important, and which less, for any given type of artificial intelligence model, can be recognized at the generating entity 230 based, for instance, on predetermined information such as a lookup table.

In some preferred embodiments, the importance indicator, I, has a lower value if the contribution is low, and/or a higher value if the contribution is high.

In further specific implementations, the assistance information can alternatively or additionally comprise a robustness indicator, R, for indicating robustness of the artificial intelligence model against radio losses in the message. In other words, given an artificial intelligence model with a number of parameters, for instance with a number of neurons, this parameter indicates how much the accuracy of the model will be impacted, if one or more of those parameter is lost, due to radio channel losses. In this case, the assistance information can be thus based on the characteristics of the artificial intelligence model. As before, the determination of how much radio loss can impact the functioning of a given artificial intelligence model can be recognized at the generating entity 230 based, for instance, on predetermined information such as a lookup table.

In some preferred embodiments, the robustness indicator, R, has a lower value if the robustness is low, and/or a higher value if the robustness is high,

In further specific implementations, the assistance information can alternatively or additionally comprise an experience indicator, E, for indicating a level of training of the artificial intelligence model. This can comprise, for instance, any of number of users having contributed to training of the artificial intelligence model, a number of iterations performed for training of the artificial intelligence model, statistics of a training set used when training the artificial intelligence model. Furthermore, the experience indicator, E, can comprise a number of environments for which a user equipment trained on, such as highway, factory, sub-urban, or different traffic distribution, eMBB, xr, urllc. etc. The determination of how experienced the given artificial intelligence model is can be recognized at the generating entity 230 based, for instance, on a numerical indication of any of the parameters indicated above.

In some preferred embodiments, the experience indicator, E, has a lower value if the number of users is low, and/or a higher value if the number of users is high,

In further specific implementations, the assistance information can alternatively or additionally comprise a generality indicator, G, for indicating a capability of the artificial intelligence model to perform with high robustness against different sets of inputs.

In some preferred embodiments, the generality indicator, G, has a lower value if the capability if low, and/or a higher value if the capability is high.

That is, any of the importance indicator, the robustness indicator, the experience indicator and the generality indicator can be expressed numerically by the generating entity 230. In this manner, the generating entity can numerically indicate to the mapping entity how the message should be mapped in a manner which is normalized across a plurality of types of messages and of artificial intelligence models. This allows the mapping entity 240 to then be able to allocate resources accordingly.

In some embodiments, the mapping step can be executed based on a function which has any of the importance indicator, the robustness indicator, the experience indicator and the generality indicator as input. That is, resources can be allocated as a function of any of the indicators. For instance, the importance indicator can be used to decide on the allocation of a first type of resources, and/or the robustness indicator can be used to decide on the allocation of a second type of resources, and/or the experience indicator can be used to decide on the allocation of a third type of resources, and/or the generality indicator can be used to decide on the allocation of a fourth type of resources.

Alternatively, or in addition, any of the importance indicator, the robustness indicator, the experience indicator and the generality indicator can be combined in a priority value, and the mapping step can be executed based on the priority value. In preferred embodiments, the priority value can be obtained by a function having the indicators as input. Preferably, the function is a monotonic function configured to increase if at least one of the input values increases, and/or configured to decrease if at least one of the input values decreases. In some further referred embodiment the function is a sum, preferably weighted, of one or more of the indicators, preferably all. In those embodiments where the mapping step S35 is based on the priority value, resources of the wireless communications network 200 are allocated to the message as a function of the priority value. That is, resources of the wireless communications network 200 are allocated so that a priority value results in a higher priority of use of resources by the message.

However, as previously indicated, in some embodiments the mapping can be executed based on the value of one or more indicators, not necessarily combined in the priority value.

For instance, in a specific exemplary implementation, the mapping step S35 can be configured to increase a logical channel priority, LCP, of the message, preferably to a level above “MAC CE for BSR” and/or below “C-RNTI of CCCH”, if the importance indicator, I, has a value higher than a predetermined threshold. This reflects that the artificial intelligence message is contributing to a higher valued procedure for the connectivity, and hence it should be given higher priority due to the higher importance. For example, if the Al-Message is related to AMF service that enable HP NF making sure that HP UE is reachable, it will be more important that LP NF requesting LP service. In other words, if the importance indicator, I, is higher, then LCP can increase by one level, which can advantageously increase the possibility that AI-MSG will make it in the current MAC PDU.

Alternatively, or in addition, in a specific exemplary implementation, the mapping step S35 can be configured to increase MAC Bucket Size, preferably via Prioritized Bit Rate and/or Bucket Size Duration, if the importance indicator, I, and/or the experience indicator, E, has a value higher than a predetermined threshold. That is, if the importance indicator, I, or the experience indicator, E, is higher, then PBR can be increased, so that more bits of the message will advantageously fit in a single PDU.

Alternatively, or in addition, in a specific exemplary implementation, the mapping step S35 can be configured to increase CSI-RS occasions for multiple UEs or for specific group of UEs, if the robustness indicator, R, and/or the experience indicator, E, and/or the generality indicator, G, has a value higher than a predetermined threshold. If any of R or E or G is higher, then more CSI-RS signals, in time and/or frequency, can be used to estimate the CSI for a given specific UE, or group of UE. This advantageously improves the accuracy of CSI, hence better scheduling.

Alternatively, or in addition, in a specific exemplary implementation, the mapping step S35 can be configured to allocate broadcast channel to a group of UEs to enable generality via federated learning. In particular, the artificial intelligence model can be broadcasted to the group of UEs, if the experience indicator E and/or the generality indicator G has a value higher than a predetermined threshold. This is particularly advantageous since, if E is higher, then a specific group of UEs that have this experience can be used for federated learning aggregation and local training. Similarly, if G is higher, then larger group of UEs that have this experience can be used for federated learning aggregation and local training. Both approaches advantageously help improving the message.

Alternatively, or in addition, in a specific exemplary implementation, the mapping step S35 can be configured to decrease a BLER target of a Link adaptation, or MCS selection, if the importance indicator, I, and/or the robustness indicator, R, has a value higher than a predetermined threshold. This is particularly advantageous since, if R, or I, is higher, then lower MCS and BLER can be achieved, this a better reception rate.

All of the above parameters will be clear to those skilled in the art in the field of 5G networks. Exemplary definitions can be found, for instance, in TS 38.321 .

In a further exemplary implementation, the artificial intelligence model can be configured to exchange RAN-AI, that is, RAN Artificial Intelligence, messages with one or more UE(s), for exchanging data between the artificial intelligence model and the one or more UE(s). In particular, the artificial intelligence model can be configured to output RAN-AI actuation messages, for controlling Radio Resource Management function, preferably for MAC and PHY layers

As illustrated in figure 3A, in some further embodiments, the method 300 can further comprise a step S38 of transmitting, from the service consumer 220 to the mapping entity 240, a mapping configuration. The mapping configuration can in particular include information which allows defining how the mapping of the resources is performed on the basis of the assistance information previously described. That is, instead of having a predetermined configuration deriving the mapping from the assistance information, such configuration can be provided by the service consumer 220. In such embodiments, the step S35 of mapping is based on the assistance information and on the mapping configuration. This can be particularly advantageous in case information available at the service consumer 220 could be used to change the mapping of the message.

As illustrated in figure 3B, in some further embodiments, the method 300 can further comprise a step S39 of transmitting, from the service consumer 220 to the mapping entity 240, a correction value. The method can further comprise a step S40 of changing, at the mapping entity 240, any of the importance indicator, I, the robustness indicator, R, the experience indicator, E, or the generality indicator, G, by an amount corresponding to the correction value. Similarly as above, this can be particularly advantageous in case information available at the service consumer 220 could be used to change the mapping of the message.

As illustrated in figure 3C, in some further embodiments, the method 300 can further comprise a step S41 of obtaining, at the generating entity 230, a generating policy for configuring the way in which the generating entity 230 performs the generating step. As illustrated, the generating policy can be provided by the PCF 13. In preferred embodiments, the generating policy could configure the way in which the characteristics of the artificial intelligence model, and/or of the message are encoded into the assistance information.

As illustrated in figure 3D, in some further embodiments, the method 300 can further comprise a step S42 of obtaining, at the mapping entity 240, a mapping policy for configuring the way in which the mapping entity 240 performs the mapping step. As illustrated, the mapping policy can be provided by the PCF 13. In preferred embodiments, the mapping policy could configure the way in which the mapping of the resources is performed, based on the value of the assistance information.

This could be advantageous as different policies on how to prioritize messages related to artificial intelligence might be different at different segments of the network 200. Thus, a message that was mapped to a higher priority in a first segment might be mapped to a lower priority in a second segment. This could be the case, for instance, because in a network segment a certain Al functionality could be more or less important. As an example, a message related to an handover optimization implemented by Al could be more important in a network segment with high mobility, like on a highway, whereas it could less important in a segment such as a private network for indoor automation. Alternatively, or in addition, it could be because the set of slices that the Al functionality is applied to has changed. The flexibility to modify the mapping therefore advantageously allows those cases to be targeted, by providing at least two different mapping policies to two different segments of the network 200.

Embodiments above have been described with reference to method 300 as comprising a plurality of entities and steps. It will be clear that not all those entities and/or steps need to be implemented for the invention to be carried out, and/or that embodiments of the invention can also relate to one or more of the entities carrying out only one or more of the described steps. In particular, as visible in figure 4, a method 400 can be implemented for operating the generating entity 230 in a wireless communications network 200. The wireless communications network 200 can, as previously described, comprise the service provider 210 implementing an artificial intelligence model for providing services to at least one service consumer 220. It will be clear that, in addition thereto, the wireless communications network 200 can comprise any of the entities previously described, and in particular the generating entity.

As visible in figure 4, the method 400 can comprise a step S32 of obtaining characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model, a step S33 of generating assistance information related to characteristics of the artificial intelligence model and/or characteristics of the message, and a step S34 of transmitting to the mapping entity 240, the assistance information. Any of the considerations previously made for any of those steps in the description of method 300 can also apply to method 400.

Similarly, as visible in figure 5, a method 500 can be implemented for operating the mapping entity 240 in the wireless communications network 200. The considerations previously made for wireless communications network 200 and method 400 also apply for method 500. In this case, the wireless communications network 200 can in particular comprise the mapping entity 240.

As visible in figure 5, the method 500 can comprise a step S34 of obtaining assistance information related to characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model, and a step S35 of mapping the message to resources of the wireless communications network (200), based on the assistance information. Any of the considerations previously made for any of those steps in the description of method 300 can also apply to method 500.

While the invention has been described in terms of method steps, it will be clear that embodiments of the invention can relate also to entities, or devices. In general, it is clear that a generic computing device can comprise a memory and a microprocessor, wherein the memory can be configured to cause the microprocessor to carry out any of the steps previously described.

An embodiment can in particular refer to the generating entity 230. In particular, the generating entity 230 can, as illustrated in figure 6, comprise a processor 230-1 and a memory 230-3, the memory comprising instructions configured to, when executed by the processor 230-1 , cause the processor to carry out one or more steps. In general, the memory and processor can be configured to carry out any of the steps previously described in association with the generating entity 230. In some embodiments those can in particular comprise steps S32, S33 and S34. Thus, the steps can comprise a step S32, of obtaining characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model. The steps can further comprise a step S33 of generating assistance information related to characteristics of the artificial intelligence model and/or characteristics of the message. The steps can also comprise a step S34 of transmitting to the mapping entity 240, the assistance information.

Another embodiment can refer to the mapping entity 240. In particular, as visible in figure 7, the mapping entity 240 can comprise a processor 240-1 and a memory 240-3, the memory comprising instructions configured to, when executed by the processor 240-1 , cause the processor to carry out any of the steps previously described in association with the mapping entity 240. In some embodiments those can in particular comprise steps S34 and S35. Thus, the steps can comprise a step S34 of obtaining assistance information related to characteristics of the artificial intelligence model and/or characteristics of a message related to the artificial intelligence model. Additionally, the steps can comprise a step S35 of mapping the message to resources of the wireless communications network 200, based on the assistance information.

Both the generating entity 230 and the mapping entity 240 can furthermore comprise input/output means, respectively 230-2 and 240-2, for sending or receiving data with other entities in the network 200 and/or network functions, as well as entities outside of the network 200.

It has thus been described how, when an entity in the wireless telecommunications network 200 implements an artificial intelligence model, messages to and/from this entity, or more generally messages related to, or originated from, or directed to, the artificial intelligence model can be mapped to resources in a manner which takes into account the characteristics of the artificial intelligence model and/or of the message. This allows an efficient mapping of those messages to the resources of the wireless telecommunications network 200, which would not be possible with the static mapping of the prior art.

Various embodiments have been described, each with a set of features. It is understood this has been done for explanatory purposes and not to limit the invention to the specific embodiments and/or to the specific combination of features described. Further embodiments of the invention can be obtained by combining any of the features previously described, even if described a belonging to different embodiments.