EP4659526A1 - Data-driven access point coordination in uplink and downlink coexistence - Google Patents

Abstract

A computer-implemented method (200) for enabling access point assignment in a wireless communications network (100) where multiple access points (110) simultaneously serve multiple wireless devices (121). The method comprises obtaining (210) an uplink-downlink schedule in which each wireless device of the multiple wireless devices (121) is scheduled for either uplink or downlink; monitoring (220) link characteristics between the multiple access points (110) and the multiple wireless devices (121) over time; and when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determining (250) an uplink or downlink assignment of one or more access points of the multiple access points (110), respectively, based on the obtained uplink-downlink schedule and on the monitored link characteristics using a machine learning model, wherein the machine learning model is trained to provide an uplink or downlink assignment of one or more access points, respectively, based on uplink-downlink schedules and on link characteristics.

Description

DATA-DRIVEN ACCESS POINT COORDINATION IN UPLINK AND DOWNLINK COEXISTENCE TECHNICAL FIELD The present disclosure relates generally to the field of wireless communication. More particularly, it relates to a method for uplink and downlink assignments for access points (APs) in a wireless communications network where multiple APs simultaneously serve multiple wireless devices; such wireless communications network is sometimes referred to as a distributed multiple-input and multiple-output (D-MIMO) network. BACKGROUND In wireless communications, multiple-input and multiple-output (MIMO) is a technology utilizing multiple transmission antennas and receiving antennas to increase capacity. Distributed MIMO (D-MIMO), also known as cell-free MIMO, Radio Stripes etc., is a key technology candidate for the sixth generation (6G) physical layer in wireless communications. The basic idea in D- MIMO is to distribute antennas geographically and have them operate together. A typical architecture is that multiple distributed antenna units (also known as access points, APs) are interconnected and configured in such a way that they may cooperate phase-coherently. Each AP in turn may comprise multiple antenna elements that are also configured to operate phase- coherently together, so that all antennas of all APs together effectively form a large, coherently operating antenna array. The link resulting from such spatial distribution of APs typically yields higher spatial degrees-of-freedom (compared to non-distributed, i.e., co-located, MIMO setups). Such degrees-of-freedom may be exploited in a number of ways, e.g., they may be exploited to ensure fairer signal to noise ratio (SNR) distributions among user equipments (UEs), larger number spatially multiplexed data streams, and larger link robustness. It is envisioned that future D-MIMO deployments will provide an excess number of distributed APs (i.e. many more APs than active UEs). Each distributed AP is further expected to operate with half-duplex technology since full-duplex technology is still relatively immature and expensive to realize. Nevertheless, an excess number of half-duplex APs may effectively be utilized to realize a network yielding some of the advantages of full-duplex-like communications (without using a real full-duplex hardware). Such advantages are e.g. communications with lower latency and potential higher spectrum efficiency. To elaborate, suppose there are a large number of half-duplex APs distributed over a service area and each AP operates either in downlink (DL) or uplink (UL) mode at the same time and frequency according to the current needs. With that, some APs will be receiving in UL and, at the same time and frequency, other APs will be transmitting in DL. From a system-level point of view, the network may be thought as performing full-duplex communications. Such a way of operating a D-MIMO network, and additional challenges resulting from it, has some resemblance with dynamic time division duplex (DTDD), which has been discussed and standardized over the last decade in the context of traditional cellular systems. Here, UL and DL transmissions are dynamically switched depending on the UL/DL demand at each cell. A drawback resulting from such switching flexibility is that of additional inter-cell cross-link interference. Such interference comprise DL-to-UL interference between two base stations (e.g., gNB-to-gNB interference), and UL-to-DL interference between two UEs (i.e. UE-to-UE interference). In traditional cellular systems, there is no central coordination and no phase coherent joint signal processing between gNBs operating at different locations. A similar type of cross-link interference will occur in a D-MIMO network that operates as described above, except that it will be “intra-cell” cross-link interference in the form of AP-to- AP interference and UE-to-UE interference. In this case, cell means the service area for the D-MIMO network. To summarize, a D-MIMO network combined with DTDD concept offers a more flexible duplex option that is able to harvest full duplex-like benefits in the communication performance. In the literature, a few approaches for AP assignment, i.e., assigning each AP in the wireless communications network to operate either in UL or DL, have been suggested. Shuto Fukue et al. propose an optimization-based approach that jointly design AP assignment and beamforming weights by means of non-convex optimization techniques in “Joint Access Configuration and Beamforming for Cell-Free Massive MIMO Systems With Dynamic TDD”, IEEE Access, vol.10, Apr.2022. Yue Zhu et al. suggest using a specific type of convex relaxation methods (i.e., parallel convex relaxation) to improve the spectrum efficiency by designing AP assignment in “Optimization of Duplex Mode Selection for Network-Assisted Full-Duplex Cell-Free Massive MIMO Systems”, IEEE Commun. Lett., vol.25, no.11, Nov.2021. Jiamin Li et al. use a random AP assignment method in the context of channel estimation in DTDD D-MIMO in “Network-Assisted Full-Duplex Distributed Massive MIMO Systems With Beamforming Training Based CSI Estimation”, IEEE Trans. Wireless Commun., vol.20, no.4, Apr.2021. Dongming Wang et al. propose a user scheduling method to avoid the aforementioned cross interference in “Performance of Network-Assisted Full-Duplex for Cell-Free Massive MIMO”, IEEE Trans. Commun., vol.68, no.3, Mar.2020. However, there is a need for improved ways of implementing AP assignment. SUMMARY It is an object of the present disclosure to mitigate, alleviate or eliminate one or more of the above-identified deficiencies and disadvantages in the prior art and solve at least the above- mentioned problem. In particular, an object is to provide improved wireless communications networks where multiple access points (APs) simultaneously serve multiple wireless devices. This object is obtained at least in part by a computer-implemented method for enabling AP assignment in a wireless communications network where multiple APs simultaneously serve multiple wireless devices. The computer-implemented method comprises obtaining an uplink- downlink schedule in which each wireless device of the multiple wireless devices is scheduled for either uplink (UL) or downlink (DL). The computer-implemented method also comprises monitoring link characteristics between the multiple APs and the multiple wireless devices over time. The computer-implemented method further comprises, when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determining an UL or DL assignment of one or more APs of the multiple APs, respectively, based on the obtained uplink-downlink schedule and on the monitored link characteristics using a machine learning model (ML), wherein the ML model is trained to provide an UL or DL assignment of one or more APs, respectively, based on uplink-downlink schedules and on link characteristics. The disclosed computer-implemented method enables a wireless communications network with low communication fronthaul overhead, low computational cost, and low latency. The disclosed method may determine UL or DL assignment for the APs only when it is deemed likely to have an effect of the overall performance of the wireless communications network. Consequently, AP assignments may be kept when the channel statistics remain relatively unchanged, i.e., when the condition that at least one link characteristic changes more than the predetermined amount is not met. The monitored link characteristics over time may represent long-term channel statistics. The disclosed method may only determine UL or DL assignments for the APs once per time it takes for the long-term channel statistics involving active wireless devices and APs to change. A change in a link characteristic may occur if e.g. a wireless device moves from one position to another. After such change, the path gain of that wireless device to the APs will likely change. In that case, an update of AP assignments may lead to improved performance of the wireless communications network, such as improved overall total data capacity in the network. Consequently, it is desired in that case to determine an UL or DL assignment. Furthermore, when the wireless device moves from the first position to the second position, it is likely that the inter-cell interference will change. Such movement thus also may motivate an update in AP assignment. Link characteristics may also change for other reasons. Using an already trained ML model during an online deployment of the wireless communications network enables an UL or DL assignment of the APs with little processing cost and time compared to searching for an optimal assignment online. Depending on how the ML model is trained, the disclosed method may also determine AP assignments enabling the wireless communications network to present similar performance (e.g. measured in terms of throughput) as an exhaustive search for an optimal set of AP assignments. According to some aspects, the ML model has been trained using supervised learning. If a desired set of AP assignments are known for a given scenario or state of a particular wireless communications network, the ML model may advantageously be trained using supervised learning to be able to assign APs for other similar scenarios and for similar wireless communications networks. Here, a state in the particular wireless communications network may be parameterized in terms of the on link characteristics between multiple APs and multiple wireless devices in the particular wireless communications network, and in terms of an uplink- downlink schedule. In particular, the ML model may have been trained using a training set comprising one or more input-output pairs. An input-output pair comprises an uplink-downlink schedule and link characteristics as input, and UL or DL assignment of one or more APs, respectively, as output. In other words, in an input-output pair, a particular state of the particular wireless communications network, parametrized as the link characteristics and as the uplink- downlink schedule, is paired with a desired output, namely, a desired UL or DL assignment of one or more APs, respectively, for that particular state of the particular wireless communications network. According to some aspects, the one or more input-output pairs have been obtained using an exhaustive search. In other words, a desired UL or DL assignment of one or more APs, respectively, for a particular state of a particular wireless communications network may be found by evaluation of all possible assignments. The evaluation may e.g. comprise of finding the assignments that result in the highest possible throughput of the particular wireless communications network. The exhaustive search is a separate step from training the ML model. The ML model may be subsequently trained using obtained one or more input-output pairs. This way, the ML model is likely to assign APs such the particular wireless communications network (or any similar wireless communications network) present similar performance (e.g. measured in terms of throughput) as the exhaustive search for an optimal set of AP assignments. Performing an exhaustive search may be very time consuming. Therefore, the ML model may alternatively have been trained using reinforcement learning. This way, the trained ML model may perhaps not find as good of an assignment as when the ML model had been trained using supervised learning. However, reinforcement learning does not require any a priori knowledge of a desired assignment. Consequently, reinforcement learning is typically less time consuming compared to supervised learning. When using reinforcement learning, the ML model may have been trained using a predetermined utility in terms of link quality and using a training set comprising one or more input sets. Here, an input set comprises an uplink-downlink schedule and link characteristics and an uplink-downlink schedule. In other words, the ML model is directly trained to provide a desired UL or DL assignment of one or more APs, respectively, for a particular state of a particular wireless communications network. Here, the particular state of the particular wireless communications network is parameterized as the link characteristics and as the uplink- downlink schedule. Similarly, the predetermined utility can be seen as a parameterization of the desired UL or DL assignment. The predetermined utility can be seen as a reward function for reinforcement learning. The predetermined utility may comprises any of path gain, received power, transmitted power, signal-to-noise ratio, SNR, SINR, spatial channel correlation, channel rank and data capacity. Any such predetermined utility may be used to effectively train the ML model to assign the APs to achieve high performance in the wireless communications network comprising the APs, e.g. in terms of data capacity. In particular, the predetermined utility may comprise information of interference. This interference may be between APs, between wireless devices, and/or between APs and wireless devices. This may help in training the ML model to an assignment that is close to a global optimum solution, in terms of e.g. data capacity, while being computationally efficient. According to some aspects, the computer-implemented method further comprises determining whether the ML model should be retrained based on the monitored link characteristics, and retraining the ML model based on the monitored link characteristics when it is determined to retrain the ML model. This way, the ML model may be adapted if e.g. deployment scenario of the wireless communications network changes during online deployment. Consequently, the retrained ML model may result in an improved performance of the wireless communications network during online deployment. The retraining may also be based on an original training set used for an initial training of the ML model. When basing the retraining of the ML model on the monitored link characteristics, the retrained ML model may advantageously take any changes in the deployment scenario into account. According to some aspects, the method comprises determining that the ML model is to be retrained when a performance metric of the wireless communications network based on the monitored link characteristics is below a first threshold value. The performance metric may e.g. be throughput. In that case, when the throughput is below first threshold value, retraining of the ML model is triggered. The performance metric may also comprise information such as signal-to-interference-plus-noise ratio, SINR. According to some aspects, the ML model is further trained to output an expected performance metric for any given DL or UL assignment outputted by the ML model. In that case, the method computer-implemented may comprise determining that the ML model is to be retrained when a performance metric of the wireless communications network based on the monitored link characteristics differs from the expected performance metric more than a second threshold value. This provides a convenient way triggering retraining if deemed necessary. The expected performance metric comprises the same type of data as the performance metric of the wireless communications network based on the monitored link characteristics. The expected performance metric may comprise any of signal-to-interference-plus-noise ratio, SINR, and throughput. According to some aspects, the monitored link characteristics comprises one or more path losses between different APs of the multiple APs, and/or between any AP of the multiple APs and any wireless device of the multiple wireless devices, and/or between different wireless devices of the multiple wireless devices. In that case, the computer-implemented method may comprise determining that the ML model is to be retrained when one or more of the one or more path losses have changed more than a third threshold value in a first time interval. This provides a convenient way of monitoring how the ML model is performing during online deployment, and triggering retraining if deemed necessary. For example, an investigation if one or more of the one or more path losses have changed more than a third threshold value may occur every day. More generally, the second time interval may be 1-10 days. However, the second time interval may be any number. The second time interval may be static or dynamic. For example, the predetermined time interval may be based on past link characteristics of the monitored link characteristics. This may additionally reduce computational overhead and latency on average since the monitoring is not done when it is likely to be unnecessary, i.e., when nothing is likely to have changed in the network. According to some aspects, the computer-implemented method further comprises assigning the one or more access points of the multiple access points to operate in either uplink or downlink based on the determined uplink or downlink assignment of the one or more access points of the multiple access points. For example, if the method is performed by a data processing unit, the data processing unit may communicate the determined UL or DL assignment to each corresponding AP, via e.g. high-capacity backhaul links. According to some aspects, the monitored link characteristics comprise any of path gain, received power, transmitted power, signal-to-noise ratio, SNR, signal-to-interference-plus- noise ratio, SINR, spatial channel correlation, and channel rank. Other link characteristics are also possible. According to some aspects, the computer-implemented method further comprises monitoring whether the at least one link characteristic of the monitored link characteristics changes more than the predetermined amount in a second time interval. This may further reduce computational overhead and latency on average. For example, an investigation if the at least one link characteristic has changed may occur every 1 seconds. More generally, however, the second time interval may be 0.1-10 seconds. However, the predetermined time interval may be any number. The second time interval may be static or dynamic. For example, the predetermined time interval may be based on past link characteristics of the monitored link characteristics. This may additionally reduce computational overhead and latency on average since the monitoring is not done when it is likely to be unnecessary, i.e., when nothing is likely to have changed in the network. In an example scenario, a wireless device has been stationary for 10 minutes. In that case, the predetermined time interval is currently 10 seconds. Then, the wireless device moves, which changes the link characteristics. The predetermined time interval is then dynamically changed to 1 second. This is an example of a reactive change of the predetermined time interval. The predetermined time interval may alternatively, or in combination of, be dynamically changed proactively. This may additionally reduce computational overhead and latency on average for similar reasons. The predetermined time interval may e.g. be based on an expected time the respective wireless devices are expected to move less than a predetermined distance. There is also disclosed herein a node for access point (AP) assignment in a wireless communications network where multiple APs simultaneously serve multiple wireless devices. The node is associated with the above-discussed advantages. The node comprises a processing circuitry and a memory. The processing circuitry is configured to obtain an uplink- downlink schedule in which each wireless device of the multiple wireless devices is scheduled for either uplink (UL) or downlink (DL). The processing circuitry is further configured to monitor link characteristics between the multiple APs and the multiple wireless devices over time. The processing circuitry is also configured to, when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determine an UL or DL assignment of one or more APs of the multiple APs, respectively, based on the obtained uplink- downlink schedule and on the monitored link characteristics using a machine learning model (ML), wherein the ML model is trained to provide an UL or DL assignment of one or more APs, respectively, based on uplink-downlink schedules and on link characteristics. There is also disclosed herein a computer program product comprising instructions which, when executed on at least one processing circuitry, cause the at least one processing circuitry to carry out the method according to the discussion above. The computer program is associated with the above-discussed advantages. There is also disclosed herein a computer program carrier carrying a computer program product according to the discussion above, wherein the computer program carrier is one of an electronic signal, optical signal, radio signal, or computer-readable storage medium. The computer program carrier is associated with the above-discussed advantages. BRIEF DESCRIPTION OF THE DRAWINGS With reference to the appended drawings, below follows a more detailed description of embodiments of the present disclosure cited as examples. In the drawings: Figure 1 is schematic illustrations of a wireless communications networks; Figures 2-4 are flow charts illustrating methods; Figure 5 shows spectrum efficiency for various access point assignment methods; and Figure 6 schematically illustrates an access point or a data processing unit. DETAILED DESCRIPTION The present disclosure is described below with reference to the accompanying drawings, in which certain aspects of the present disclosure are shown. The present disclosure may, however, be embodied in many different forms and should not be construed as limited to the embodiments and aspects set forth herein; rather, these embodiments are provided by way of example so that this disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art. Like numbers refer to like elements throughout the description. It is to be understood that the present disclosure is not limited to the embodiments described herein and illustrated in the drawings; rather, the skilled person will recognize that many changes and modifications may be made within the scope of the appended claims. Figure 1 depicts a wireless communications network 100 in which embodiments herein may operate. In some embodiments, the wireless communications network 100 may be a radio communications network, such as, 6G, NR or NR+ telecommunications network. However, the wireless communications network 100 may also employ technology of any one of 3/4/5G, LTE, LTE-Advanced, WCDMA, GSM/EDGE, WiMax, UMB, GSM, or any other similar network or system. The wireless communications network 100 may also employ technology transmitting on millimeter-waves (mmW), such as, e.g. an Ultra Dense Network, UDN. In some embodiments, the wireless communications network 100 may also employ transmissions supporting WiFi transmissions, e.g. the wireless communications standard IEEE 802.11ad or similar, or other non-cellular wireless transmissions. In particular, the wireless communications network in Figure 1 is a distributed multiple-input and multiple-output (D-MIMO) network, where access points (APs) 110 are distributed, i.e., geographically spread out over an area, in a planned or random fashion. The APs are arranged in communication with a data processing unit 140 through e.g. high-capacity backhaul links (such as fiber optic cables). The data processing unit 140 may e.g. be a remote standalone server, a cloud-implemented server, a distributed server, dedicated data processing resources in a server farm, or similar. The APs 110 may serve wireless devices 121 in at least coverage area 115, which may also be called a cell. The APs 110 may correspond to any type of network node or radio network node capable of communicating with a wireless device and/or with another network node, such as, a base station (BS), a radio base station, gNB, eNB, eNodeB, a Home NodeB, a Home eNodeB, a femto Base Station (BS), or a pico BS in the wireless communications network 100. As is also shown in Figure 1, a plurality of wireless devices 121 are located within the cell 115. Each wireless device 121 in the plurality of wireless devices is configured to communicate within the wireless communications network 100 via the APs 110 over radio links served by the APs 110. The wireless devices 121 may transmit data over an air or radio interface to one or more APs 110 in uplink (UL) transmissions and the APs 110 may transmit data over an air or radio interface to one or more wireless devices 121 in downlink (DL) transmissions. The wireless devices 121 may refer to any type of wireless devices or user equipment (UE) communicating with a network node and/or with another wireless device in a cellular, mobile or radio communication network or system. Examples of such wireless devices are mobile phones, cellular phones, Personal Digital Assistants (PDAs), smart phones, tablets, sensors equipped with a UE, Laptop Mounted Equipment (LME) (e.g. USB), Laptop Embedded Equipment (LEE), Machine Type Communication (MTC) devices, or Machine to Machine (M2M) device, Customer Premises Equipment (CPE), target device, device-to-device (D2D) wireless device, wireless device capable of machine to machine (M2M) communication. Each distributed AP in the example wireless communications network 100 in Figure 1 is operating with half-duplex. The circles 130 represent wireless devices and APs operating in UL. The remainder of wireless devices and APs are operating in DL. As previously noted, dynamic time division duplex (DTDD) D-MIMO operation may result in a cross-link interference problem similar to that existing in between cells in conventional cellular communications networks when each cell performs DTDD communications. In the conventional cellular DTDD, each cell may dynamically change its duplex mode depending on the demand from users located within the cell, resulting in the fact that DL base stations (e.g., Next Generation Node B, gNB) cause interference towards neighboring UL base stations, and UL users cause interference towards DL users in different neighboring cells. Also as previously noted, a similar type of (intra-cell) cross-interference may occur in DTDD D-MIMO, e.g. AP-to-AP interference. However, it should be noted that there is typically more flexibility to address such interference in D-MIMO systems since it is much easier to coordinate APs in the same cell, compared with coordinating base stations in different cells. As part of the developing of the embodiments described herein, it has been realized that there is a need for improved ways of assigning APs to operate in either UL or DL in a D-MIMO system. In DTDD D-MIMO, a suitable set of APs should be assigned to operate in DL and another suitable set of APs should be assigned to operate in UL. This is because the configurations of the APs directly affects how each AP interferes with each other. Such interference between APs should be kept under control. We hereafter refer to this problem as an AP assignment problem, or UL/DL AP assignment problem. AP assignment may be done at a given point in time based on the instantaneous channel state information (CSI). However, such instantaneous CSI-based assignments may pose the following issues: • high computational cost due to the fact that AP UL/DL configuration needs to be updated every time the instantaneous CSI changes (i.e. once per channel coherent time), • high communication overhead on fronthaul links between APs and their central computing unit, and • high latency due to multiple re-configurations of APs based on short-term CSI. In contrast to such approaches, where the UL/DL AP assignment needs to be updated once per channel coherence time (e.g. once per new radio, NR, slot in the worst case), the disclosed approach may only need an AP assignment update once per time it takes for the long-term channel statistics to change. This results in a system yielding benefits similar to those of full- duplex communications, but with 1) lower communication fronthaul overhead, 2) the computational cost, and 3) latency. The current disclosure addresses the UL/DL AP assignment problem by disclosing data-driven AP assignment methodologies that overcome the aforementioned bottlenecks by means of a machine learning (ML) model, which is also called an artificial intelligence (AI) framework. Examples embodiments of the disclosed methodologies provide different ML approaches with varying degrees of required data and required processing complexity that solve the UL/DL AP assignment problem. The disclosed methodologies may perform AP assignment for wireless communications network where multiple APs simultaneously serve multiple wireless devices, and where each AP of the multiple APs and each wireless device of the multiple wireless devices perform half- duplex communications. The AP assignment is based on a ML model and current channels link statistics (such as path loss), e.g. for each pair of APs and wireless devices, between different APs, and/or between different wireless devices in the wireless communications network. An ML model may also be called an intelligent agent (IA). A selected ML model is trained based on a training data set. The training data set may comprise a desired AP assignment for a given channel condition, e.g. by means of an offline exhaustive search or by means of some target performance metric (such as throughput). After training, assignment of one or more APs in the wireless communications network to either be an UL AP or DL AP is based on the trained ML model. Training of the certain ML model (i.e., before online transmission) may be done offline and may advantageously be used in the disclosed methodologies. If the ML model is trained specifically for the online wireless communications network, the deployment scenario of the online wireless communications network should preferably not change more than a tolerable amount. In other words, some changes in the deployment scenario may be acceptable. The limit of acceptable changes may e.g. be determined based on a performance metric (such as throughput) of the online wireless communications network. An example of a change of the deployment scenario is when a new building is built in the service area 115. In that case, propagation paths in the wireless communications network may change and the ML model may not provide an AP assignment resulting in a desired performance of the online wireless communications network. Changes in the propagation paths may be indicated from changes in e.g. path gains between APs and wireless devices, between different APs, and or between different wireless devices. However, it also possible to train the ML model to be more generic such that it can take changes of the deployment scenario into account. It is also possible to train the ML model to be able to provide AP assignments for different wireless communications networks. An initial training or a retraining of the ML model may be done in different ways. In one example, supervised training is used. In this case, a desired AP assignment is available for a given set of training data. This desired AP assignment may be obtained in different ways, such as an exhaustive search. In another example, reinforcement learning is used. In this case, the ML model is trained based on a reward. This reward may also be called a utility. The utility may e.g. be based on a sum of estimated signal-to-interference-plus-noise-ratios (SINRs) at each wireless device. The ML model may be trained to maximize the utility. As mentioned, the ML model may be trained offline. In that case, the processing cost and time for AP assignment during online transmission in the wireless communications network is significantly smaller than other assignment methods, such as an exhaustive search of a desired AP assignment performed when the wireless communications network is online. Retraining of the ML model may be done offline or online. Moreover, as is shown further below, simulations show that the performance of the embodiments of the disclosed methodologies are also comparable to an ideal case, where an exhaustive search with instantaneous CSI knowledge is performed during online deployment. In other words, the embodiments of the disclosed methodologies herein may present almost as good performance as a global optimum solution for AP assignment, while requiring significantly less computational resources in terms of computational cost and latency. The system shown in Figure 1 is a DTDD D-MIMO system serving multiple UL/DL UEs at the same time. The disclosed methodologies herein are suitable for such DTDD D-MIMO system. In general, however, the disclosed methodologies are suitable for AP assignment in a wireless communications network where multiple APs 110 simultaneously serve multiple wireless devices 121. Although the disclosed methodologies are suitable for any types of wireless devices, UEs are used as an example in embodiments disclosed herein. Below follows a system model of an example DTDD D-MIMO system. For simplicity, is it assumed that each AP is equipped with a number ^^ antennas and each UE is equipped with a number ^^ antennas, in which all the radio devices are assumed to operate in half-duplex mode. However, the teachings of this disclosure is also applicable to the more general case where the number of antennas vary per UE and/or per AP. Furthermore, it is assumed that the network knows the current UL/DL demand from all UEs (i.e., which UE would like to operate in UL/DL). In other words, the data processing unit 140 and/or APs may be configured to obtain UL/DL demands from UEs in the wireless communications network 100. A general objective in any AP assignment method may be to find an optimal set of APs serving in the UL and DL, respectively. Preferably, a global optimum is found. However, other assignments than the global optimum (such as local optimums) may be more desirable if they may be found with less computational power without sacrificing much overall performance. Note that since each AP is assumed to only be capable of half-duplex communications, an AP assigned to UL group may not transmit in the DL simultaneously, and vice versa. For the sake of later convenience, let ℒ _^^ and ℒ _^^ be the set of APs operating in UL and DL, respectively. Since the wireless communications network 100 serves UL and DL at the same time and frequency, the system may suffer from cross-link interference, i.e., interference from DL APs to UL APs, and interference from UL UEs to DL UEs. To elaborate, denoting ^^ _^^ as an UL UE index with ^^ _^^ being the set of UL UEs, the ^^ _^^-th UL signal ( ^^̂ _{^^ ^^}) at the central computing unit, after collecting the received signal from each UL AP, may be written as where ^^^H ^_{^ ^^} ∈ ℂ^{1× ^^} is the combining vector at the ^^-th AP with ^^ ∈ ℒ _^^ to detect signal from the ^^ _^^-th UL UE; ^^ _{^^ ^^} ∈ ℂ ^{^^× ^^} denotes the channel matrix between ^^-th UE and ^^-th AP; ^^ _{^^ ^^} ∈ ℂ ^{^^×1} is the precoding vector at ^^ _^^-th UL UE; ^^̈ _{^^ ^^} ^′ ∈ ℂ ^{^^× ^^} is the cross-AP channel matrix between ^^- th UL AP and ^^^′-th denotes the precoding vector at ^^^′-th DL AP towards ^^^′-th DL UE; ^^ _^^ ∈ ℂ^{1× ^^} is the noise vector at ^^-th UL AP. Similarly, the received signal ( ^^ _{^^ ^^}) at ^^ _^^-th DL UE may be written as where ^^^H _{^^ ^^} ∈ ℂ^{1× ^^} is the combining vector at ^^ _^^-th DL UE, ^^ _{^^ ^^} ∈ ℂ^{1× ^^} is the noise vector at ^^ _^^- th DL UE, the channel UL/DL reciprocity is assumed, and ^^̇ _{^^ ^^ ^^} ^′ ∈ ℂ ^{^^× ^^} is the interference channel matrix from the ^^^′-th UL UE to ^^ _^^-th DL UE. Below follows a description of supervised learning that may be used to train the ML model used in the disclosed methodologies. Supervised learning is a ML framework that learns to map an input to an output based on given input-output examples. An input-output example may also be called an input-output pair or labelled data. For example, an input of an input-output may comprise a matrix of path gains between APs and UEs (e.g., between APs and UEs, between different UEs, and/or between different APs) as input and a corresponding desired AP assignment output. In the disclosed methodologies, the input comprises link characteristics between the multiple APs and the multiple UEs. The channel characteristic can be seen as parameterization of the state of the wireless communications network. Different input-pairs may represent different scenarios, such as different matrixes of path gains between APs and UEs. More generally, an input-output pair may comprise link characteristics between multiple APs and the multiple UEs as input and UL or DL assignment of one or more APs of the multiple APs 110, respectively, as output. As an example, the input-output pairs may be obtained from any search method including an exhaustive search. The “exhaustive search” may also called “brute-force search” or “generate and test”. The exhaustive search is a general problem-solving technique that systematically enumerates all possible candidates (i.e. AP assignment possibilities in the disclosed methodologies) for an input, and checks whether each candidate satisfies a problem statement. The problem statement may e.g. to find a global or local maximum of an estimated performance metric of the wireless communications network (or equivalently a minimum of an undesired metric), such as throughput. Below, ^^ _^^ denotes a set of a desired AP assignment and ℐ _^^ denotes a set comprising link characteristics, where the subscript ^^ denotes an index of an input-output pair in the training data set. In context of supervised learning, ^^ _^^ may be seen as the output it is desired to map the input ℐ _^^ to. When it comes to the ML model choice, any ML model that is suitable to binary classification tasks may be used, since the AP assignment problem may be considered as a binary classification problem by, e.g., regarding indices of DL APs as 1 and those of UL APs as 0. As an example, a logistic regression model with or without a multi-layer neural network may be used for the binary classification problem. Other suitable ML models for binary problems are Bayesian networks, decision trees, and support vector machines. A ML model typically comprises a set of model parameters that are tuned during the training of the ML model. A ML model may be trained based on the input-output pairs as described above, solving the following optimization problem: where ^^ denotes the set of model parameters of the chosen ML model, ^^ is the index set of the training data set, and ℒ(ℐ _^^, ^^ _^^) describes a loss function determined by the input-output pair, which may be, for example, the binary cross entropy (BCE), which is defined as log(1 − ^^ _^^( ^^))) where the ^^ is some input set and ^^ is some output set ( ^^ may be ℐ _^^ and ^^ may be ^^ _^^), ^^ _^^ is the ^^-th element of the output set (e.g., ^^ _^^ = 0 if DL and ^^ _^^ = 1 if UL), and ^^ _^^(∙) denotes the output of the ML model for the ^^-th element. Below follows a description of reinforcement learning that may be used to train the ML model used in the disclosed methodologies. In case sets of a desired AP assignment (i.e., ^^ _^^) is not available, reinforcement learning methods may be utilized instead of the supervised counterpart described above. In reinforcement learning, a certain utility is obtained from the input data ℐ _^^, where the utility is e.g. maximized (globally or locally) through a gradient descent. In this case, other ML models than the ones suitable for supervised learning are typically used. Examples of ML models for reinforcement learning are Q-learning and deep Q-learning. These ML models also typically comprises a set of model parameters that are tuned during the training of the ML model. As an example of reinforcement learning, an ML model may be trained by maximizing an expected SINR (e.g. containing only of path gain information) from ℐ _^^. To elaborate further, an expected SINR of a ^^ _^^-th UL UE may be with ^^ ^^ _{^^ ^^ ^^} denoting the pathgain between ^^ _^^-th UL UE and ^^-th AP with ^^ ∈ ℒ _^^, ^^ ^^ _{^^ ^^} being the pathgain between ^^-th UL UE with ^^ ∈ ^^ _^^, ^^ ≠ ^^ _^^ and ^^-th AP with ^^ ∈ ℒ _^^, and ^^ ^^̈ _{^^ ^^}′ denoting the pathgain between ^^-th AP with ^^ ∈ ℒ _^^ and ^^^′-th AP with ^^^′ ∈ ℒ _^^. Based on such a “pseudo- SINR” expression, the ML model may be trained such that a sum of the pseudo-SINR expression for all UL UEs is maximized (globally or locally). In this example, no information of a desired set of AP assignment is required to train the ML model. Given the training strategy, i.e., using supervised learning or reinforcement learning, the ML model may be trained via a widely used first order optimization method such as gradient descent algorithms with the backpropagation technique that computes the gradient of a certain objective function (e.g. based on the ted desired AP assignments or on the utility) with respect to each parameter of the ML model. Note that the backpropagation numerically computes the gradient from the loss value with the current ML parameters. Here, the loss value may be the loss function discussed above for the supervised learning, or the utility for reinforcement learning, such as the pseudo-SINR expression above. In this context, the training of the disclosed ML-based AP assignment methodology may be done by means of a publicly available frameworks for training ML models (e.g., Tensorflow, Pytorch, JAX, etc.). More information about the backpropagation can be found, e.g., in https:// en.wikipedia.org/ wiki/ Backpropagation, as published on 2023-01-18. With reference to Figure 2, there is disclosed herein a computer-implemented method 200 for enabling AP assignment in a wireless communications network 100 where multiple APs 110 simultaneously serve multiple wireless devices 121. The method may e.g. be performed by an AP and/or by a data processing unit 140. Herein, simultaneously may mean using the same time resources and/or frequency resources. In particular, the multiple APs 110 may be distributed in a service area 115. As mentioned, the wireless communications network 100 may be a D-MIMO network. There are two or more APs and two or more wireless devices in the wireless communications network 100. However, there are typically a larger amount of APs and wireless devices, and the number of APs are typically 2-10 times larger than the number of wireless devises. The method comprises obtaining an uplink-downlink schedule in which each wireless device of the multiple wireless devices 121 is scheduled for either UL or DL. The uplink-downlink schedule comprises scheduling information about each wireless devices of the multiple wireless devices. The wireless devices scheduled for UL are different wireless devices than those scheduled for DL. The computer-implemented method 200 further comprises monitoring 220 link characteristics between the multiple APs 110 and the multiple wireless devices 121 over time. Herein, “between multiple APs 110 and the multiple wireless devices 121” includes any of: between different APs among the multiple APs; between different wireless devices among the multiple wireless devices 121; and between APs among the multiple APs and wireless devices among the multiple wireless devices 121. The monitored link characteristics are link characteristics obtained from the wireless communications network 100 when online. The monitored link characteristics may e.g. be based on communication links between respective APs to respective wireless devices and/or communication links between respective APs to respective other APs and/or communication links between respective wireless devices to respective other wireless devices. In other words, the monitored link characteristics may e.g. be based on communication links between pairs of APs, pairs of wireless devices, and/or pairs of wireless devices and APs. The communication links may also be called channels or communication channels. The communication links may comprise UL and/or DL transmissions, including pilots. When a first wireless device transmits a pilot signal, this pilot signal is likely received by a plurality of APs (a first group). The APs in the first group may then extract link characteristics from these pilots, such as path gains of the respective communication links. APs in a second group, that do not receive the pilot (e.g., because the received signal strength is too low), may determine that the respective links between said wireless device and the APs in the second group are unsuitable for future UL/DL transmissions. Being unsuitable is also a link characteristic. Being unsuitable may be represented by a large path loss (i.e. low path gain) value, such as 1000 dB. The link characteristics (such as path gain) may be extracted from any signals between the different APs of the multiple APs, between different wireless devices of the multiple wireless devices, and/or between APs of the multiple APs and wireless devices of the multiple wireless devices. The monitored link characteristics may in general comprise data indicative of channels statistics in the wireless communications network 100. For example, the monitored link characteristics may comprise any of path gain, received power, transmitted power, signal-to- noise ratio, SNR, signal-to-interference-plus-noise ratio, SINR, spatial channel correlation, and channel rank. The link characteristics monitored over time may be seen as representing long-term channel statistics. As mentioned, the disclosed method may update AP assignments once per time it takes for the long-term channel statistics to change. In other words, the AP assignments may be updated when at least one link characteristic changes more than a predetermined amount. Note that a given set of wireless devices are scheduled for UL and another set of wireless devices are scheduled for DL. Here, a set comprises one or more wireless devices. If any of these two sets changes, an updated assignment of the APs to operate in either UL or DL may be desired. The computer-implemented method 200 further comprises, when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determining 250 an UL or DL assignment of one or more access points of the multiple access points 110, respectively, based on the obtained uplink-downlink schedule and on the monitored link characteristics using a machine learning model. In other words, UL or DL assignments for the one or more APs is determined based on the on the monitored link characteristics and the obtained uplink-downlink schedule. Preferably, all APs of the multiple APs in the wireless communications network 100 is determined to operate in either UL or DL. A change in a link characteristic may occur if a wireless device moves from one position to another. After such change, the respective path gains of that wireless device to one or more APs of the multiple APs will likely change. This may mean that an update of AP assignments leads to improved performance of the wireless communications network, such as improved data capacity. The predetermined amount may e.g. be 1 dB difference in path gain. Here, the path gain may e.g. be between one pair of an AP and a wireless device, or a sum of path gains for several pairs. Furthermore, when the wireless device moves from the first position to the second position, it is likely that the inter-cell interference will change. Which also may motivate an update in AP assignment. The ML model is trained to provide an UL or DL assignment of one or more APs, respectively, based on uplink-downlink schedules and on link characteristics. In other words, the ML model has been trained such that the ML model may provide UL or DL assignment of the one or more APs of the multiple APs 110 as an output, when using the obtained uplink-downlink schedule and the monitored link characteristics as input. The ML model may have been trained specifically for the wireless communications network 100. In particular, the ML model may have been trained specifically for the multiple APs 110 in the wireless communications network 100. In that case, the ML model may have been trained with training data previously obtained from the wireless communications network 100 and/or previously simulated for the wireless communications network 100. However, it is possible that the ML model has been trained to be used for different wireless communications networks with respective different amounts of distributed APs. Furthermore, the ML model may have been trained using different sets of wireless devices in each different wireless communications network. Furthermore, multiple uplink-downlink schedules for each set of wireless devices may have been used during training of the ML model. The ML model may have been trained for different scenarios where different link characteristics between APs and wireless devices have been used as input. Thus, the trained ML model may be trained generally to provide respective UL or DL for any number of APs, any uplink-downlink schedules, and various link characteristics as input. The computer-implemented method 200 may further comprise assigning 260 the one or more access points of the multiple access points 110 to operate in either uplink or downlink based on the determined uplink or downlink assignment of the one or more access points of the multiple access points 110. In other words, the determined assignment for a particular AP is communicated to that particular AP. Consequently, the particular AP is assigned to operate in UL or DL according to the determined assignment. If, e.g., the method is performed by a data processing unit 140, the data processing unit may communicate the determined UL or DL assignment to each corresponding AP, via e.g. high-capacity backhaul links. The ML model may have been trained offline. The training may e.g. have been trained using supervised learning. In that case, the ML model may have been trained using a training set comprising one or more input-output pairs. Here, an input-output pair comprises link characteristics and an uplink-downlink schedule as input. The input-output pair also comprises UL or DL assignment of one or more APs of the multiple APs 110, respectively, as output. In an input-out pair, a particular state of a particular wireless communications network, parametrized as the link characteristics between APs and wireless devices of particular wireless communications network and as the uplink-downlink schedule – which comprises scheduling information about each wireless devices in the particular wireless communications network, is paired with a desired output, namely, a desired UL or DL assignment of one or more APs of the multiple APs, respectively, for that particular state of the particular wireless communications network. If the ML model has been specifically trained for the wireless communications network 100, the input of an input-output pair may comprises previously obtained (e.g. from measurements or simulations) link characteristics between the multiple APs 110 and the multiple wireless devices 121 of the wireless communications network 100, and the output of an input-output pair may comprises UL or DL assignment of one or more APs of the multiple APs 110 of the wireless communications network 100, respectively. The link characteristics for training the ML model preferably comprise same type of data as the motored link characteristics in the wireless communications network 100. The one or more input-output pairs may have been obtained using an exhaustive search. In other words, a desired UL or DL assignment of one or more APs of multiple APs for a particular state of a particular wireless communications network may be found by evaluation all possible assignments. The exhaustive search is a separate step from training the ML model. The ML model may be subsequently trained using obtained one or more input-output pairs. Alternatively, the ML model may have been trained using reinforcement learning. In that case, the ML model may have been trained using a predetermined utility in terms of link quality and using a training set comprising one or more input sets. Here, an input set comprises link characteristics between and an uplink-downlink schedule. In other words, the ML model is directly trained to provide a desired UL or DL assignment of one or more APs, respectively, for a particular state of a particular wireless communications network. Here, the particular state of the particular wireless communications network is parameterized as the link characteristics and as the uplink-downlink schedule. Similarly, the predetermined utility can be seen as a parameterization of the desired UL or DL assignment. The predetermined utility can be seen as a reward function for the reinforcement learning. If the ML model has been specifically trained for the wireless communications network 100, an input set may comprises previously obtained (e.g. from measurements or simulations) link characteristics between the multiple APs 110 and the multiple wireless devices 121 of the wireless communications network 100. In that case, the predetermined utility in terms of link quality may be a predetermined utility in terms of link quality of the wireless communications network 100. The predetermined utility may comprises any of path gain, received power, transmitted power, signal-to-noise ratio, SNR, SINR, spatial channel correlation, channel rank and data capacity. In particular, the predetermined utility comprises information of interference. This interference may be between APs, between wireless devices, and/or between APs and wireless devices As mentioned, the deployment scenario of the wireless communications network 100 change during online deployment. Consequently, the trained ML model may not provide an AP assignment resulting in a desired performance of the wireless communications network 100. Therefore, it may be desired to retrain the ML model. Retraining may use an original training set that was used in the offline training of the ML model, as well as new training data based on the monitored link characteristics obtained during online deployment. The retraining of the ML model may be triggered by a condition in the monitored link characteristics. In other words, the computer-implemented method 200 may further comprising determining 230 whether the ML model should be retrained based on the monitored link characteristics, and thereafter retraining 240 the ML model based on the monitored link characteristics when it is determined to retrain the ML model. As mentioned, the retraining may also be based on the original training set. According to an example, the ML model is retrained using the original training set that is weighted based on the monitored link characteristics. In another example, the ML model is retrained using a weighted combination of the original training set and the monitored link characteristics. As an example, the computer-implemented method 200 may comprise determining 231 that the ML model is to be retrained when a performance metric of the wireless communications network 100 based on the monitored link characteristics is below a first threshold value. The performance metric may e.g. be throughput. In that case, when the throughput is below first threshold value, retraining of the ML model is triggered. The performance metric may also comprise information such as signal-to-interference-plus-noise ratio, SINR. As another example, the ML model is further trained to output an expected performance metric for any given DL or UL assignment outputted by the ML model. In that case, the computer- implemented method 200 may comprise determining 232 that the ML model is to be retrained when a performance metric of the wireless communications network 100 based on the monitored link characteristics differs from the expected performance metric more than a second threshold value. The expected performance metric comprises the same type of data as the performance metric of the wireless communications network 100 based on the monitored link characteristics. The expected performance metric may comprise any of signal- to-interference-plus-noise ratio, SINR, and throughput. In yet another example, the monitored link characteristics comprises one or more path losses between different APs of the multiple APs 110, and/or between any AP of the multiple APs 110 and any wireless device of the multiple wireless devices 121, and/or between different wireless devices of the multiple wireless devices 121. In that case, the computer-implemented method 200 may comprise determining 233 that the ML model is to be retrained when one or more of the one or more path losses have changed more than a third threshold value in a first time interval. As an example, the ML model may be determined to be retrained when all path losses (e.g. a sum of all path losses) has changed more than the third threshold. Here, all path losses are the path losses between different APs of the multiple APs 110, and/or between any AP of the multiple APs 110 and any wireless device of the multiple wireless devices 121, and/or between different wireless devices of the multiple wireless devices. For example, an investigation if the one or more path losses mentioned in the previous paragraph have changed may occur every day. More generally, however, the first time interval may be 1-10 days. However, the first time interval may be any number. The first time interval may be static or dynamic. For example, the first time interval may be based on past link characteristics of the monitored link characteristics. In an example scenario, a path gain between two different APs has remained the same for several weeks. In that case, the first time interval may currently be set to 10 days. Then, a new building emerges that obstructs the channel between said two APs, which changes the link characteristics. The first time interval is then dynamically changed to 1 day. This is an example of a reactive change of the first time interval. The computer-implemented method 200 may comprise monitoring 211 whether the at least one link characteristic of the monitored link characteristics changes more than the predetermined amount in a second time interval. For example, an investigation if the at least one link characteristic has changed may occur every 1 seconds. More generally, however, the second time interval may be 0.1-10 seconds. However, the second time interval may be any number. The second time interval may be static or dynamic. For example, the second time interval may be based on past link characteristics of the monitored link characteristics. In an example scenario, a wireless device has been stationary for 10 minutes. In that case, the second time interval may be currently set to 10 seconds. Then, the wireless device moves, which changes the link characteristics. The second time interval is then dynamically changed to 1 second. This is an example of a reactive change of the second time interval. The second time interval may alternatively, or in combination of, be dynamically changed proactively. In other words, the second time interval may be based on an expected time the respective wireless devices of the multiple wireless devices 121 are expected to move less than a predetermined distance. Figure 3 shows a flowchart of an example embodiment of a method 300 including steps from AP assignment to data transmission. Figure 4 shows a flowchart of an example of method 400 for training the ML model. The example training method of Figure 4 may be used together with the assignment step 340 in Figure 3. In Figure 3, the method starts at step 310. At step 320, the method determines whether a current AP assignment should be changed in the wireless communications network 100. If there has been no significant change in the wireless communications network 100, there may be no need to change the AP assignment. If, on the other hand, there is change, such as one or more wireless devices moving within the wireless communications network, it may be desired to change the AP assignment. If it is decided to keep the current AP assignment, the method proceeds to step 350, which is discussed below. If it is decided to change the AP assignment, the method proceeds to step 340 to make a new AP assignment. Thereafter, the method proceeds to step 350. At step 350, the method 300 performs a channel estimation between the APs and the wireless devices. This step may utilize data of the wireless communications network also used for the assignment step 340. Next, the method proceeds to step 360 that updates beam weights used in the D-MIMO system and transmits/receives data in UL and DL. At step 370, the method checks if any data remains in the buffer and if the channel estimation should be updated. If that is the case, the method goes back to step 320. If that is not the case, the method stops at 380. In Figure 4, the method 400 starts at step 410. In step 420, the method checks at step 420 whether a re-training criterion is triggered. Different ways of determining if the ML model should be retrained is discussed above. The method 400 may be triggered between steps 320 and 340 of the method 300 in Figure 3. If, at step 420 in the method 400, no re-training is triggered, the method 400 stops at step 430. If, on the other hand, a re-training criterion is triggered, the method proceeds to step 440, where an available training set is obtained. At step 450, the method checks if a training set for supervised learning is available. If available, the method proceeds to step 460 to train the ML model using supervised learning. If, on the other hand, a training set for supervised learning is not available, the method proceeds to step 470 to train the ML model using reinforcement learning. After the training of the ML model is completed, the method goes back to step 420 to check if the re-training condition is still triggered when using the newly trained ML model. Figure 5 shows a spectrum efficiency comparison for various AP assignment methods. The assumptions are as follows. The transmit power is set to 100 mW at UEs and 100 mW at APs. The data rate is calculated by the Shannon channel capacity formula with SINR formulations by equation (1) and (2). The path loss is modelled by the 3GPP Indoor Hotspot path loss model (namely, 3GPP TR 38.901 V 14.0.0 InH - Office in Table 7.4.1-1 and LOS/NLOS probability according to Table 7.4.2-1), while the fading is modelled as a Rayleigh random variable. It is assumed that 16 APs are distributed in a square grid fashion, while 5 UEs are distributed randomly within a square service area with one-side length of 50 meter. The UL/DL demand/scheduling is randomly selected for each UE. The beamforming weights are designed such that the mean square error (MSE) of equations (1) and (2) is minimized (i.e., MMSE) with instantaneous channel information. Figure 5 compares the following assignment methods: 1. A TDD scenario where the available time is divided into UL and DL in an equal manner. This particular assignment is denoted as “TDD” in Figure 5 and is plotted as a dashed line. The SINR is calculated under the condition that there is no interference from UL-UEs to DL-UEs and from ^{DL-APs to UL-APs.} In other words, this method represents a conventional TDD scenario where UL and DL is separated in time. 2. A DTDD scenario where the AP assignment is done randomly. This particular assignment is denoted “Random-AP assign. (DTDD)” in Figure 5 and is plotted as a solid line. In this scenario, the received signals comprise cross-link interference as shown in equation (1) and (2). Whether each AP is assigned to UL/DL is determined by a Bernoulli process with the equal probability being UL/DL. In order to ensure at least one AP serving UL/DL, one AP is randomly selected to serve DL/UL in case all APs are chosen to be UL/DL. 3. A DTDD scenario where the AP UL/DL assignment is done based on an exhaustive search. This particular assignment is denoted as “Bruteforce-AP assign. (DTDD)” in Figure 5 and is plotted with a solid line marked with “x”. The exhaustive search is implemented to find an AP UL/DL assignment set that maximizes the sum of throughput among UEs. Since all possible combinations are too large to search for within a reasonable time (i.e., 2¹⁶ combinations), a set of unique 2¹⁵ unique combinations are generated, and the one out of 2¹⁵ combinations that maximizes the sum throughput are found through the exhaustive search. This represents a type of optimum solution of possible assignments. 4. A DTDD scenario where the AP UL/DL assignment is done according to an example embodiment of the method disclosed herein. This particular assignment is denoted as “ML with supervised learning” in Figure 5 and is plotted with circles. Note that this assignment method overlaps with the assignment method denoted “Bruteforce-AP assign. (DTDD)” in Figure 5. This example embodiment corresponds to a case where the system operates in DTDD mode with an ML-based AP assignment. To elaborate, a logistic regression with the binary cross entropy is used as the objective function, where the input data is a) UEs-APs pathloss, b) UE- UE pathloss, c) AP-AP pathloss, and d) UEs’ UL/DL demand (i.e., which UE is in UL/DL) while the output data that the model tries to map the input to is the AP assignment of the exhaustive search denoted “Bruteforce” in Figure 5. As for the ML model, a 2-layers fully connected neural network with 1 hidden layer with the Rectified Linear Unit Activation Function between the input layer and hidden layer is used. The number of parameters of the considered ML model is 54192, which may be considered a small neural network. The number of the training samples used is 4000 samples, which is collected by running the exhaustive search 4000 times. The Rectified Linear Unit Activation Function is described in https:// en.wikipedia.org/ wiki/ Rectifier_(neural_networks), as published on 2023-01-18. From Figure 5, it can be observed that the exhaustive search based AP assignment method, with the line denoted “Bruteforce-AP assign. (DTDD)” presents the best performance. However, at the cost of a very large computation overhead. The disclosed ML-based approach (denoted “ML with supervised learning”) presents the same the performance as the method based on the exhaustive search (denoted “Bruteforce-AP assign. (DTDD)”). Note that the processing cost of the disclosed ML approach after training is much smaller than that of the exhaustive search, since a logistic regression model is used in this case. In other words, the disclosed ML-based approach requires less computational overhead during online deployment. There is also disclosed herein a node 110, 140 for enabling AP assignment in a wireless communications network 100 where multiple access points 110 simultaneously serve multiple wireless devices 121. The node may e.g. be an AP 110 and/or a data processing unit 140. Figure 6 shows a schematic block diagram of embodiments of an AP 110. The schematic block diagram in Figure 6 also represents embodiments of a data processing unit 140, which in this example have the same type of components as the AP 110. The embodiments of the node 110, 140 may be considered as independent embodiments or may be considered in any combination with each other. It should also be noted that, although not shown in Figure 6, the node 110, 140 may comprise known conventional features for such devices, such as a power source like a battery or mains connection. If the node is an AP, a conventional feature may e.g. be an antenna arrangement. The node 110, 140 may comprise processing circuitry 610 and a memory 620. The processing circuitry 610 may comprise a receiving module 611 and a transmitting module 612. The receiving module 611 and the transmitting module 612 may comprise radio frequency circuitry and baseband processing circuitry capable of transmitting and receiving a radio signal in the wireless communications network 100. The receiving module 611 and the transmitting module 612 may also form part of a single transceiver. It should also be noted that some or all of the functionality described in the embodiments above as being performed by the node 110, 140 may be provided by the processing circuitry 610 executing instructions stored on a computer- readable medium, such as, e.g. the memory 620 shown in Figure 6. Alternative embodiments of the node 110, 140 may comprise additional components, such as, an obtaining module 613, a monitoring module 614, a determining module 615, an assigning module 616 and/or a retraining module 617, responsible for providing functionality to support the embodiments of the node 110, 140 described herein. The node 110, 140, processing circuitry 610, or obtaining module 613 is configured to obtain an uplink-downlink schedule in which each wireless device of the multiple wireless devices 121 is scheduled for either uplink or downlink. The node 110, 140, processing circuitry 610, or monitoring module 614 is configured to monitor link characteristics between the multiple access points 110 and the multiple wireless devices 121 over time. The node 110, 140, processing circuitry 610, or assigning module 616 is configured to, when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determine an uplink or downlink assignment of one or more access points of the multiple access points 110, respectively, based on the obtained uplink-downlink schedule and on the monitored link characteristics using a machine learning model, wherein the machine learning model is trained to provide an uplink or downlink assignment of one or more access points, respectively, based on uplink-downlink schedules and on link characteristics. In said node 110, 140, the ML model may have been trained using supervised learning. Furthermore, the ML model may have been trained using a training set comprising one or more input-output pairs, where an input-output pair comprises an uplink-downlink schedule and link characteristics as input, and UL or DL assignment of one or more APs, respectively, as output. In that case, the one or more input-output pairs may have been obtained using an exhaustive search. Alternatively, the ML model may have been trained using reinforcement learning. In that case, the ML model may have been trained using a predetermined utility in terms of link quality and using a training set comprising one or more input sets, where an input set comprises an uplink- downlink schedule link characteristics between the multiple APs 110 and the multiple wireless devices 121. The node 110, 140, processing circuitry 610, or determining module 615 may be configured to determine whether the ML model should be retrained based on the monitored link characteristics. In that case, the node 110, 140, processing circuitry 610, or retraining module 617 may be configured to retrain the ML model based on the monitored link characteristics when it is determined to retrain the ML model. The node 110, 140, processing circuitry 610, or determining module 615 may be configured to determine that the ML model is to be retrained when a performance metric of the wireless communications network 100 based on the monitored link characteristics is below a first threshold value. Alternatively, the ML model may be further trained to output an expected performance metric for any given DL or UL assignment outputted by the ML model. In that case, the node 110, 140, processing circuitry 610, or determining module 615 may be configured to determine that the ML model is to be retrained when a performance metric of the wireless communications network 100 based on the monitored link characteristics differs from the expected performance metric more than a second threshold value. Here, the expected performance metric and the performance metric of the wireless communications network 100 may be based on the monitored link characteristics comprise any of signal-to-interference-plus-noise ratio, SINR, and throughput. In yet another alternative, the monitored link characteristics comprises one or more path losses between different APs of the multiple APs 110, and/or between any AP of the multiple APs 110 and any wireless device of the multiple wireless devices 121, and/or between different wireless devices of the multiple wireless devices. In that case, the node 110, 140, processing circuitry 610, or determining module 615 may be configured to determine that the ML model is to be retrained when one or more of the one or more path losses have changed more than a third threshold value in a first time interval. Here, the predetermined utility may comprises any of path gain, received power, transmitted power, signal-to-noise ratio, SNR, SINR, spatial channel correlation, channel rank and data capacity. Alternatively, or in combination of, the predetermined utility may comprise information of interference. In the node 110, 140, the monitored link characteristics may comprise any of path gain, received power, transmitted power, SNR, SINR, spatial channel correlation, and channel rank. The node 110, 140 or processing circuitry 610 may be configured to assign the one or more access points of the multiple access points 110 to operate in either uplink or downlink based on the determined uplink or downlink assignment of the one or more access points of the multiple access points 110. The node 110, 140, processing circuitry 610, or monitoring module 614 may be configured to monitor whether the at least one link characteristic of the monitored link characteristics changes more than the predetermined amount in a second time interval. The first and/or second time intervals may be based on past link characteristics of the of the monitored link characteristics. Furthermore, the first time interval may be based on respective times wireless devices of the multiple wireless devices 121 are expected to move less than a predetermined distance. The first time interval may be 1-10 days. The second time interval may be 0.1-10 seconds. The methods disclosed herein may be implemented through one or more processors, such as the processing circuitry 610 in the node 110, 140 depicted in Figure 6, together with computer program code for performing the functions and actions of the embodiments herein. The program code may also be provided as a computer program product, for instance in the form of a data carrier carrying computer program code or code means for performing the embodiments herein when being loaded into the processing circuitry 610 in the node 110, 140. The computer program code may e.g. be provided as pure program code in the node 110, 140 or on a server and downloaded to the node. Thus, it should be noted that the modules of the node 110, 140 may in some embodiments be implemented as computer programs stored in memory, e.g. in the memory modules 620 in Figure 6, for execution by processors or processing modules, e.g. the processing circuitry 610 of Figure 6. Those skilled in the art will also appreciate that the processing circuitry 610 and the memory 620 described above may refer to a combination of analog and digital circuits, and/or one or more processors configured with software and/or firmware, e.g. stored in a memory, that when executed by the one or more processors such as the processing circuitry 610 perform as described above. One or more of these processors, as well as the other digital hardware, may be included in a single application- specific integrated circuit (ASIC), or several processors and various digital hardware may be distributed among several separate components, whether individually packaged or assembled into a system-on-a-chip (SoC). The description of the example embodiments provided herein have been presented for purposes of illustration. The description is not intended to be exhaustive or to limit example embodiments to the precise form disclosed, and modifications and variations are possible in light of the above teachings or may be acquired from practice of various alternatives to the provided embodiments. The examples discussed herein were chosen and described in order to explain the principles and the nature of various example embodiments and its practical application to enable one skilled in the art to utilize the example embodiments in various manners and with various modifications as are suited to the particular use contemplated. The features of the embodiments described herein may be combined in all possible combinations of methods, apparatus, modules, systems, and computer program products. It should be appreciated that the example embodiments presented herein may be practiced in any combination with each other. It should be noted that the word “comprising” does not necessarily exclude the presence of other elements or steps than those listed and the words “a” or “an” preceding an element do not exclude the presence of a plurality of such elements. It should further be noted that any reference signs do not limit the scope of the claims, that the example embodiments may be implemented at least in part by means of both hardware and software, and that several “means”, “units” or “devices” may be represented by the same item of hardware. It should also be noted that the various example embodiments described herein are described in the general context of method steps or processes, which may be implemented in one aspect by a computer program product, embodied in a computer-readable medium, including computer-executable instructions, such as program code, executed by computers in networked environments. A computer-readable medium may include removable and non- removable storage devices including, but not limited to, Read Only Memory (ROM), Random Access Memory (RAM), compact discs (CDs), digital versatile discs (DVD), etc. Generally, program modules may include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Computer-executable instructions, associated data structures, and program modules represent examples of program code for executing steps of the methods disclosed herein. The particular sequence of such executable instructions or associated data structures represents examples of corresponding acts for implementing the functions described in such steps or processes. The embodiments herein are not limited to the above-described preferred embodiments. Various alternatives, modifications and equivalents may be used. Therefore, the above embodiments should not be construed as limiting.

Claims

CLAIMS 1. A computer-implemented method (200) for enabling access point assignment in a wireless communications network (100) where multiple access points (110) simultaneously serve multiple wireless devices (121), the computer-implemented method (200) comprising: obtaining (210) an uplink-downlink schedule in which each wireless device of the multiple wireless devices (121) is scheduled for either uplink or downlink; monitoring (220) link characteristics between the multiple access points (110) and the multiple wireless devices (121) over time; and when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determining (250) an uplink or downlink assignment of one or more access points of the multiple access points (110), respectively, based on the obtained uplink-downlink schedule and on the monitored link characteristics using a machine learning model, wherein the machine learning model is trained to provide an uplink or downlink assignment of one or more access points, respectively, based on uplink-downlink schedules and on link characteristics.

2. The computer-implemented method (200) according to claim 1, wherein the machine learning model has been trained using supervised learning.

3. The computer-implemented method (200) according to claim 2, wherein the machine learning model has been trained using a training set comprising one or more input-output pairs, where an input-output pair comprises an uplink-downlink schedule and link characteristics as input, and uplink or downlink assignment of one or more access points, respectively, as output.

4. The computer-implemented method (200) according to claim 3, wherein the one or more input-output pairs have been obtained using an exhaustive search.

5. The computer-implemented method (200) according to claim 1, wherein the machine learning model has been trained using reinforcement learning.

6. The computer-implemented method (200) according to claim 5, wherein the machine learning model has been trained using a predetermined utility in terms of link quality and using a training set comprising one or more input sets, where an input set comprises an uplink- downlink schedule and link characteristics.

7. The computer-implemented method (200) according to any previous claim, further comprising determining (230) whether the machine learning model should be retrained based on the monitored link characteristics, and retraining (240) the machine learning model based on the monitored link characteristics when it is determined to retrain the machine learning model.

8. The computer-implemented method (200) according to claim 7, comprising determining (231) that the machine learning model is to be retrained when a performance metric of the wireless communications network (100) based on the monitored link characteristics is below a first threshold value.

9. The computer-implemented method (200) according to claim 7, wherein the machine learning model is further trained to output an expected performance metric for any given downlink or uplink assignment outputted by the machine learning model, wherein the computer-implemented method (200) comprises determining (232) that the machine learning model is to be retrained when a performance metric of the wireless communications network (100) based on the monitored link characteristics differs from the expected performance metric more than a second threshold value.

10. The computer-implemented method (200) according to claim 8 or 9, wherein the expected performance metric and the performance metric of the wireless communications network (100) based on the monitored link characteristics comprise any of signal-to- interference-plus-noise ratio, SINR, and throughput.

11. The computer-implemented method (200) according to claim 7, wherein the monitored link characteristics comprises one or more path losses between different access points of the multiple access points (110), and/or between any access point of the multiple access points (110) and any wireless device of the multiple wireless devices (121), and/or between different wireless devices of the multiple wireless devices, wherein the computer-implemented method (200) comprises determining (233) that the machine learning model is to be retrained when one or more of the one or more path losses have changed more than a third threshold value in a first time interval.

12. The computer-implemented method (200) according to any previous claim when dependent on claim 6, wherein the predetermined utility comprises any of path gain, received power, transmitted power, signal-to-noise ratio, SNR, SINR, spatial channel correlation, channel rank and data capacity.

13. The computer-implemented method (200) according to any previous claim when dependent on claim 6, wherein the predetermined utility comprises information of interference.

14. The computer-implemented method (200) according to any previous claim, wherein the monitored link characteristics comprise any of path gain, received power, transmitted power, SNR, SINR, spatial channel correlation, and channel rank.

15. The computer-implemented method (200) according to any previous claim, further comprising assigning (260) the one or more access points of the multiple access points (110) to operate in either uplink or downlink based on the determined uplink or downlink assignment of the one or more access points of the multiple access points (110).

16. The computer-implemented method (200) according to any previous claim, further comprising monitoring (221) whether the at least one link characteristic of the monitored link characteristics changes more than the predetermined amount in a second time interval.

17. The computer-implemented method (200) according to claim 11 or 16, wherein the first and/or second time intervals are based on past link characteristics of the of the monitored link characteristics.

18. The computer-implemented method (200) according to any of claims 16-17, wherein first time interval is based on respective times wireless devices of the multiple wireless devices (121) are expected to move less than a predetermined distance.

19. The computer-implemented method (200) according to claim 11 wherein the first time interval is 1-10 days.

20. The computer-implemented method (200) according to any of claims 16-19, wherein the second time interval is 0.1-10 seconds.

21. A node (110, 140) for enabling access point assignment in a wireless communications network (100) where multiple access points (110) simultaneously serve multiple wireless devices (121), wherein the node (110, 140) comprises a processing circuitry (710) and a memory (720), the processing circuitry being configured to obtain an uplink-downlink schedule in which each wireless device of the multiple wireless devices (121) is scheduled for either uplink or downlink, monitor link characteristics between the multiple access points (110) and the multiple wireless devices (121) over time, and when at least one link characteristic of the monitored link characteristics changes more than a predetermined amount, determine an uplink or downlink assignment of one or more access points of the multiple access points (110), respectively, based on the obtained uplink-downlink schedule and on the monitored link characteristics using a machine learning model, wherein the machine learning model is trained to provide an uplink or downlink assignment of one or more access points, respectively, based on uplink-downlink schedules and on link characteristics.

22. The node (110, 140) according to claim 21, wherein the machine learning model has been trained using supervised learning.

23. The node (110, 140) according to claim 22, wherein the machine learning model has been trained using a training set comprising one or more input-output pairs, where an input- output pair comprises an uplink-downlink schedule and link characteristics as input, and uplink or downlink assignment of one or more access points, respectively, as output.

24. The node (110, 140) according to claim 23, wherein the one or more input-output pairs have been obtained using an exhaustive search.

25. The node (110, 140) according to claim 21, wherein the machine learning model has been trained using reinforcement learning.

26. The node (110, 140) according to claim 25, wherein the machine learning model has been trained using a predetermined utility in terms of link quality and using a training set comprising one or more input sets, where an input set comprises an uplink-downlink schedule and link characteristics.

27. The node (110, 140) according to any of claims 21-26, wherein the processing circuitry is further configured to determine whether the machine learning model should be retrained based on the monitored link characteristics, and retrain the machine learning model based on the monitored link characteristics when it is determined to retrain the machine learning model.

28. The node (110, 140) according to claim 27, wherein the processing circuitry is configured to determine that the machine learning model is to be retrained when a performance metric of the wireless communications network (100) based on the monitored link characteristics is below a first threshold value.

29. The node (110, 140) according to claim 27, wherein the machine learning model is further trained to output an expected performance metric for any given downlink or uplink assignment outputted by the machine learning model, wherein the processing circuitry is configured to determine that the machine learning model is to be retrained when a performance metric of the wireless communications network (100) based on the monitored link characteristics differs from the expected performance metric more than a second threshold value.

30. The node (110, 140) according to claim 28 or 29, wherein the expected performance metric and the performance metric of the wireless communications network (100) based on the monitored link characteristics comprise any of signal-to-interference-plus-noise ratio, SINR, and throughput.

31. The node (110, 140) according to claim 27, wherein the monitored link characteristics comprises one or more path losses between different access points of the multiple access points (110), and/or between any access point of the multiple access points (110) and any wireless device of the multiple wireless devices (121), and/or between different wireless devices of the multiple wireless devices, wherein the processing circuitry is configured to determine that the machine learning model is to be retrained when one or more of the one or more path losses have changed more than a third threshold value in a first time interval.

32. The node (110, 140) according to any of claims 21-31 when dependent on claim 26, wherein the predetermined utility comprises any of path gain, received power, transmitted power, signal-to-noise ratio, SNR, SINR, spatial channel correlation, channel rank and data capacity.

33. The node (110, 140) according to any of claims 21-32 when dependent on claim 26, wherein the predetermined utility comprises information of interference.

34. The node (110, 140) according to any of claims 21-33, wherein the monitored link characteristics comprise any of path gain, received power, transmitted power, SNR, SINR, spatial channel correlation, and channel rank.

35. The node (110, 140) according to any of claims 21-34, wherein the processing circuitry is further configured to assign the one or more access points of the multiple access points (110) to operate in either uplink or downlink based on the determined uplink or downlink assignment of the one or more access points of the multiple access points 110.

36. The node (110, 140) according to any of claims 21-35, wherein the processing circuitry is further configured to monitor whether the at least one link characteristic of the monitored link characteristics changes more than the predetermined amount in a second time interval.

37. The node (110, 140) according to claim 31 or 36, wherein the first and/or second time intervals are based on past link characteristics of the of the monitored link characteristics.

38. The node (110, 140) according to any of claims 36-37, wherein first time interval is based on respective times wireless devices of the multiple wireless devices (121) are expected to move less than a predetermined distance.

39. The node (110, 140) according to claim 31, wherein the first time interval is 1-10 days.

40. The node (110, 140) according to any of claims 36-39, wherein the second time interval is 0.1-10 seconds.

41. The node (110, 140) according to any of claims 21-40, wherein the node is a network node (110) and/or a remote data processing unit (140).

42. A computer program product comprising instructions which, when executed on at least one processing circuitry (710), cause the at least one processing circuitry to carry out the method according to any of claims 1-20.

43. A computer program carrier carrying a computer program product according to claim 42, wherein the computer program carrier is one of an electronic signal, optical signal, radio signal, or computer-readable storage medium.