Iterative process to enhance separate transmit / receive training for the channel state feedback of the encoder.

Abstract

The user device (UE) computes the raw downlink (DL) channel between the base station and the UE based on the first Channel State Information Reference Signal (CSI-RS), determines, based on the raw DL channel estimate, the precoder information indicating the precoder that will cause the UE to experience the desired effective DL channel, encodes the precoder information into a bit string using the machine learning (ML) model encoder at the UE, sends the bit string to the base station, computes the actual effective DL channel that the UE experiences based on the second CSI-RS, and calculates the error between the desired effective DL channel and the actual effective DL channel. The UE then adjusts its encoder based on the error and / or sends, to the base station, the error information corresponding to the error. The invention also proposes related network-side functions.

Description

AN ITERATIVE PROCEDURE FOR SEPARATE RECEIVE / TRANSMIT TRAININGENHANCEMENT FOR PRECODER CHANNEL STATE INFORMATION FEEDBACKTECHNICAL FIELD

[0001] This application relates generally to wireless communication systems, including system implementing artificial intelligence (AI) / machine learning (ML) models for the purpose of sending and / or receiving channel state information (CSI) feedback.BACKGROUND

[0002] Wireless mobile communication technology uses various standards and protocols to transmit data between a base station and a wireless communication device. Wireless communication system standards and protocols can include, for example, 3rd Generation Partnership Project (3GPP) Long Term Evolution (LTE) (e.g., 4G), 3GPP New Radio (NR) (e.g.. 5G), and Institute of Electrical and Electronics Engineers (IEEE) 802. 11 standard for Wireless Local Area Networks (WLAN) (commonly known to industry groups as Wi-Fi®).

[0003] As contemplated by the 3GPP, different wireless communication systems' standards and protocols can use various radio access networks (RANs) for communicating between a base station of the RAN (which may also sometimes be referred to generally as a RAN node, a network node, or simply a node) and a wireless communication device known as a user equipment (UE). 3GPP RANs can include, for example, Global System for Mobile communications (GSM), Enhanced Data Rates for GSM Evolution (EDGE) RAN (GERAN), Universal Terrestrial Radio Access Network (UTRAN), Evolved Universal Terrestrial Radio Access Network (E-UTRAN), and / or Next-Generation Radio Access Network (NG-RAN).

[0004] Each RAN may use one or more radio access technologies (RATs) to perform communication between the base station and the UE. For example, the GERAN implements GSM and / or EDGE RAT, the UTRAN implements Universal Mobile Telecommunication System (UMTS) RAT or other 3GPP RAT, the E-UTRAN implements LTE RAT (sometimes simply referred to as LTE), and NG-RAN implements NR RAT (sometimes referred to herein as 5G RAT, 5G NR RAT, or simply NR). In certain deployments, the E- UTRAN may also implement NR RAT. In certain deployments, NG-RAN may also implement LTE RAT.

[0005] A base station used by a RAN may correspond to that RAN. One example of an E- UTRAN base station is an Evolved Universal Terrestrial Radio Access Network (E- UTRAN) Node B (also commonly denoted as evolved Node B, enhanced Node B, eNodeB,or eNB). One example of an NG-RAN base station is a next generation Node B (also sometimes referred to as a g Node B or gNB).

[0006] A RAN provides its communication services with external entities through its connection to a core network (CN). For example, E-UTRAN may utilize an Evolved Packet Core (EPC) while NG-RAN may utilize a 5G Core Network (5GC).BRIEF DESCRIPTION OF THE SEVERAL VIEWS OF THE DRAWINGS

[0007] To easily identify the discussion of any particular element or act, the most significant digit or digits in a reference number refer to the figure number in which that element is first introduced.

[0008] FIG. 1 illustrates a diagram showing an example of a two-sided AI / ML model for CSI compression and decompression, according to embodiments discussed herein.

[0009] FIG. 2 illustrates a diagram showing a training that occurs in a manner that is joint at one of the UE or the network.

[0010] FIG. 3 illustrates a diagram showing a training that occurs in a manner that is joint and across each of the UE and the network (each of the UE and the network are active participants in the training).

[0011] FIG. 4 illustrates a flow diagram showing that a first training of the encoder of the UE side of the model may occur at the UE, while a second training of the decoder of the network side of the model may occur at the network.

[0012] FIG. 5 illustrates a flow diagram showing an example procedure between a network and a UE for training an AI / ML model for CSI feedback using joint training at the network.

[0013] FIG. 6 illustrates a flow diagram for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network (e.g., a base station of the network) and a UE, according to embodiments discussed herein.

[0014] FIG. 7 illustrates a flow diagram for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network (e.g., a base station of the network) and a UE, according to embodiments discussed herein.

[0015] FIG. 8 illustrates a flow diagram for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network (e.g., a base station of the network) and a UE, according to embodiments discussed herein.

[0016] FIG. 9 illustrates a flow diagram for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network (e.g., a base station of the network) and a UE, according to embodiments discussed herein.

[0017] FIG. 10 illustrates a flow diagram for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between the network (e.g., a base station of the network) and the UE, according to embodiments discussed herein.

[0018] FIG. 11 illustrates a flow diagram for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network (e.g., a base station of the network) and a UE, according to embodiments discussed herein.

[0019] FIG. 12 illustrates a diagram showing an explicit calculation of CQI / RI as may be carried out by a UE.

[0020] FIG. 13 illustrates a diagram for using an encoder of an AI / ML model at a UE to generate CQI / RI corresponding to input, according to embodiments herein.

[0021] FIG. 14 illustrates a diagram showing an example of reporting CSI / CQI / RI after AI / ML model alignment between a network and a UE, according to embodiments described herein.

[0022] FIG. 15 illustrates a method of a UE, according to embodiments herein.

[0023] FIG. 16 illustrates a method of a UE, according to embodiments herein.

[0024] FIG. 17 illustrates a method of a base station, according to embodiments herein.

[0025] FIG. 18 illustrates a method of a UE, according to embodiments herein.

[0026] FIG. 19 illustrates a method of a UE, according to embodiments herein.

[0027] FIG. 20 illustrates a method of a base station, according to embodiments herein.

[0028] FIG. 21 illustrates an example architecture of a wireless communication system, according to embodiments disclosed herein.

[0029] FIG. 22 illustrates a system for performing signaling between a wireless device and a network device, according to embodiments disclosed herein.DETAILED DESCRIPTION

[0030] Various embodiments are described with regard to a UE. However, reference to a UE is merely provided for illustrative purposes. The example embodiments may be utilizedwith any electronic component that may establish a connection to a network and is configured with the hardware, software, and / or firmware to exchange information and data with the network. Therefore, the UE as described herein is used to represent any appropriate electronic component.

[0031] FIG. 1 illustrates a diagram 100 showing an example of a two-sided artificial intelligence (Al)Zmachine learning (ML) model 102 for channel state information (CSI) compression and decompression, according to embodiments discussed herein. The model 102 includes the (logical) UE side 104 (a portion of the model 102 that exists at a UE 106) and the (logical) network side 108 (a portion of the model 102 that exists at, for example, a base station of a network 110), as illustrated.

[0032] The UE side 104 illustrates that the UE 106 includes an encoder 112. The encoder 112 is configured to accept an input 116 and to provide a bitstream 118 that is based on that input 116 as output. As illustrated, in some cases, the input 116 may include downlink (DL) channel information, such as a raw DL channel estimate. In some cases, the input 116 may include precoder information, such as a precoder W and / or a set of eigenvectors corresponding to a precoder W.

[0033] The bitstream 118 is then transmitted from the UE 106 to the network 110.

[0034] The network side 108 illustrates that the network 110 (e.g., a base station of the network 110) includes a decoder 114. The decoder 114 is configured to accept the bitstream 118 as input, and to decode information therein as output 120 to the network 110 for further processing. In this way, the original information from the input 116 is known to the network 110. Accordingly, in some cases, the output 120 may include downlink (DL) channel information, such as a raw DL channel estimate. In some cases, the output 120 may include precoder information, such as a precoder W and / or a set of eigenvectors corresponding to a precoder W.

[0035] Based on the encoding mechanism used by the encoder 112, the bitstream 118 may be smaller than the raw data as presented in the input 116. The encoder 112 may thus be understood to ‘‘compress'’ the input 116 into the bitstream, which is correspondingly understood to represent “compressed” information. The result is that the transmission of the bitstream 118 from the UE 106 to the network 110 results in the use of fewer radio resources than the alternative case where the raw data (e.g., as found in the input 116) is sent from the UE 106 to the network side 108 without such encoding / compression.

[0036] The model 102 (including the UE side 104 and the network side 108) may be a trained AI / ML model. For the purposes of training a two-sided model for the use case ofCSI compression (as in the model 102), various AI / ML model training collaborations may be considered.

[0037] Herein, the term “joint training” may mean that a generation model (e.g., an encoder 112) and a reconstruction model (e.g., a decoder 114) are each trained in a same loop for forward propagation and backward propagation. Joint training may be done either at a single network entity or across multiple network entities. For examplejoint training may occur through a gradient exchange between netw ork entities.

[0038] Herein, the term “separate training” may mean that sequential and / or parallel training is performed with respect to the generation model (e.g., an encoder 112) and the reconstruction model (e.g., a decoder 114). Sequential training may start with training at the UE and then proceed to training at the network side, or may start with training at the network side and then proceed to training at the UE side. In parallel training, both training at the UE side and training at the netw ork side are performed at the same time.

[0039] A first type of AI / ML model training collaboration may include joint training of a two-sided model at a single side / network entity. In such cases, the training can be either UE-sided or network-sided. FIG. 2 illustrates a diagram 200 showing a training 202 that occurs in a manner that is joint 204 and that is performed at one of the UE 106 or the network 110.

[0040] As another example, a second type of AI / ML model training collaboration may include joint training of the two-sided model both at the netw ork side and at the UE side, respectively. FIG. 3 illustrates a diagram 300 showing a training 302 that occurs in a manner that is joint 304 and across each of the UE 106 and the network 110 (each of the UE 106 and the network 110 are active participants in the training 302).

[0041] As another example, a third type of AI / ML model training collaboration may include separate training each of the UE 106 and the network 110. FIG. 4 illustrates a diagram 400 show ing that a first training 402 of the encoder 112 of the UE side 104 of the model 102 may occur at the UE 106, while a second training 404 of the decoder 114 of the network side 108 of the model 102 may occur at the netw ork 110.

[0042] FIG. 5 illustrates a flow diagram 500 show ing an example procedure betw een a network 502 and a UE 504 for training an AI / ML model for CSI feedback using joint training at the network.

[0043] The network 502 sends 506 a first channel state information reference signal (CSI- RS) to the UE 504. The UE 504 receives the first CSI-RS and uses it to compute precoder information (e g., to compute the eigenvectors of the channel).

[0044] The UE sends 508 the precoder information to the network 502 as training data, as illustrated. The network 502 uses the training data from the UE (and its own knowledge of the characteristics of the CSI-RS as sent) for the joint model training / generation 510, in which both a UE side (including, e.g., an encoder) and a network side (including, e.g., a decoder) of the AI / ML model is trained.

[0045] After the AI / ML model is trained at the network 502, the network 502 transfers 512 the UE side of the AI / ML model to the UE 504

[0046] With the UE side of the AI / ML model at the UE, the inference stage may now be used. The network 502 sends 514 the UE 504 a second CSI-RS. The UE 504 measures the channel and computes precoder information. The UE 504 then makes an encoder inference 516 by feeding the precoder information to the encoder of the UE side of the AI / ML model. The UE 504 then reports 518 the encoder results to network 502 in the form of compressed CSI, as shown.

[0047] The network 502 will then apply the compressed CSI to the decoder of the network side of the AI / ML model to generate a decoder inference 520 to recover the precoder information as was encoded at the UE 504 during the encoder inference 516.

[0048] In some cases, when performing joint training of a two-sided model at a single side / network entity as described herein, a mechanism for exchanging the AI / ML model between the network and the UE is used. A new data format / representation / procedure for signaling the AI / ML model as between the UE and the network may need to be specified. Further, the transfer of the AI / ML model between the entities will use air interface resources. Still further, when the model is outdated and retraining is needed, a new corresponding transfer will need to take place (a process which can become undesirably costly in terms of resources). Further, it may be that there are intellectual property issues corresponding to the AI / ML model that make allowing for this transfer between the UE and the network undesirable. Still further, a new data format / representation / procedure for signaling training data as between the UE and the network may need to be specified.

[0049] In other cases, when performing joint training of a two-sided model both at the network side and at the UE side, it may be that there are large forward propagation and backward propagation information exchange(s) over the air interface to facilitate the joint training across the network entities. Further, a new data format / representation / procedure for signaling this information as between the UE and the network may need to be specified.

[0050] Embodiments disclosed herein relate to an AI / ML model training collaboration procedure that uses separate and parallel training. In a such procedure, as compared to acase of joint training of a two-sided model at a single side / network entity, no training data transfer or model transfer is needed (these have been indicated with an “X” in the flow diagram 500 to show that these would not be needed in such a case).

[0051] Further, in such a procedure, as compared to a case of joint training of the two- sided model both at the network side and at the UE side, a backward propagation information exchange is not needed, thus saving air interface resources.

[0052] An Al / ML model training collaboration procedure that uses separate and parallel training, the training may be performed in an iterative manner in order to calibrate the AI / ML model.

[0053] FIG. 6 illustrates a flow diagram 600 for the training and use of Al / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network 602 (e.g., a base station of the network 602) and a UE 604, according to embodiments discussed herein.

[0054] The flow diagram 600 begins with a training phase 606. As part of the training phase 606, the network 602 sends 608 the UE 604 a CSI-RS. This CSI-RS is used by the UE 604 to generate a DL channel estimate 612 at the UE. In the embodiment of the flow diagram 600, the DL channel estimate 612 is used to determine precoder information in the form of eigenvectors 614 corresponding to a precoder for the channel. The eigenvectors 614 are used to train a UE AI / ML model 616 that includes a UE encoder 618 and a UE decoder 620 (note that in this case both are sited at the UE). As illustrated, the UE encoder 618 is configured to encode the eigenvectors 614 into a bitstream 622 that can be decoded by the UE decoder 620 back into the eigenvectors 614. The bitstream 622 may be smaller / compressed as compared to raw representations of the eigenvectors 614.

[0055] Continuing with the training phase 606: the UE 604 sends 610 the network 602 an uplink (UL) sounding reference signal (SRS). This UL SRS is used by the net ork 602 to generate a UL channel estimate 624 at the network. In the embodiment of the flow diagram 600, the UL channel estimate 624 is used to determine precoder information in the form of eigenvectors 626 corresponding to a precoder for the channel. The eigenvectors 626 are used to train a network AI / ML model 628 that includes a network encoder 630 and a network decoder 632 (note that in this case both are sited at the network 602). As illustrated, the network encoder 630 is configured to encode the eigenvectors 626 into a bitstream 634 that can be decoded by the network decoder 632 back into the eigenvectors 626. The bitstream 634 may be smaller / compressed as compared to raw representations of the eigenvectors 626.

[0056] The flow diagram 600 then proceeds to the model alignment phase 636. The purpose of the model alignment phase 636 is to align each of the UE AI / ME model 616 and the network AI / ML model 628, such that (at least) precoder information (eigenvectors) as encoded into a compressed bitstream by the UE encoder 618 of the UE AI / ML model 616 can be properly (or at least sufficiently accurately) decoded from the compressed bitstream by the network decoder 632 of the network AI / ML model 628.

[0057] As illustrated, the network 602 sends 638 a CSI-RS for DL channel estimation to the UE 604. The UE 604 may use this CSI-RS to generate a raw DL channel estimate (denoted Hm. in the flow diagram 600). The raw DL channel estimate is then used to determine precoder information (denoted WUE in the flow diagram 600). This precoder information represents a precoder that, if used by the base station, should cause the UE 604 to experience a desired effective DL channel (denoted Heff in the flow diagram 600) on DL transmissions using that precoder (e.g., Heff = HDLWUE). In the example of the flow diagram 600, this precoder information may be or include eigenvectors corresponding to the precoder for the channel.

[0058] The precoder information is then encoded by the UE encoder 618 into a bitstream of compressed CSI (this bitstream is also denoted PMIUE in the flow diagram 600). The UE 604 then sends 640 the bitstream to the network 602.

[0059] The network 602 then decodes precoder information from the bitstream using the network decoder 632 (precoder information as decoded at the network 602 is denoted WNW in the flow diagram 600). Note that due to differences in the (separately trained) UE AI / ML model 616 and the network AI / ML model 628, precoder information as decoded from the bitstream at the network 602 (WNW) may be different than the precoder information that was encoded into the bitstream at the UE 604 (WUE).

[0060] The network 602 then generates a beamformed CSI-RS based on the precoder information decoded from the bitstream (WNW) and sends 642 it to the UE 604. Upon receiving this beamformed CSI-RS, the UE 604 uses it to compute an effective DL channel that is perceived (an “actual effective DL channel,” denoted Heff in the flow diagram 600). The UE 604 then computes a similarity metric (e.g., generalized cosine similarity' (GCS)) between the desired effective DL channel (Heff) that was previously determined and the actual effective DL channel (Heff) that was determined through the use of the UE encoder 618 and the network decoder 632 and the sending 642 of the beamformed CSI-RS as has been described, in order to determine an error 644 (e.g., a “model mismatch,” as denoted in the flow diagram 600).

[0061] If the error 644 is above a threshold, the error 644 may be used to drive an update to the UE AI / ML model 61 at the UE side (e.g., an update 646 to the UE encoder 618 of the UE AI / ML model 616, as illustrated) to reduce the error 644, thereby increasing the model alignment between the UE AI / ML model 616 and the network AI / ML model 628.

[0062] After the UE AI / ML model 616 has been updated at the UE 604, various functions previously described may be repeated. For example, the UE encodes the precoder information into a bitstream using the UE encoder 618. This bitstream may be different than that previously sent, due to changes to the UE encoder 618 of the UE AI / ML model 616 from its prior state in order to compensate for the prior error 644, as just described. The UE 604 then sends 648 this (new) bitstream to the network 602. The network then (again) decodes the bitstream into a (new) network-side precoder information WNW and uses this new precoder information to send another beamformed CSI-RS to the UE 604, in the manner that has been described. The UE can then compute a (new) actual effective DL channel Heff and compare it to the (original) desired effective DL channel FLff to calculate a (new) error 644.

[0063] While the error 644 remains above a threshold (e.g., some small value), another update 646 corresponding to the error 644 may be carried out with respect to the UE encoder 618 of the UE AI / ML model 616, and further repetitions such as that just described may be performed, in an attempt to ultimately drive the error 644 to within the threshold.

[0064] If at any point the error 644 falls within the threshold, the UE 604 may determine that the UE encoder 618 of the UE AI / ML model 616 is accurate (e.g., is aligned w ith the network decoder 632 of the network AI / ML model 628, such that a bitstream encoded at the UE 604 by the UE encoder 618 can be decoded by the network decoder 632 of the network AI / ML model 628 with acceptable accuracy) and drops any upcoming repetitions. Further, the UE 604 sends 650 an indication to the network 602 that the UE encoder 618 of the UE AI / ML model 616 is accurate, such that the network does not expect such future repetitions for model calibration purposes.

[0065] The flow diagram 600 then proceeds to the operation phase 652. In the operation phase, the network 602 sends 654 a CSI-RS to the UE 604. The sending of this CSI-RS may be for a substantive channel sounding (that is for purposes other than model alignment).

[0066] The UE 604 performs a channel estimation using this CSI-RS and calculates corresponding precoder information. This precoder information is then encoded into a bitstream using the UE encoder 618 of the UE AI / ML model 616. The UE 604 then sends 656 the bitstream to the network 602, which uses the network decoder 632 of the networkAI / ML model 628 to decode it back into precoder information. Because of the prior operations of the model alignment phase 636 as described, this precoder information as understood at each of the UE 604 and the network 602 are in alignment (e.g., are the same, or at least are similar within an acceptable accuracy), such that the substantive channel sounding is itself accurate.

[0067] FIG. 7 illustrates a flow diagram 700 for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network 702 (e.g., a base station of the network 702) and a UE 704, according to embodiments discussed herein.

[0068] The flow diagram 700 begins with the training phase 606. The training phase 606 and related elements thereof as illustrated in the flow diagram 700 may be as explained in relation to the flow diagram 600.

[0069] The flow diagram 700 then proceeds to the model alignment phase 706. The purpose of the model alignment phase 706 is to align each of the UE AI / ML model 616 and the network AI / ML model 628, such that precoder information (eigenvectors) as encoded into a compressed bitstream by the UE encoder 618 of the UE AI / ML model 616 can be properly decoded from the compressed bitstream by the network decoder 632 of the network AI / ML model 628.

[0070] As illustrated, the network 702 sends 708 a CSI-RS for DL channel estimation to the UE 704. The result may be a raw DL channel estimate (denoted HDL in the flow diagram 700). The raw DL channel estimate is then used to determine precoder information (denoted WUE in the flow diagram 700). This precoder information represents a precoder that, if used by the base station, should cause the UE 704 to experience a desired effective DL channel (denoted Herr in the flow diagram 700) on DL transmissions using that precoder (e.g., Herr = HDLWUE). In the example of the flow diagram 700, this precoder information may be or include eigenvectors corresponding to the precoder for the channel.

[0071] The precoder information is then encoded by the UE encoder 618 into a bitstream of compressed CSI (this bitstream is also denoted PMIUE in the flow diagram 700). The UE 704 then sends 710 the bitstream to the network 702.

[0072] The network 702 then decodes precoder information from the bitstream using the network decoder 632 (precoder information as decoded at the network 702 is denoted WNW in the flow diagram 700). Note that due to differences in the (separately trained) UE AI / ML model 61 and the network AI / ML model 628, precoder information as decoded from thebitstream at the network 702 (WNW) may be different than the precoder information that was encoded into the bitstream at the UE 704 (WUE).

[0073] The network 602 then generates a beamformed CSI-RS based on the precoder information decoded from the bitstream and sends 712 it to the UE 704. Upon receiving this beamformed CSI-RS, the UE 704 uses it to compute actual effective DL channel that is perceived (Heff in the flow diagram 700). The UE 704 then computes a similarity metric (e.g., GCS) between the desired effective DL channel (Heff) that was previously determined and the actual effective DL channel (Heff) that was determined through the use of the UE encoder 618 and the network decoder 632 and the sending 712 of the beamformed CSI-RS, as has been described in order to determine an error 714.

[0074] If the error is above a threshold (e.g.. some small value), the error 714 may be encoded using an error encoder 716 to generate an error code 718 corresponding to the error 714. Note that the error encoder 716 may be an encoder that is other than the UE encoder 618 of the UE AI / ML model 616. It is contemplated that the error encoder 716 may be, in some cases, an encoder of another AI / ML model (other than the UE AI / ML model 616) that is in use at the UE.

[0075] The UE 704 then sends 720 the error code 718 to the network 702. The error code 718 may be used to drive an update 722 to the network AI / ML model 628 at the network side (e.g., an update 722 to the network decoder 632 of the network AI / ML model 628, as illustrated) to reduce the error 714, thereby increasing the model alignment between the UE AI / ML model 616 and the network AI / ML model 628.

[0076] After the network AI / ML model 628 has been updated at the network 702, various functions may be repeated. For example, the network 702 may again decode the (originally received) bitstream using the network decoder 632. The resulting precoder information WNW may be different than the prior WNW due to changes to the network decoder 632 of the network AI / ML model 628 from its prior state as driven by the error code 718, as just described. The network 702 then generates and sends a (new) beamformed CSI-RS according to this new WNW. Using this new beamformed CSI-RS, the UE computes a (new) actual effective DL channel Heff and compares it to the (original) desired effective DL channel Heff to calculate a (new) error 714. If the error 714 remains above the threshold, the error encoder 716 may be used to generate another error code 718 that is sent to the network 702 to drive another update 722 to the network AI / ML model 628.

[0077] Note that while the error 714 remains above the threshold, further repetitions such as that just described may be performed, in an attempt to ultimately drive the error 714 to within the threshold.

[0078] If at any point the error 714 falls within the threshold, the UE 704 may determine that the network decoder 632 of the network AI / ML model 628 is accurate (e.g., is aligned with the UE encoder 618 of the UE AI / ML model 616, such that a bitstream encoded at the UE 704 by the UE encoder 618 can be decoded by the network decoder 632 of the network AI / ML model 628 with acceptable accuracy) and drops any upcoming repetitions. Further, the UE 704 sends 724 an indication to the network 702 that the UE encoder 618 of the UE AI / ML model 616 is accurate, such that the network does not expect such future repetitions for model calibration purposes.

[0079] The flow diagram 700 then proceeds to the operation phase 652. The operation phase 652 may proceed as explained in relation to the flow diagram 600.

[0080] FIG. 8 illustrates a flow diagram 800 for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network 802 (e.g., a base station of the network 802) and a UE 804, according to embodiments discussed herein.

[0081] The flow diagram 800 begins with the training phase 606. The training phase 606 and related elements thereof as illustrated in the flow diagram 800 may be as explained in relation to the flow diagram 600.

[0082] The flow diagram 800 then proceeds to the model alignment phase 806. The purpose of the model alignment phase 806 is to align each of the UE AI / ML model 616 and the network AI / ML model 628, such that precoder information (eigenvectors) as encoded into a compressed bitstream by the UE encoder 618 of the UE AI / ML model 616 can be properly decoded from the compressed bitstream by the network decoder 632 of the network AI / ML model 628.

[0083] As illustrated, the network 802 sends 808 a CSI-RS for DL channel estimation to the UE 804. The result may be a raw DL channel estimate (denoted HDL in the flow diagram 800). The raw DL channel estimate is then used to determine precoder information (denoted WUE in the flow diagram 800). This precoder information represents a precoder that, if used by the base station, should cause the UE 804 to experience a desired effective DL channel (denoted Heff in the flow diagram 800) on DL transmissions using that precoder (e.g., Heff = HDLWUE). In the example of the flow diagram 800, this precoder information may be or include eigenvectors corresponding to the precoder for the channel.

[0084] The precoder information is then encoded by the UE encoder 618 into a bitstream of compressed CSI (this bitstream is also denoted PMIUE in the flow diagram 800). The UE 804 then sends 810 the bitstream to the network 802.

[0085] The network 802 then decodes precoder information from the bitstream using the network decoder 632 (precoder information as decoded at the network 802 is denoted WNW in the flow diagram 800). Note that due to differences in the (separately trained) UE AI / ML model 616 and the network AI / ML model 628, precoder information as decoded from the bitstream at the network 802 (WNW) may be different than the precoder information that was encoded into the bitstream at the UE 804 (WUE).

[0086] The network 602 then generates a beamformed CSI-RS based on the precoder information decoded from the bitstream and sends 812 it to the UE 804. Upon receiving this beamformed CSI-RS, the UE 804 uses it to compute an actual effective DL channel that is perceived (Hcff in the flow diagram 800). The UE 804 then computes a similarity metric (e.g., GCS) between the desired effective DL channel (Herr) that was previously determined and the actual effective DL channel (Heff) that was determined through the use of the UE encoder 618 and the network decoder 632 and the sending 812 of the beamformed CSI-RS, as has been described in order to determine an error 814.

[0087] If the error 814 is above a threshold (e.g., some small value), the error 814 may be used to drive an update to the UE AI / ML model 616 at the UE side (e.g., an update 816 to the UE encoder 618 of the UE AI / ML model 616, as illustrated) to reduce the error 814, thereby increasing the model alignment between the UE AI / ML model 616 and the network AI / ML model 628.

[0088] Further, the error 814 may also be encoded using an error encoder 818 to generate an error code 820 corresponding to the error 814. Note that the error encoder 818 may be an encoder that is other than the UE encoder 618 of the UE AI / ML model 616. It is contemplated that the error encoder 818 may be, in some cases, an encoder of another AI / ML model (other than the UE AI / ML model 616) that is in use at the UE 804.

[0089] The UE 804 then sends 822 the error code 820 to the network 802. The error code 820 may be used to drive an update 824 to the network AI / ML model 628 at the network side (e.g., an update 824 to the network decoder 632 of the network AI / ML model 628, as illustrated) to reduce the error 814, thereby increasing the model alignment between the UE AI / ML model 616 and the network AI / ML model 628.

[0090] After the UE encoder 618 has been updated at the UE AI / ML model 616 and the network AI / ML model 628 has been updated at the network AI / ML model 628, variousfunctions may be repeated. For example, the UE 804 encodes the precoder information into a bitstream using the UE encoder 618. This bitstream may be different than that previously sent, due to changes to the UE encoder 618 of the UE AI / ML model 616 from its prior state in order to compensate for the prior error 814. as just described. The UE 804 then sends 826 this (new) bitstream to the network 802.

[0091] The network then decodes the (new) bitstream into a (new) network-side precoder information WNW (using its updated network decoder 632) and uses this new precoder information to send another beamformed CSI-RS to the UE 804. in the manner that has been described. The UE 804 can then compute a (new) actual effective DL channel Heff and compare it to the (original) desired effective DL channel Heff to calculate a (new) error 814.

[0092] While the error 814 remains above the threshold, another update 816 corresponding to the error 814 may be carried out with respect to the UE encoder 618 of the UE AI / ML model 616, and another error code 820 may be generated by the error encoder 818 using the error 814 and sent to the network 802 in order to drive another update 824 of the network decoder 632 of the network AI / ML model 628. Then further repetitions such as that just described may be performed, in an attempt to ultimately drive the error 814 to within the threshold.

[0093] If at any point the error 814 falls within the threshold, the UE 804 may determine that each of the UE encoder 618 of the UE AI / ML model 616 and the network decoder 632 of the network AI / ML model 628 are accurate (e.g., are aligned such that a bitstream encoded at the UE 804 by the UE encoder 618 of the UE AI / ML model 616 can be decoded at the network 802 by the network decoder 632 of the network AI / ML model 628 with acceptable accuracy) and drops any upcoming repetitions. Further, the UE 804 sends 828 an indication to the network 802 that the UE encoder 618 of the UE AI / ML model 616 and / or the network decoder 632 of the network AI / ML model 628 is / are accurate, such that the network does not expect such future repetitions for model calibration purposes.

[0094] The flow diagram 800 then proceeds to the operation phase 652. The operation phase 652 may proceed as explained in relation to the flow diagram 600.

[0095] FIG. 9 illustrates a flow diagram 900 for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network 902 (e.g., a base station of the network 902) and a UE 904, according to embodiments discussed herein.

[0096] The flow diagram 900 begins with a training phase 906. As part of the training phase 906, the network 902 sends 908 the UE 904 a CSI-RS. This CSI-RS is used by the UE904 to generate a DL channel estimate 912 at the UE 904. The DL channel estimate 912 is used to train a UE AI / ML model 914 that includes a UE encoder 916 and a UE decoder 918 (note that in this case both are sited at the UE). As illustrated, the UE encoder 916 is configured to encode DL channel estimate 912 into a bitstream 920 that can be decoded by the UE decoder 918 back into DL channel estimate 912. The bitstream 920 may be smaller / compressed as compared to a raw representation of the DL channel estimate 912.

[0097] Continuing with the training phase 906: the UE 904 sends 910 the network 902 an UL SRS. This UL SRS is used by the network 902 to generate a UL channel estimate 922 at the network. In the embodiment of the flow diagram 900, the UL channel estimate 922 is used to train a network AI / ML model 924 that includes a network encoder 926 and a network decoder 928 (note that in this case both are sited at the network). As illustrated, the network encoder 926 is configured to encode the UL channel estimate 922 into a bitstream 920 that can be decoded by the network decoder 928 back into the UL channel estimate 922. The bitstream 930 may be smaller / compressed as compared to a raw representation of the UL channel estimate 922.

[0098] The flow diagram 900 then proceeds to the model alignment phase 932. The purpose of the model alignment phase 932 is to align each of the UE AI / ML model 914 and the network AI / ML model 924, such that a raw DL channel estimate as encoded into a compressed bitstream by the UE encoder 916 of the UE AI / ML model 914 can be properly decoded from the compressed bitstream by the network decoder 928 of the network AI / ML model 924.

[0099] As illustrated, the network 902 sends 934 a CSI-RS for DL channel estimation to the UE 904. The UE 904 may use this CSI-RS to generate a raw DL channel estimate (denoted HDL in the flow diagram 900). The raw DL channel estimate is then used to determine precoder information (denoted WUE in the flow diagram 900). This precoder information represents a precoder that, if used by the base station, should cause the UE to experience a desired effective DL channel (denoted Heff in the flow diagram 900) on DL transmissions using that precoder (e.g., Heff = HDLWUE). In the example of the flow diagram 900, this precoder information may be or include eigenvectors corresponding to the precoder for the channel.

[0100] The raw DL channel estimate is then encoded by the UE encoder 916 into a bitstream of compressed CSI (this bitstream is also denoted PMIUE in the flow diagram 900). The UE 904 then sends 936 the bitstream to the network 902.

[0101] The network 902 then decodes a raw DL channel estimate from the bitstream using the network decoder 928 (a raw DL channel estimate as decoded at the network 902 is denoted HNW in the flow diagram 900). Note that due to differences in the (separately trained) UE AI / ML model 914 and the network AI / ML model 924, the raw DL channel estimate as decoded from the bitstream at the network 902 (HNW) may be different than the raw DL channel estimate that was encoded into the bitstream at the UE 904 (HDL).

[0102] The raw DL channel estimate at the network 902 (HNW) is then used by the network 902 to determine precoder information (denoted WNW in the flow diagram 900). This precoder information represents a precoder that, if used by the base station, should cause the UE 904 to experience a desired effective DL channel on DL transmissions using that precoder, assuming that HNW accurately reflects the channel conditions from the UE perspective (e.g., assuming that HNW is equal to or at least near HDL). In the example of the flow diagram 900, this precoder information may be or include eigenvectors corresponding to the precoder.

[0103] The network 902 then generates a beamformed CSI-RS based on the precoder information (WNW) and sends 938 it to the UE 904. Upon receiving this beamformed CSI- RS, the UE 904 uses it to compute an actual effective DL channel (denoted Heff in the flow diagram 900). The UE 904 then computes a similarity metric (e.g., a GCS) between the desired effective DL channel (Heft) that was previously determined and the actual effective DL channel (Heff) that was determined through the use of the UE encoder 916 and the network decoder 928 and the sending 938 of the beamformed CSI-RS, as has been described in order to determine an error 940 (e.g., a "model mismatch,” as denoted in the flow diagram 900).

[0104] If the error 940 is above a threshold, the error 940 may be used to drive an update to the UE AI / ML model 914 at the UE side (e.g., an update 942 to the UE encoder 916 of the UE AI / ML model 914, as illustrated) to reduce the error 940, thereby increasing the model alignment between the UE AI / ML model 914 and the network AI / ML model 924.

[0105] After the UE AI / ML model 914 has been updated at the UE 904, various functions previously described may be repeated. For example, the UE encodes the raw DL channel estimate into a bitstream using the UE encoder 916. This bitstream may be different than that previously sent, due to changes to the UE encoder 916 of the UE AI / ML model 914 from its prior state in order to compensate for the prior error 940, as just described. The UE 904 then sends 944 this (new) bitstream to the network 602. The network then (again) decodes the bitstream into a (new) network-side raw DL channel estimate HNW and uses it toderive (new) network-side precoder information WNW. This new network-side precoder information is used to send another beamformed CSI-RS to the UE 904, in the manner that has been described. The UE can then compute an (new) actual effective DL channel Heff and compare it to the (original) desired effective DL channel Heff to calculate a (new) error 940.

[0106] While the error 940 remains above a threshold (e.g., some small value), another update 942 corresponding to the error 940 may be carried out with respect to the UE encoder 916 of the UE AI / ML model 914, and further repetitions such as that just described may be performed, in an attempt to ultimately drive the error 940 to within the threshold.

[0107] If at any point the error 940 falls within the threshold, the UE 904 may determine that the UE encoder 916 of the UE AI / ML model 914 is accurate (e.g., is aligned with the network decoder 928 of the network AI / ML model 924, such that a bitstream encoded at the UE 904 by the UE encoder 916 can be decoded by the network decoder 928 of the network AI / ML model 924 with acceptable accuracy) and drops any upcoming repetitions. Further, the UE 904 sends 946 an indication to the network 902 that the UE encoder 916 of the UE AI / ML model 914 is accurate, such that the network does not expect such future repetitions for model calibration purposes.

[0108] The flow diagram 900 then proceeds to the operation phase 948. In the operation phase, the network 902 sends 950 a CSI-RS to the UE 904. The sending of this CSI-RS may be for a substantive channel sounding (that is for purposes other than model alignment).

[0109] The UE 904 performs a raw DL channel estimation using this CSI-RS and encodes it into a bitstream using the UE encoder 916 of the UE AI / ML model 914. The UE 904 then sends 952 the bitstream to the network 902, which uses the network decoder 928 of the network AI / ML model 924 to decode it back into a raw DL channel estimate (which may then be used to derive precoder information used for DL transmissions). Because of the prior operations of the model alignment phase 932 as described, this raw DL channel estimate as understood at each of the UE 904 and the network 902 are in alignment (e.g., are the same, or at least are similar within an acceptable accuracy), such that the substantive channel sounding is itself accurate.

[0110] FIG. 10 illustrates a flow diagram 1000 for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between the network 1002 (e.g., a base station of the network 1002) and the UE 1004, according to embodiments discussed herein.

[0111] The flow diagram 1000 begins with the training phase 906. The training phase 906 and related elements thereof as illustrated in the flow diagram 1000 may be as explained in relation to the flow diagram 900.

[0112] The flow diagram 1000 then proceeds to the model alignment phase 1006. The purpose of the model alignment phase 1006 is to align each of the UE Al / ME model 914 and the network AI / ML model 924, such that a raw DL channel estimate as encoded into a compressed bitstream by the UE encoder 916 of the UE AI / ML model 914 can be properly decoded from the compressed bitstream by the network decoder 928 of the network AI / ML model 924.

[0113] As illustrated, the network 1002 sends 1008 a CSI-RS for DL channel estimation to the UE 704. The result may be a raw DL channel estimate (denoted HDL in the flow diagram 1000). The raw DL channel estimate is then used to determine precoder information (denoted WUE in the flow diagram 1000). This precoder information represents a precoder that, if used by the base station, should cause the UE 1004 to experience a desired effective DL channel (denoted Heff in the flow diagram 1000) on DL transmissions using that precoder (e.g., Heff = HDLWUE). In the example of the flow diagram 1000, this precoder information may be or include eigenvectors corresponding to the precoder for the channel.

[0114] The raw DL channel estimate is then encoded by the UE encoder 916 into a bitstream of compressed CSI (this bitstream is also denoted PMIUE in the flow diagram 1000). The UE 1004 then sends 1010 the bitstream to the network 1002.

[0115] The network 1002 then decodes a raw DL channel estimate from the bitstream using the network decoder 928 (a raw DL channel estimate as decoded at the network 1002 is denoted HNW in the flow diagram 1000). Note that due to differences in the (separately trained) UE AI / ML model 914 and the network AI / ML model 924, the raw DL channel estimate as decoded from the bitstream at the network 1002 (HNW) may be different than the raw DL channel estimate that was encoded into the bitstream at the UE 1004 (HDL).

[0116] The raw DL channel estimate at the network 1002 (HNW) is then used at the network 1002 to determine precoder information (denoted WNW in the flow diagram 1000). This precoder information represents a precoder that, if used by the base station, should cause the UE 1004 to experience a desired effective DL channel on DL transmissions using that precoder, assuming that HNW accurately reflects the channel conditions from the UE perspective (e.g., assuming that HNW is equal to or at least near HDL). In the example of the flow diagram 1000, this precoder information may be or include eigenvectors corresponding to the precoder.

[0117] The network 1002 then generates a beamformed CSI-RS based on the precoder information (WNW) and sends 1012 it to the UE 1004. Upon receiving this beamformed CSI- RS, the UE 1004 uses it to compute an effective DL channel that is perceived (an “actual effective DE channel,” denoted Heff in the flow diagram 1000). The UE 1004 then computes a similarity- metric (e.g., GCS) between the desired effective DL channel (Heft) that was previously determined and the actual effective DL channel (Heff) that was determined through the use of the UE encoder 916 and the network decoder 928 and the sending 1012 of the beamformed CSI-RS, as has been described in order to determine an error 1014.

[0118] If the error 1014 is above a threshold (e.g., some small value), the error 1014 maybe encoded using an error encoder 1016 to generate an error code 1018 corresponding to the error 1014. Note that the error encoder 1016 may be an encoder that is other than the UE encoder 916 of the UE AI / ML model 914. It is contemplated that the error encoder 1016 may be, in some cases, an encoder of another AI / ML model (other than the UE AI / ML model 914) that is in use at the UE.

[0119] The UE 1004 then sends 1020 the error code 1018 to the network 1002. The error code 1018 may be used to drive an update 1022 to the network AI / ML model 924 at the network side (e.g., an update 1022 to the network decoder 928 of the network AI / ML model 924, as illustrated) to reduce the error 1014, thereby increasing the model alignment between the UE AI / ML model 914 and the network AI / ML model 924.

[0120] After the network AI / ML model 924 has been updated at the network 1002, various functions may be repeated. For example, the network 1002 may again decode the (originally- received) bitstream using the network decoder 928. The resulting raw DL channel estimate HNW may be different than the prior HNW due to changes to the network decoder 928 of the network AI / ML model 924 from its prior state as driven by the error code 1018, as just described. Using this (new) HNW, the network 1002 determines a new network-side precoder information WNW. The network 1002 then generates and sends a (new) beamformed CSI-RS according to this new WNW. Using this new beamformed CSI-RS, the UE computes a (new) actual effective DL channel Heff and compares it to the (original) desired effective DL channel Heff to calculate a (new) error 1014.

[0121] Note that while the error 1014 remains above the threshold, further repetitions such as that just described may be performed in an attempt to ultimately drive the error 1014 to within the threshold.

[0122] If at any point the error 1014 falls within the threshold, the UE 1004 may determine that the network decoder 928 of the network AI / ML model 924 is accurate (e.g., is alignedwith the UE encoder 916 of the UE AI / ML model 914, such that a bitstream encoded at the UE 1004 by the UE encoder 916 can be decoded by the network decoder 928 of the network AI / ML model 924 with acceptable accuracy) and drops any upcoming repetitions. Further, the UE 1004 sends 1024 an indication to the network 1002 that the UE encoder 916 of the UE AI / ML model 914 is accurate, such that the network does not expect such future repetitions for model calibration purposes.

[0123] The flow diagram 1000 then proceeds to the operation phase 948. The operation phase 652 may proceed as explained in relation to the flow diagram 900.

[0124] FIG. 11 illustrates a flow diagram 1100 for the training and use of AI / ML models for CSI compression, where the AI / ML models are trained using a collaboration procedure that uses separate training as between a network 1102 (e.g., a base station of the network 1102) and a UE 1104, according to embodiments discussed herein.

[0125] The flow diagram 1100 begins with the training phase 906. The training phase 906 and related elements thereof as illustrated in the flow diagram 1100 may be as explained in relation to the flow diagram 900.

[0126] The flow diagram 1100 then proceeds to the model alignment phase 1106. The purpose of the model alignment phase 1106 is to align each of the UE AI / ML model 914 and the network AI / ML model 924, such that a raw DL channel estimate as encoded into a compressed bitstream by the UE encoder 916 of the UE AI / ML model 914 can be properly decoded from the compressed bitstream by the network decoder 928 of the network AI / ML model 924.

[0127] As illustrated, the network 1102 sends 1108 a CSLRS for DL channel estimation to the UE 1104. The result may be a raw DL channel estimate (denoted HDL in the flow diagram 1100). The raw DL channel estimate is then used to determine precoder information (denoted WUE in the flow diagram 1100). This precoder information represents a precoder that, if used by the base station, should cause the UE 1104 to experience a desired effective DL channel (denoted Heff in the flow diagram 1100) on DL transmissions using that precoder (e g., Heff = HDLWUE). In the example of the flow diagram 1100, this precoder information may be or include eigenvectors corresponding to the precoder for the channel.

[0128] The raw DL channel estimate is then encoded by the UE encoder 916 into a bitstream of compressed CSI (this bitstream is also denoted PMIUE in the flow diagram 1 100). The UE 1104 then sends 11 10 the bitstream to the network 902.

[0129] The network 1102 then decodes a raw DL channel estimate from the bitstream using the network decoder 928 (a raw DL channel estimate as decoded at the network 1102 is denoted HNW in the flow diagram 1100). Note that due to differences in the (separately trained) UE AI / ML model 914 and the network AI / ML model 924, the raw DL channel estimate as decoded from the bitstream at the network 1102 (HNW) may be different than the raw DL channel estimate that was encoded into the bitstream at the UE 1104 (HDI ).

[0130] The raw DL channel estimate at the network (HNW) is then used at the network 1102 to determine precoder information (denoted WNW in the flow diagram 1100). This precoder information represents a precoder that, if used by the base station, should cause the UE 1104 to experience a desired effective DL channel on DL transmissions using that precoder, assuming that HNW accurately reflects the channel conditions from the perspective of the UE 1104 (e.g., assuming that HNW is equal to or at least near HDL). In the example of the flow diagram 1100, this precoder information may be or include eigenvectors corresponding to the precoder.

[0131] The network 1102 then generates a beamformed CSI-RS based on the precoder information (WNW) and sends 1112 it to the UE 1104. Upon receiving this beamformed CSI- RS, the UE 1104 uses it to compute an effective DL channel that is perceived (an “actual effective DL channel,” denoted Heir in the flow diagram 1100). The UE 1104 then computes a similarity metric (e.g., GCS) between the desired effective DL channel (Heft) that was previously determined and the actual effective DL channel (Heff) that was determined through the use of the UE encoder 916 and the network decoder 928 and the sending 1 112 of the beamformed CSI-RS, as has been described in order to determine an error 1114.

[0132] If the error 1114 is above a threshold (e.g.. some small value), the error 1114 maybe used to drive an update to the UE AI / ML model 914 at the UE side (e.g., an update 1116 to the UE encoder 916 of the UE AI / ML model 914, as illustrated) to reduce the error 1114, thereby increasing the model alignment between the UE AI / ML model 914 and the network AI / ML model 924.

[0133] Further, the error 1114 may also be encoded using an error encoder 1118 to generate an error code 1120 corresponding to the error 1114. Note that the error encoder 1118 may be an encoder that is other than the UE encoder 916 of the UE AI / ML model 914. It is contemplated that the error encoder 1118 may be, in some cases, an encoder of another AI / ML model (other than the UE AI / ML model 914) that is in use at the UE 1104.

[0134] The UE 1104 then sends 1122 the error code 1120 to the network 1102. The error code 1120 may be used to drive an update 1124 to the network AI / ML model 924 at thenetwork side (e.g., an update 1124 to the network decoder 928 of the network AI / ML model 924, as illustrated) to reduce the error 1114, thereby increasing the model alignment between the UE AI / ML model 914 and the network AI / ML model 924.

[0135] After the UE encoder 916 has been updated at the UE AI / ML model 914 and the network AI / ML model 924 has been updated at the network AI / ML model 924, various functions may be repeated. For example, the UE 1104 encodes the precoder information into a bitstream using the UE encoder 916. This bitstream may be different than that previously sent, due to changes to the UE encoder 916 of the UE AI / ML model 914 from its prior state, in order to compensate for the prior error 1114, as just described. The UE 1104 then sends 1126 this (new) bitstream to the network 1102.

[0136] The network then decodes the (new) bitstream into a (new) network-side precoder information WNW (using its updated network decoder 928) and uses this new precoder information to send another beamformed CSI-RS to the UE 1104, in the manner that has been described. The UE 1104 can then compute a (new) actual effective DL channel FLff and compare it to the (original) desired effective DL channel Heff to calculate a (new) error 1114.

[0137] While the error 1114 remains above the threshold, another update 1116 corresponding to the error 1114 may be carried out with respect to the UE encoder 916 of the UE AI / ML model 914, another error code 1120 may be generated by the error encoder 1118 using the error 1114 and sent to the network 1102 in order to drive another update 1124 of the network decoder 928 of the network AI / ML model 924. Then, further repetitions such as that just described may be performed in an attempt to ultimately drive the error 1114 to within the threshold.

[0138] If at any point the error 1114 falls within the threshold, the UE 1104 may determine that each of the UE encoder 916 of the UE AI / ML model 914 and the network decoder 928 of the network AI / ML model 924 are accurate (e.g., are aligned such that a bitstream encoded at the UE 1104 by the UE encoder 916 can be decoded at the network 1102 by the network decoder 928 of the network AI / ML model 924 with acceptable accuracy) and drops any upcoming repetitions. Further, the UE 1104 sends 1128 an indication to the network 1102 that the UE encoder 916 of the UE AI / ML model 914 and / or the network decoder 928 of the network AI / ML model 924 is / are accurate, such that the network 1102 does not expect such future repetitions for model calibration purposes.

[0139] The flow diagram 1100 then proceeds to the operation phase 948. The operation phase 948 may proceed as explained in relation to the flow diagram 900.

[0140] In some wireless communication systems, an output of a CSI generation model can be explicit feedback (e.g., feeding back a compressed version of raw DL channel information. Another output of a CSI generation model can be implicit feedback (e.g., reusing or modifying a rank indication (RI) / channel quality index (CQI)Zprecoding matrix indicator (PMI) framework). In some cases, the output of a CSI generation model can be a combination of the two. Feeding back a compressed version of the raw DL channel information can enable optimal feedback report quality7.

[0141] In some cases, there are benefits of transmitting RI and CQI feedback in addition to the raw DL channel information. For example, for purposes of providing a base station a more complete understanding of the channel conditions at the UE, it may be beneficial to report both the channel and some metric of the interference and noise.

[0142] A CQI calculation uses knowledge of the interference and / or noise experienced on the channel and may not be derived solely from the raw DL channel information. For example, the CQI value can be obtained from a combination of channel measurement on CSI-RS and interference measurements on a channel state information interference measurement (CSI-IM) resource. In some cases, the output of a network-side AI / ML model (e.g., a decoder) may not perfectly match with the input of the UE-side AI / ML model (e.g., an encoder) (e.g., as is discussed herein). In such cases, a reported CQI value may be irrelevant and / or misunderstood by the base station, which could ultimately lead to performance degradation. In cases where the UE has AI / ML decoder information, it could detect the mismatch and it could possibly mitigate it by correcting the reported CQI value.

[0143] For purposes of providing the UE with RI and / or CQI information, it is contemplated that further enhancements to the CSI based AI / ML models discussed herein. In such cases, it is contemplated that CQI and / or RI information could be encoded into a bitstream along with the compressed CSI. Such an encoding can take place when the two- sided AI / ML models at each of the UE and the network have been aligned or calibrated (e.g., as is discussed herein). In some cases, a model alignment flag (an accuracy indication) will mark the start of the CSI / CQI / RI reporting in the bitstream.

[0144] In some embodiments, for CSI / CQI / RI reporting for compressed CSI based AI / ML models, a relative reduction in UE computations may be realized.

[0145] For example, consider a first case where the UE completes channel estimates and uses the channel estimates as an input to AI / ML models to encode / feedback CSI based on the channel estimates. The UE may then also complete multiple computations to estimate channel capacities and / or mutual information in order to find an optimum CQI / RIcorresponding to the channel estimate. The UE may then report CQI / RI to the network separately from the encoded CSI.

[0146] FIG. 12 illustrates a diagram 1200 showing an explicit calculation of CQI / RI as may be carried out by a UE. First, the UE calculates each of a Gaussian capacity 1202 and a substream capacity 1204 based on an (actual) effective DL channel Heff experienced at the UE. Then, the UE performs non-linear mapping 1206 to generate mutual information 1208 (e.g., capacity and / or spectral efficiency information (e.g., in bits / sec / hertz (Hz)). The mutual information 1208 is fed to a look up table (LUT) 1210 which is used to finally identify the CQI / RI 1212.

[0147] Relative to this case, UE computations can be significantly reduced at the UE if the UE instead completes raw DL channel estimates and / or precoder information and uses the raw' DL channel estimates and / or precoder information as an input to an AI / ML model that is trained to encode / feedback CSI / CQI / RI together. This use avoids the need for explicit UE computations to estimate channel capacities and / or mutual information (e.g., as described in relation to the diagram 1200).

[0148] In some embodiments, raw' DL channel estimates and / or precoder information at the UE is whitened to remove interference. The raw' DL channel estimates and / or precoder information may also be scaled by a signal to noise ratio (SNR) to more fully reflect noise conditions. In some cases, one of either whitening or scaling of the channel estimates may occur, while in other cases both whitening and scaling of the channel estimates may occur.

[0149] An AI / ML model at the UE (e.g., as is discussed herein) can be enhanced to estimate CQI / RI based on additionally training with CQI / RI in training data (e.g., as calculated, based on a raw DL channel estimations that are also being used for the training). For example, the complicated capacity' expressions could be learned directly based on inputted channel estimation information by the AI / ML. This training may cause an encoder of such a model to generate CQI / RI based on a raw DL channel estimate (and / or. by extension, precoder information corresponding to such a raw' DL channel estimate).

[0150] FIG. 13 illustrates a diagram 1300 for using an encoder 1308 of an AI / ML model at a UE 1302 to generate CQI / RI corresponding to input, according to embodiments herein. First, the UE 1302 calculates a DL channel estimate 1306. The UE may then perform whitening and / or SNR scaling 1310 on this DL channel estimate 1306 (or, in other embodiments, precoder information calculated at the UE based on the DL channel estimate 1306 may be used). The result is then provided to an AI / ML model that is trained to use an encoder 1308 to generate both compressed CSI 1312 (e.g., as is discussed herein) andpredict, based on the training, a corresponding CQI / RI indication 1314. The encoder 1308 may encode the compressed CSI 1312 into a bitstream 1316 (e.g., as has been described herein). The encoder 1308 may further add the corresponding CQI / RI indication 1314 to the bitstream 1316, as illustrated (e.g., in a non-compressed form).

[0151] Once the bitstream 1316 arrives at the network 1304, a decoder 1318 may be used to decode the compressed CSI 1312 back into a DL channel estimate (or back into corresponding precoder information, as the case may be), according to procedures that are discussed elsewhere herein. Further, the network 1304 may also receive the (e.g., noncompressed) CQI / RI indication 1314 from the bitstream 1316 such that it is informed of a corresponding CQI / RI.

[0152] FIG. 14 illustrates diagram 1400 showing an example of reporting CSI / CQI / RI after AI / ML model alignment between a network 1402 and a UE 1404, according to embodiments described herein. The particular AI / ML models used at each of the network 1402 and the UE 1404, and the alignment procedures used to align those models, may be described in any of the flow diagram 600, the flow diagram 700, the flow diagram 800, the flow diagram 900, the flow diagram 1000, and / or the flow diagram 1100.

[0153] The diagram 1400 illustrates an operation phase 1406 (e.g., that occurs after a model alignment phase). The operation phase 1406 could correspond to any of. for example, the operation phase 652 and / or the operation phase 948 as these are described with respect to varying embodiments disclosed herein.

[0154] During the operation phase 1406, the network 1402 sends 1408 a CSI-RS. The UE 1404 uses the CSI-RS to generate a DL channel estimate (or alternatively a corresponding precoder) that is encoded by an encoder of an AI / ML model for compressed CSI to generate compressed CSI, as has been described herein. Further, based on the DL channel estimate (or alternatively the corresponding precoder), the encoder may determine CQI / RI values corresponding to the compressed CSI. The UE 1404 then sends 1410 a bitstream containing the compressed CSI and the CQI / RI values to the network 1402, thereby informing the network 1402 of the CQI / RI along with the compressed CSI.

[0155] FIG. 15 illustrates a method 1500 of a UE. according to embodiments herein. The method 1500 includes computing 1502 a raw DL channel estimate between a base station and the UE based on a first CSI-RS received from the base station. The method 1500 further includes determining 1504, based on the raw DL channel estimate, precoder information representing a precoder for use by the base station such that the UE experiences a desired effective DL channel corresponding to DL transmissions from the base station. The method1500 further includes encoding 1506, the precoder information into a first bitstream using an encoder of an ML model at the UE. The method 1500 further includes sending 1508 the first bitstream to the base station. The method 1500 further includes computing 1510 a first actual effective DL channel experienced at the UE based on a second CSI-RS received from the base station in response to the first bitstream. The method 1500 further includes calculating 1512 a first error between the desired effective DL channel and the first actual effective DL channel. The method 1500 further includes performing 1514 a first adjustment of the encoder of the ML model at the UE based on the first error.

[0156] In some embodiments of the method 1500, the precoder information comprises a set of eigenvectors corresponding to the precoder.

[0157] In some embodiments, the method 1500 further comprises, after performing the first adjustment of the encoder of the ML model at the UE, encoding the precoder information into a second bitstream using the encoder of the ML model at the UE, and sending the second bitstream to the base station. Some such embodiments further comprise computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station in response to the second bitstream, calculating a second error between the desired effective DL channel and the second actual effective DL channel, and performing a second adjustment of the encoder of the ML model at the UE based on the second error. Some other such embodiments further comprise computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station in response to the second bitstream, determining that a second error between the desired effective DL channel and the second actual effective DL channel is within a threshold, and sending, to the base station, based on the determining that the second error is within the threshold, an indication that the encoder of the ML model at the UE is accurate.

[0158] In some embodiments, the method 1500 further comprises performing an initial training of the ML model at the UE using a third CSI-RS that is received from the base station prior to the first CSI-RS. Some such embodiments further comprise sending a UL SRS to the base station in response to receiving the third CSI-RS.

[0159] In some embodiments, the method 1500 further comprises determining, at the encoder of the ML model at the UE, based on the precoder information, one or more of a CQI and an RI, and including the one or more of the CQI and the RI in the first bitstream prior to sending the first bitstream to the base station.

[0160] In some embodiments, the method 1500 further comprises sending, to the base station, first error information corresponding to the first error.

[0161] FIG. 16 illustrates a method 1600 of a UE. according to embodiments herein. The method 1600 includes computing 1602 a raw DL channel estimate between a base station and the UE based on a first CSI-RS received from the base station. The method 1600 further includes determining 1604, based on the raw DL channel estimate, precoder information representing a precoder for use by the base station, such that the UE experiences a desired effective DL channel corresponding to DL transmissions from the base station. The method 1600 further includes encoding 1606 the precoder information into a first bitstream using an encoder of an ML model at the UE. The method 1600 further includes sending 1608, the first bitstream to the base station. The method 1600 further includes computing 1610 a first actual effective DL channel experienced at the UE based on a second CSI-RS received from the base station in response to the first bitstream. The method 1600 further includes calculating 1612 a first error between the desired effective DL channel and the first actual effective DL channel. The method 1600 further includes sending 1614, to the base station, first error information corresponding to the first error.

[0162] In some embodiments of the method 1600, the precoder information comprises a set of eigenvectors corresponding to the precoder.

[0163] In some embodiments, the method 1600 further includes generating the first error information by encoding the first error using a second ML model at the UE.

[0164] In some embodiments, the method 1600 further includes computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station after sending the first error information, calculating a second error between the desired effective DL channel and the second actual effective DL channel, and sending, to the base station, second error information corresponding to the second error.

[0165] In some embodiments, the method 1600 further includes computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station after sending the first error information, determining that a second error between the desired effective DL channel and the second actual effective DL channel is within a threshold, and sending, to the base station, based on the determining that the second error is within the threshold, an indication that a decoder of a second ML model at the base station is accurate.

[0166] In some embodiments, the method 1600 further includes performing an initial training of the ML model at the UE using a third CSI-RS that is received from the base station prior to the first CSI-RS. Some such embodiments further comprise sending a UL SRS to the base station in response to receiving the third CSI-RS.

[0167] In some embodiments, the method 1600 further includes determining, at the encoder of the ML model at the UE, based on the precoder information, one or more of a CQI and an RI, and including the one or more of the CQI and the RI in the first bitstream prior to sending the first bitstream to the base station.

[0168] FIG. 17 illustrates a method 1700 of a base station, according to embodiments herein. The method 1700 includes sending 1702, to a UE, a first CSLRS. The method 1700 further includes receiving 1704, from the UE, in response to the first CSI-RS, a first bitstream. The method 1700 further includes decoding 1706 first precoder information from the first bitstream using a decoder of a first ML model at the base station. The method 1700 further includes sending 1708, to the UE, a second CSI-RS that is beamformed according to a first precoder represented by the first precoder information.

[0169] In some embodiments of the method 1700, the first precoder information comprises a set of eigenvectors corresponding to the first precoder.

[0170] In some embodiments, the method 1700 further includes receiving, from the UE, in response to the second CSI-RS, a second bitstream, decoding second precoder information from the second bitstream using the decoder of the ML model at the base station, and sending, to the UE, a third CSI-RS that is beamformed according to a second precoder represented by the second precoder information.

[0171] In some embodiments, the method 1700 further includes receiving, from the UE, an indication that an encoder of a second ML model at the UE is accurate.

[0172] In some embodiments, the method 1700 further includes performing an initial training of the ML model at the base station using a UL SRS that is received from the UE prior to the sending of the first CSI-RS.

[0173] In some embodiments of the method 1700, the bitstream further includes one or more of a CQI and an RI.

[0174] In some embodiments, the method 1700 further includes receiving, from the UE, first error information corresponding to a first error between a desired effective DL channel and a first actual effective DL channel experienced by the UE corresponding to the second CSI-RS, and performing an adjustment of the decoder of the first ML model at the base station based on the first error information. In some such embodiments, the first error information comprises an encoding of the first error. Some other such embodiments further comprise decoding, after adjusting the decoder of the first ML model at the base station, the first bitstream into second precoder information using the decoder of the ML model at the base station, and sending, to the UE, a third CSI-RS that is beamformed according to asecond precoder represented by the second precoder information. Certain such embodiments further comprise receiving, from the UE, after sending the third CSI-RS, an indication that the decoder of the ML model at the base station is accurate.

[0175] FIG. 18 illustrates a method 1800 of a UE. according to embodiments herein. The method 1800 includes computing 1802 a raw DL channel estimate between a base station and the UE based on a first CSI-RS received from the base station. The method 1800 further includes determining 1804, based on the raw DL channel estimate, a desired effective DL channel corresponding to DL transmissions from the base station. The method 1800 further includes encoding 1806 the raw DL channel estimate into a first bitstream using an encoder of an ML model at the UE. The method 1800 further includes sending 1808, the first bitstream to the base station. The method 1800 further includes computing 1810 a first actual effective DL channel experienced at the UE based on a second CSI-RS received from the base station in response to the first bitstream. The method 1800 further includes calculating 1812 a first error between the desired effective DL channel and the first actual effective DL channel. The method 1800 further includes performing 1814 a first adjustment of the encoder of the ML model at the UE based on the first error.

[0176] In some embodiments, the method 1800 further includes, after performing the first adjustment of the encoder of the ML model at the UE, encoding the raw DL channel estimate into a second bitstream using the encoder of the ML model at the UE, and sending the second bitstream to the base station. Some such embodiments further comprise computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station in response to the second bitstream, calculating a second error between the desired effective DL channel and the second actual effective DL channel, and performing a second adjustment of the encoder of the ML model at the UE based on the second error. Some other such embodiments further comprise computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station in response to the second bitstream, determining that a second error between the desired effective DL channel and the second actual effective DL channel is within a threshold, and sending, to the base station, based on the determining that the second error is within the threshold, an indication that the encoder of the ML model at the UE is accurate.

[0177] In some embodiments, the method 1800 further includes performing an initial training of the ML model at the UE using a third CSI-RS that is received from the base station prior to the first CSI-RS. Some such embodiments further comprise sending a UL SRS to the base station in response to receiving the third CSI-RS.

[0178] In some embodiments, the method 1800 further includes determining, at the encoder of the ML model at the UE, based on the raw DL channel estimate, one or more of a CQI and an RI, and including the one or more of the CQI and the RI in the first bitstream prior to sending the first bitstream to the base station.

[0179] In some embodiments, the method 1800 further includes sending, to the base station, first error information corresponding to the first error.

[0180] FIG. 19 illustrates a method 1900 of a UE. according to embodiments herein. The method 1900 includes computing 1902 a raw DL channel estimate between a base station and the UE based on a first CSI-RS received from the base station. The method 1900 further includes determining 1904, based on the raw DL channel estimate, a desired effective DL channel corresponding to DL transmissions from the base station. The method 1900 further includes encoding 1906 the raw DL channel estimate into a first bitstream using an encoder of an ML model at the UE. The method 1900 further includes sending 1908 the first bitstream to the base station. The method 1900 further includes computing 1910 a first actual effective DL channel experienced at the UE based on a second CSI-RS received from the base station in response to the first bitstream. The method 1900 further includes calculating 1912 a first error between the desired effective DL channel and the first actual effective DL channel. The method 1900 further includes sending 1914, to the base station, first error information corresponding to the first error.

[0181] In some embodiments, the method 1900 further includes generating the first error information by encoding the first error using a second ML model at the UE.

[0182] In some embodiments, the method 1900 further includes computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station after sending the first error information, calculating a second error between the desired effective DL channel and the second actual effective DL channel, and sending, to the base station, second error information corresponding to the second error.

[0183] In some embodiments, the method 1900 further includes computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station after sending the first error information, determining that a second error between the desired effective DL channel and the second actual effective DL channel is within a threshold, and sending, to the base station, based on the determining that the second error is within the threshold, an indication that a decoder of a second ML model at the base station is accurate.

[0184] In some embodiments, the method 1900 further includes performing an initial training of the ML model at the UE using a third CSI-RS that is received from the base station prior to the first CSI-RS. Some such embodiments further comprise sending a UL SRS to the base station in response to receiving the third CSI-RS.

[0185] In some embodiments, the method 1900 further includes determining, at the encoder of the ML model at the UE, based on the raw DL channel estimate, one or more of a CQI and an RI, and including the one or more of the CQI and the RI in the first bitstream prior to sending the first bitstream to the base station.

[0186] FIG. 20 illustrates a method 2000 of a base station, according to embodiments herein. The method 2000 includes sending 2002, to a UE, a first CSI-RS. The method 2000 further includes receiving 2004, from the UE, in response to the first CSI-RS, a first bitstream. The method 2000 further includes decoding 2006 a first raw DL channel estimate from the first bitstream using a decoder of a first ML model at the base station. The method 2000 further includes determining 2008, based on the first raw DL channel estimate, first precoder information representing a precoder for use. such that the UE experiences a desired effective DL channel corresponding to DL transmissions from the base station under conditions of the first raw DL channel estimate. The method 2000 further includes sending 2010, to the UE, a second CSI-RS that is beamformed according to the first precoder.

[0187] In some embodiments of the method 2000, the precoder information comprises a set of eigenvectors corresponding to the precoder.

[0188] In some embodiments, the method 2000 further includes receiving, from the UE, in response to the second CSI-RS, a second bitstream, decoding a second raw DL channel estimate from the second bitstream using the decoder of the ML model at the base station, determining, based on the second raw DL channel estimate, second precoder information representing a second precoder for use by the base station such that the UE experiences the desired effective DL channel corresponding to DL transmissions from the base station under conditions of the second raw DL channel estimate, and sending, to the UE, a third CSI-RS that is beamformed according to the second precoder.

[0189] In some embodiments, the method 2000 further includes receiving, from the UE, an indication that an encoder of a second ML model at the UE is accurate.

[0190] In some embodiments, the method 2000 further includes performing an initial training of the ML model at the base station using a UL SRS that is received from the UE prior to the sending of the first CSI-RS.

[0191] In some embodiments of the method 2000. the bitstream further includes one or more of a CQI and an RI.

[0192] In some embodiments, the method 2000 further includes receiving, from the UE, first error information corresponding to a first error between the desired effective DL channel and a first actual effective DL channel experienced by the UE corresponding to the second CSI-RS, and performing an adjustment of the decoder of the first ML model at the base station based on the first error information. In some such embodiments, the first error information comprises an encoding of the first error. Some other such embodiments further comprise decoding, after adjusting the decoder of the first ML model at the base station, the first bitstream into a second raw DL channel estimate using the decoder of the ML model at the base station, determining, based on the second raw DL channel estimate, second precoder information representing a second precoder for use by the base station, such that the UE experiences the desired effective DL channel corresponding to DL transmissions from the base station under conditions of the second raw DL channel estimate, and sending, to the UE, a third CSI-RS that is beamformed according to the second precoder. Certain such embodiments further comprise receiving, from the UE, after sending the third CSI-RS, an indication that the decoder of the ML model at the base station is accurate.

[0193] FIG. 21 illustrates an example architecture of a wireless communication system 2100. according to embodiments disclosed herein. The following description is provided for an example wireless communication system 2100 that operates in conjunction with the LTE system standards and / or 5G or NR system standards, as provided by 3GPP technical specifications.

[0194] As shown by FIG. 21, the wireless communication system 2100 includes UE 2102 and UE 2104 (although any number of UEs may be used). In this example, the UE 2102 and the UE 2104 are illustrated as smartphones (e.g., handheld touchscreen mobile computing devices connectable to one or more cellular networks) but may also comprise any mobile or non-mobile computing device configured for wireless communication.

[0195] The UE 2102 and UE 2104 may be configured to communicatively couple with a RAN 2106. In embodiments, the RAN 2106 may be NG-RAN, E-UTRAN, etc. The UE 2102 and UE 2104 utilize connections (or channels) (shown as connection 2108 and connection 2110, respectively) with the RAN 2106, each of which comprises a physical communications interface. The RAN 2106 can include one or more base stations (such as base station 2112 and base station 2114) that enable the connection 2108 and connection 2110.

[0196] In this example, the connection 2108 and connection 2110 are air interfaces to enable such communicative coupling and may be consistent with RAT(s) used by the RAN 2106, such as, for example, an LTE and / or NR.

[0197] In some embodiments, the UE 2102 and UE 2104 may also directly exchange communication data via a sidelink interface 211 . The UE 2104 is shown to be configured to access an access point (shown as AP 2118) via connection 2120. By way of example, the connection 2120 can comprise a local wireless connection, such as a connection consistent with any IEEE 802. 11 protocol, wherein the AP 2118 may comprise a Wi-Fi® router. In this example, the AP 2118 may be connected to another network (for example, the Internet) without going through a CN 2124.

[0198] In embodiments, the UE 2102 and UE 2104 can be configured to communicate using orthogonal frequency division multiplexing (OFDM) communication signals with each other or with the base station 2112 and / or the base station 2114 over a multicarrier communication channel in accordance with various communication techniques, such as, but not limited to, an orthogonal frequency division multiple access (OFDMA) communication technique (e.g., for downlink communications) or a single carrier frequency division multiple access (SC-FDMA) communication technique (e.g., for uplink and ProSe or sidelink communications), although the scope of the embodiments is not limited in this respect. The OFDM signals can comprise a plurality' of orthogonal subcarriers.

[0199] In some embodiments, all or parts of the base station 2112 or base station 2114 may be implemented as one or more softw are entities running on server computers as part of a virtual network. In addition, or in other embodiments, the base station 2112 or base station 2114 may be configured to communicate with one another via interface 2122. In embodiments where the wireless communication system 2100 is an LTE system (e.g., when the CN 2124 is an EPC), the interface 2122 may be an X2 interface. The X2 interface may be defined between two or more base stations (e.g., two or more eNBs and the like) that connect to an EPC, and / or between two eNBs connecting to the EPC. In embodiments where the wireless communication system 2100 is an NR system (e.g., when CN 2124 is a 5GC), the interface 2122 may be an Xn interface. The Xn interface is defined between tw o or more base stations (e.g., two or more gNBs and the like) that connect to 5GC, between a base station 2112 (e.g., a gNB) connecting to 5GC and an eNB, and / or between two eNBs connecting to 5GC (e.g., CN 2124).

[0200] The RAN 2106 is shown to be communicatively coupled to the CN 2124. The CN 2124 may comprise one or more network elements 2126, which are configured to offervarious data and telecommunications services to customers / subscribers (e.g., users of UE 2102 and UE 2104) who are connected to the CN 2124 via the RAN 2106. The components of the CN 2124 may be implemented in one physical device or separate physical devices, including components to read and execute instructions from a machine-readable or computer-readable medium (e.g.. a non-transitory machine-readable storage medium).

[0201] In embodiments, the CN 2124 may be an EPC, and the RAN 2106 may be connected with the CN 2124 via an SI interface 2128. In embodiments, the SI interface 2128 may be split into two parts, an S I user plane (Sl-U) interface, which carries traffic data between the base station 2112 or base station 2114 and a serving gateway (S-GW), and the SI -MME interface, which is a signaling interface between the base station 2112 or base station 2114 and mobility management entities (MMEs).

[0202] In embodiments, the CN 2124 may be a 5GC, and the RAN 2106 may be connected with the CN 2124 via an NG interface 2128. In embodiments, the NG interface 2128 may be split into two parts, an NG user plane (NG-U) interface, which carries traffic data between the base station 2112 or base station 2114 and a user plane function (UPF), and the SI control plane (NG-C) interface, which is a signaling interface between the base station 2112 or base station 2114 and access and mobility management functions (AMFs).

[0203] Generally, an application server 2130 may be an element offering applications that use internet protocol (IP) bearer resources with the CN 2124 (e.g., packet switched data services). The application server 2130 can also be configured to support one or more communication services (e.g., VoIP sessions, group communication sessions, etc.) for the UE 2102 and UE 2104 via the CN 2124. The application server 2130 may communicate with the CN 2124 through an IP communications interface 2132.

[0204] FIG. 22 illustrates a system 2200 for performing signaling 2234 between a wireless device 2202 and a network device 2218, according to embodiments disclosed herein. The system 2200 may be a portion of a wireless communications system, as herein described. The wireless device 2202 may be, for example, a UE of a wireless communication system. The network device 2218 may be, for example, a base station (e.g., an eNB or a gNB) of a wireless communication system.

[0205] The wireless device 2202 may include one or more processor(s) 2204. The processor(s) 2204 may execute instructions such that various operations of the wireless device 2202 are performed, as described herein. The processor(s) 2204 may include one or more baseband processors implemented using, for example, a central processing unit (CPU), a digital signal processor (DSP), an application specific integrated circuit (ASIC), acontroller, a field programmable gate array (FPGA) device, another hardware device, a firmware device, or any combination thereof configured to perform the operations described herein.

[0206] The wireless device 2202 may include a memory 2206. The memory 2206 may be a non-transitory computer-readable storage medium that stores instructions 2208 (which may include, for example, the instructions being executed by the processor(s) 2204). The instructions 2208 may also be referred to as program code or a computer program. The memory 2206 may also store data used by, and results computed by. the processor(s) 2204.

[0207] The wireless device 2202 may include one or more transceiver(s) 2210 that may include radio frequency (RF) transmitter circuitry and / or receiver circuitry that use the antenna(s) 2212 of the wireless device 2202 to facilitate signaling (e.g.. the signaling 2234) to and / or from the wireless device 2202 with other devices (e.g., the network device 2218), according to corresponding RATs.

[0208] The wireless device 2202 may include one or more antenna(s) 2212 (e.g., one, two. four, or more). For embodiments with multiple antenna(s) 2212, the wireless device 2202 may leverage the spatial diversity of such multiple antenna(s) 2212 to send and / or receive multiple different data streams on the same time and frequency resources. This behavior may be referred to as, for example, multiple input multiple output (MIMO) behavior (referring to the multiple antennas used at each of a transmitting device and a receiving device that enable this aspect). MIMO transmissions by the wireless device 2202 may be accomplished according to precoding (or digital beamforming) that is applied at the wireless device 2202 that multiplexes the data streams across the antenna(s) 2212, according to known or assumed channel characteristics such that each data stream is received with an appropriate signal strength relative to other streams and at a desired location in the spatial domain (e.g., the location of a receiver associated with that data stream). Certain embodiments may use single user MIMO (SU-MIMO) methods (where the data streams are all directed to a single receiver) and / or multi user MIMO (MU-MIMO) methods (where individual data streams may be directed to individual (different) receivers in different locations in the spatial domain).

[0209] In certain embodiments having multiple antennas, the wireless device 2202 mayimplement analog beamforming techniques, whereby phases of the signals sent by the antenna(s) 2212 are relatively adjusted such that the (joint) transmission of the antenna(s) 2212 can be directed (this is sometimes referred to as beam steering).

[0210] The wireless device 2202 may include one or more interface(s) 2214. The interface(s) 2214 may be used to provide input to or output from the wireless device 2202. For example, a wireless device 2202 that is a UE may include interface(s) 2214 such as microphones, speakers, a touchscreen, buttons, and the like in order to allow for input and / or output to the UE by a user of the UE. Other interfaces of such a UE may be made up of transmitters, receivers, and other circuitry (e.g., other than the transceiver(s) 2210 / antenna(s) 2212 already described) that allow for communication between the UE and other devices and may operate according to known protocols (e.g., Wi-Fi®, Bluetooth®, and the like).

[0211] The wireless device 2202 may include a CSI compression module 2216. The CSI compression module 2216 may be implemented via hardware, software, or combinations thereof. For example, the CSI compression module 2216 may be implemented as a processor, circuit, and / or instructions 2208 stored in the memory 2206 and executed by the processor(s) 2204. In some examples, the CSI compression module 2216 may be integrated within the processor(s) 2204 and / or the transceiver(s) 2210. For example, the CSI compression module 2216 may be implemented by a combination of software components (e.g.. executed by a DSP or a general processor) and hardware components (e.g., logic gates and circuitry) within the processor(s) 2204 or the transceiver(s) 2210.

[0212] The CSI compression module 2216 may be used for various aspects of the present disclosure, for example, aspects of FIG. 6. FIG. 7, FIG. 8, FIG. 9, FIG. 10, FIG. 11, FIG. 13, FIG. 15, FIG. 16, FIG. 18, and / or FIG. 19. For example, the CSI compression module 2216 may configure the wireless device 2202 to compute a raw DL channel estimate, determine precoder information representing a precoder for a desired effective DL channel, encode the raw DL channel estimate using an encoder of an AI / ML model into a bitstream, encode the precoder using an encoder of an AI / ML model into a bitstream, compute an actual effective DL channel experienced at the UE based on a CSI-RS received in response to sending a bitstream, calculate an error between the desired effective DL channel and the actual effective DL channel, adjust an AI / ML model based on an error, and / or encode an error into an error code and send the error code to a base station, as is described herein.

[0213] The network device 2218 may include one or more processor(s) 2220. The processor(s) 2220 may execute instructions such that various operations of the network device 2218 are performed, as described herein. The processor(s) 2220 may include one or more baseband processors implemented using, for example, a CPU, a DSP, an ASIC, acontroller, an FPGA device, another hardware device, a firmware device, or any combination thereof configured to perform the operations described herein.

[0214] The network device 2218 may include a memory’ 2222. The memory L may be a non-transitory computer-readable storage medium that stores instructions 2224 (which may include, for example, the instructions being executed by the processor(s) 2220). The instructions 2224 may also be referred to as program code or a computer program. The memory' 2222 may also store data used by, and results computed by. the processor(s) 2220.

[0215] The network device 2218 may include one or more transceiver(s) 2226 that may include RF transmitter circuitry' and / or receiver circuitry that use the antenna(s) 2228 of the network device 2218 to facilitate signaling (e.g., the signaling 2234) to and / or from the network device 2218 with other devices (e.g., the wireless device 2202), according to corresponding RATs.

[0216] The network device 2218 may include one or more antenna(s) 2228 (e g., one, two, four, or more). In embodiments having multiple antenna(s) 2228, the network device 2218 may perform MIMO, digital beamforming, analog beamforming, beam steering, etc., as has been described.

[0217] The network device 2218 may include one or more interface(s) 2230. The interface(s) 2230 may be used to provide input to or output from the network device 2218. For example, a network device 2218 that is a base station may include interface(s) 2230 made up of transmitters, receivers, and other circuitry (e.g., other than the transceiver(s) 2226 / antenna(s) 2228 already described) that enables the base station to communicate with other equipment in a core network, and / or that enables the base station to communicate with external networks, computers, databases, and the like for purposes of operations, administration, and maintenance of the base station or other equipment operably connected thereto.

[0218] The network device 2218 may include a CSI compression module 2232. The CSI compression module 2232 may be implemented via hardware, software, or combinations thereof. For example, the CSI compression module 2232 may be implemented as a processor, circuit, and / or instructions 2224 stored in the memory 2222 and executed by the processor(s) 2220. In some examples, the CSI compression module 2232 may be integrated within the processor(s) 2220 and / or the transceiver(s) 2226. For example, the CSI compression module 2232 may be implemented by a combination of software components (e.g., executed by a DSP or a general processor) and hardware components (e.g., logic gates and circuitry) within the processor(s) 2220 or the transceiver(s) 2226.

[0219] The CSI compression module 2232 may be used for various aspects of the present disclosure, for example, aspects of FIG. 6, FIG. 7, FIG. 8, FIG. 9, FIG. 10, FIG. 11, FIG.13, FIG. 17, and / or FIG. 20. For example, the CSI compression module 2232 may configure the network device 2218 to decode a raw DL channel estimate from a bitstream received from a UE using a decoder of an AI / ML model, decode precoder information representing a precoder for a desired effective DL channel from a bitstream received from a UE using a decoder of an AI / ML model, determine precoder information representing a precoder for a desired effective DL channel based on a raw' DL channel estimate, generate and send a C SIRS to a UE based on precoder information, and / or receive an error code corresponding to an error between a desired effective DL channel and an actual effective DL channel from the UE and use it to adjust an decoder of an AI / ML model, as is described herein.

[0220] Embodiments contemplated herein include an apparatus comprising means to perform one or more elements of any of the method 1500, the method 1600, the method 1800, and / or the method 1900. This apparatus may be, for example, an apparatus of a UE (such as a wireless device 2202 that is a UE, as described herein).

[0221] Embodiments contemplated herein include one or more non-transitory computer- readable media comprising instructions to cause an electronic device, upon execution of the instructions by one or more processors of the electronic device, to perform one or more elements of any of the method 1500, the method 1600, the method 1800, and / or the method 1900. This non-transitory computer-readable media may be, for example, a memory of a UE (such as a memory 2206 of a wdreless device 2202 that is a UE, as described herein).

[0222] Embodiments contemplated herein include an apparatus comprising logic, modules, or circuitry to perform one or more elements of any one of the method 1500, the method 1600, the method 1800, and / or the method 1900. This apparatus may be, for example, an apparatus of a UE (such as a w ireless device 2202 that is a UE, as described herein).

[0223] Embodiments contemplated herein include an apparatus comprising: one or more processors and one or more computer-readable media comprising instructions that, when executed by the one or more processors, cause the one or more processors to perform one or more elements of the any one of the method 1500, the method 1600, the method 1800, and / or the method 1900. This apparatus may be, for example, an apparatus of a UE (such as a wireless device 2202 that is a UE, as described herein).

[0224] Embodiments contemplated herein include a signal as described in or related to one or more elements of any of the method 1500, the method 1600, the method 1800. and / or the method 1900.

[0225] Embodiments contemplated herein include a computer program or computer program product comprising instructions, wherein execution of the program by a processor is to cause the processor to carry out one or more elements of any of the method 1500, the method 1600, the method 1800, and / or the method 1900. The processor may be a processor of a UE (such as a processor(s) 2204 of a wireless device 2202 that is a UE, as described herein). These instructions may be, for example, located in the processor and / or on a memory of the UE (such as a memory 2206 of a wireless device 2202 that is a UE, as described herein).

[0226] Embodiments contemplated herein include an apparatus comprising means to perform one or more elements of any of the method 1700 and / or the method 2000. This apparatus may be, for example, an apparatus of a base station (such as a network device 2218 that is a base station, as described herein).

[0227] Embodiments contemplated herein include one or more non-transitory computer- readable media comprising instructions to cause an electronic device, upon execution of the instructions by one or more processors of the electronic device, to perform one or more elements of any of the method 1700 and / or the method 2000. This non-transitory computer- readable media may be, for example, a memory of a base station (such as a memory 2222 of a network device 2218 that is a base station, as described herein).

[0228] Embodiments contemplated herein include an apparatus comprising logic, modules, or circuitry to perform one or more elements of any of the method 1700 and / or the method 2000. This apparatus may be, for example, an apparatus of a base station (such as a network device 2218 that is a base station, as described herein).

[0229] Embodiments contemplated herein include an apparatus comprising: one or more processors and one or more computer-readable media comprising instructions that, when executed by the one or more processors, cause the one or more processors to perform one or more elements of any of the method 1700 and / or the method 2000. This apparatus may be, for example, an apparatus of a base station (such as a network device 2218 that is a base station, as described herein).

[0230] Embodiments contemplated herein include a signal as described in or related to one or more elements of any of the method 1700 and / or the method 2000.

[0231] Embodiments contemplated herein include a computer program or computer program product comprising instructions, wherein execution of the program by a processing element is to cause the processing element to carry out one or more elements of any of the method 1700 and / or the method 2000. The processor may be a processor of a base station(such as a processor(s) 2220 of a network device 2218 that is a base station, as described herein). These instructions may be, for example, located in the processor and / or on a memory of the base station (such as a memory' 2222 of a network device 2218 that is a base station, as described herein).

[0232] For one or more embodiments, at least one of the components set forth in one or more of the preceding figures may be configured to perform one or more operations, techniques, processes, and / or methods as set forth herein. For example, a baseband processor as described herein in connection with one or more of the preceding figures may be configured to operate in accordance with one or more of the examples set forth herein. For another example, circuitry associated with a UE, base station, network element, etc. as described above in connection with one or more of the preceding figures may be configured to operate in accordance with one or more of the examples set forth herein.

[0233] Any of the above described embodiments may be combined with any other embodiment (or combination of embodiments), unless explicitly stated otherwise. The foregoing description of one or more implementations provides illustration and description, but is not intended to be exhaustive or to limit the scope of embodiments to the precise form disclosed. Modifications and variations are possible in light of the above teachings or may be acquired from practice of various embodiments.

[0234] Embodiments and implementations of the systems and methods described herein may include various operations, which may be embodied in machine-executable instructions to be executed by a computer system. A computer system may include one or more general- purpose or special-purpose computers (or other electronic devices). The computer system may include hardware components that include specific logic for performing the operations or may include a combination of hardware, software, and / or firmware.

[0235] It should be recognized that the systems described herein include descriptions of specific embodiments. These embodiments can be combined into single systems, partially combined into other systems, split into multiple systems, or divided or combined in other ways. In addition, it is contemplated that parameters, attributes, aspects, etc. of one embodiment can be used in another embodiment. The parameters, attributes, aspects, etc. are merely described in one or more embodiments for clarity, and it is recognized that the parameters, attributes, aspects, etc. can be combined with or substituted for parameters, attributes, aspects, etc. of another embodiment unless specifically disclaimed herein.

[0236] It is well understood that the use of personally identifiable information should follow privacy policies and practices that are generally recognized as meeting or exceedingindustry or governmental requirements for maintaining the privacy of users. In particular, personally identifiable information data should be managed and handled so as to minimize risks of unintentional or unauthorized access or use, and the nature of authorized use should be clearly indicated to users.

[0237] Although the foregoing has been described in some detail for purposes of clarity, it will be apparent that certain changes and modifications may be made without departing from the principles thereof. It should be noted that there are many alternative ways of implementing both the processes and apparatuses described herein. Accordingly, the present embodiments are to be considered illustrative and not restrictive, and the description is not to be limited to the details given herein, but may be modified within the scope and equivalents of the appended claims.

Claims

CLAIMS1. A method of a user equipment (UE), comprising: computing a raw downlink (DL) channel estimate between a base station and the UE based on a first channel state information reference signal (CSI-RS) received from the base station; determining, based on the raw DL channel estimate, precoder information representing a precoder for use by the base station such that the UE experiences a desired effective DL channel corresponding to DL transmissions from the base station; encoding the precoder information into a first bitstream using an encoder of a machine learning (ML) model at the UE; sending the first bitstream to the base station; computing a first actual effective DL channel experienced at the UE based on a second CSI-RS received from the base station in response to the first bitstream; calculating a first error between the desired effective DL channel and the first actual effective DL channel; and performing a first adjustment of the encoder of the ML model at the UE based on the first error.

2. The method of claim 1, wherein the precoder information comprises a set of eigenvectors corresponding to the precoder.

3. The method of claim 1 , further comprising: after performing the first adjustment of the encoder of the ML model at the UE, encoding the precoder information into a second bitstream using the encoder of the ML model at the UE; and sending the second bitstream to the base station.

4. The method of claim 3, further comprising: computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station in response to the second bitstream; calculating a second error between the desired effective DL channel and the second actual effective DL channel; and performing a second adjustment of the encoder of the ML model at the UE based on the second error.

5. The method of claim 3, further comprising:computing a second actual effective DL channel experienced at the UE using a third CSI-RS received from the base station in response to the second bitstream; determining that a second error between the desired effective DL channel and the second actual effective DL channel is within a threshold; and sending, to the base station, based on the determining that the second error is within the threshold, an indication that the encoder of the ML model at the UE is accurate.

6. The method of claim 1, further comprising performing an initial training of the ML model at the UE using a third CSI-RS that is received from the base station prior to the first CSI- RS.

7. The method of claim 6, further comprising sending an uplink (UL) sounding reference signal (SRS) to the base station in response to receiving the third CSI-RS.

8. The method of claim 1, further comprising: determining, at the encoder of the ML model at the UE, based on the precoder information, one or more of a channel quality index (CQI) and a rank indicator (RI); and including the one or more of the CQI and the RI in the first bitstream prior to sending the first bitstream to the base station.

9. The method of claim 1, further comprising sending, to the base station, first error information corresponding to the first error.

10. A method of a user equipment (UE), comprising: computing a raw downlink (DL) channel estimate between a base station and the UE based on a first channel state information reference signal (CSI-RS) received from the base station; determining, based on the raw DL channel estimate, precoder information representing a precoder for use by the base station such that the UE experiences a desired effective DL channel corresponding to DL transmissions from the base station; encoding the precoder information into a first bitstream using an encoder of a machine learning (ML) model at the UE; sending the first bitstream to the base station; computing a first actual effective DL channel experienced at the UE based on a second CSI-RS received from the base station in response to the first bitstream; calculating a first error between the desired effective DL channel and the first actual effective DL channel; andsending, to the base station, first error information corresponding to the first error.

11. The method of claim 10, wherein the precoder information comprises a set of eigenvectors corresponding to the precoder.

12. The method of claim 10, further comprising generating the first error information by encoding the first error using a second ML model at the UE.

13. The method of claim 10, further comprising: computing a second actual effective DL channel experienced at the UE using a third CSLRS received from the base station after sending the first error information; calculating a second error between the desired effective DL channel and the second actual effective DL channel; and sending, to the base station, second error information corresponding to the second error.

14. The method of claim 10, further comprising: computing a second actual effective DL channel experienced at the UE using a third CSLRS received from the base station after sending the first error information; determining that a second error between the desired effective DL channel and the second actual effective DL channel is within a threshold; and sending, to the base station, based on the determining that the second error is within the threshold, an indication that a decoder of a second ML model at the base station is accurate.

15. The method of claim 10, further comprising performing an initial training of the ML model at the UE using a third CSI-RS that is received from the base station prior to the first CSLRS.

16. The method of claim 15, further comprising sending an uplink (UL) sounding reference signal (SRS) to the base station in response to receiving the third CSLRS.

17. The method of claim 10, further comprising: determining, at the encoder of the ML model at the UE, based on the precoder information, one or more of a channel quality index (CQI) and a rank indicator (RI); and including the one or more of the CQI and the RI in the first bitstream prior to sending the first bitstream to the base station.

18. A method of a base station, comprising: sending, to a user equipment (UE), a first channel state information reference signal (CSI-RS); receiving, from the UE, in response to the first CSI-RS, a first bitstream; decoding first precoder information from the first bitstream using a decoder of a first machine learning (ML) model at the base station; and sending, to the UE, a second CSI-RS that is beamformed according to a first precoder represented by the first precoder information.

19. The method of claim 18, wherein the first precoder information comprises a set of eigenvectors corresponding to the first precoder.

20. The method of claim 18, further comprising: receiving, from the UE, in response to the second CSI-RS, a second bitstream; decoding second precoder information from the second bitstream using the decoder of the ML model at the base station; and sending, to the UE, a third CSI-RS that is beamformed according to a second precoder represented by the second precoder information.

21. The method of claim 18, further comprising receiving, from the UE, an indication that an encoder of a second ML model at the UE is accurate.

22. The method of claim 18, further comprising performing an initial training of the ML model at the base station using an uplink (UL) sounding reference signal (SRS) that is received from the UE prior to the sending of the first CSI-RS.

23. The method of claim 18, wherein the bitstream further includes one or more of a channel quality index (CQI) and a rank indicator (RI).

24. The method of claim 18, further comprising: receiving, from the UE, first error information corresponding to a first error between a desired effective DL channel and a first actual effective DL channel experienced by the UE corresponding to the second CSI-RS; and performing an adjustment of the decoder of the first ML model at the base station based on the first error information.

25. The method of claim 24, wherein the first error information comprises an encoding of the first error.

26. The method of claim 24, further comprising: decoding, after adjusting the decoder of the first ML model at the base station, the first bitstream into second precoder information using the decoder of the ML model at the base station; and sending, to the UE, a third CSI-RS that is beamformed according to a second precoder represented by the second precoder information.

27. The method of claim 26, further comprising receiving, from the UE. after sending the third CSI-RS, an indication that the decoder of the ML model at the base station is accurate.

28. An apparatus comprising means to perform the method of any of claim 1 to claim 27.

29. A computer-readable media comprising instructions to cause an electronic device, upon execution of the instructions by one or more processors of the electronic device, to perform the method of any of claim 1 to claim 27.

30. An apparatus comprising logic, modules, or circuitry to perform the method of any of claim 1 to claim 27.

Images