WO2023201695A1 - Automatic model selection - Google Patents

Automatic model selection Download PDF

Info

Publication number
WO2023201695A1
WO2023201695A1 PCT/CN2022/088430 CN2022088430W WO2023201695A1 WO 2023201695 A1 WO2023201695 A1 WO 2023201695A1 CN 2022088430 W CN2022088430 W CN 2022088430W WO 2023201695 A1 WO2023201695 A1 WO 2023201695A1
Authority
WO
WIPO (PCT)
Prior art keywords
model
determining
determined
feature
candidate
Prior art date
Application number
PCT/CN2022/088430
Other languages
French (fr)
Inventor
Huaisong Zhu
Pingping WU
Original Assignee
Telefonaktiebolaget Lm Ericsson (Publ)
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Telefonaktiebolaget Lm Ericsson (Publ) filed Critical Telefonaktiebolaget Lm Ericsson (Publ)
Priority to PCT/CN2022/088430 priority Critical patent/WO2023201695A1/en
Publication of WO2023201695A1 publication Critical patent/WO2023201695A1/en

Links

Images

Classifications

    • HELECTRICITY
    • H04ELECTRIC COMMUNICATION TECHNIQUE
    • H04BTRANSMISSION
    • H04B1/00Details of transmission systems, not covered by a single one of groups H04B3/00 - H04B13/00; Details of transmission systems not characterised by the medium used for transmission
    • H04B1/38Transceivers, i.e. devices in which transmitter and receiver form a structural unit and in which at least one part is used for functions of transmitting and receiving
    • H04B1/40Circuits
    • H04B1/401Circuits for selecting or indicating operating mode
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/08Learning methods
    • G06N3/0985Hyperparameter optimisation; Meta-learning; Learning-to-learn
    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06NCOMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00Computing arrangements based on biological models
    • G06N3/02Neural networks
    • G06N3/04Architecture, e.g. interconnection topology
    • G06N3/0464Convolutional networks [CNN, ConvNet]

Definitions

  • the present disclosure is related to the field of telecommunication, and in particular, to a communication device and a method for automatic model selection.
  • RAN Radio Access Network
  • 5G fifth generation
  • NR New Radio
  • Modulation mappers are one of the most important functional blocks that can significantly impact the performance of a 5G RAN.
  • a variety of different modulation mappers are defined in 3rd Generation Partnership Project (3GPP) technical specifications (TS) for 5G RAN technology, such as, ⁇ /2 -Binary Phase Shift Keying ( ⁇ /2-BPSK) , BPSK, Quadrature Phase Shift Keying (QPSK) , 16 Quadrature Amplitude Modulation (16QAM) , 64QAM, 256 QAM, 1024QAM, etc.
  • 3GPP 3rd Generation Partnership Project
  • TS technical specifications
  • ⁇ /2-BPSK BPSK
  • QPSK Quadrature Phase Shift Keying
  • 16QAM 16 Quadrature Amplitude Modulation
  • 64QAM 64QAM
  • 256 QAM 256 QAM
  • 1024QAM 1024QAM
  • PSK Phase Shift Keying
  • 5G technology implements BPSK and QPSK as the lower order modulation formats. Although they will provide the slower data throughput, they will also provide the most robust link and as such it can be used when signal levels are low or when interference is high.
  • PSK Phase Shift Keying
  • Another form of PSK called ⁇ /2-BPSK is used in conjunction with Discrete Fourier Transform spread Orthogonal Frequency Division Multiplexing (DFT-s-OFDM) on the uplink.
  • DFT-s-OFDM Discrete Fourier Transform spread Orthogonal Frequency Division Multiplexing
  • Quadrature Amplitude Modulation QAM enables the data throughput to be increased. Formats used within 5G mobile communications system include 16QAM, 64QAM and 256QAM. Further, with the evolution of 5G technology, other QAM formats may also be used, for example 1024QAM. The higher the order of modulation, the greater the throughput, although the penalty is the noise resilience. Therefore, 256QAM or higher order formats are only used when link quality is good, and it reduces to 64QAM and then 16QAM etc., as the link deteriorates. It is a balance between data throughput and resilience.
  • QAM is widely used for data transmission as it enables better levels of spectral efficiency than other formats of modulation.
  • QAM uses two carriers on the same frequency shifted by 90° which are modulated by two data streams: an I or In-phase component and a Q or Quadrature component.
  • 3GPP defines some modulation mappers in Subclause 5.1 of 3GPP TS 38.211, definitions of corresponding demodulation de-mappers are not provided by 3GPP and therefore they are vendor specific.
  • a method at a communication device in a telecommunications network for model selection comprises: determining one of multiple modes to be used for model selection; and selecting, in the determined mode, a model from a set of candidate models to be used by a feature.
  • the multiple modes comprise at least one of: a random mode in which a model is to be randomly selected; and a greedy mode in which a model with the highest performance value is to be selected.
  • the feature comprises at least one of: soft bit computation.
  • the method further comprises: performing the feature by at least using the selected model.
  • the set of candidate models comprise one or more trained Machine Learning (ML) models.
  • at least one of the step of determining the mode and the step of selecting the model is performed in response to each invocation of the feature.
  • the method before the step of selecting the model, further comprises: determining the set of candidate models at least based on one or more measurements related to the feature.
  • the one or more measurements comprise, for a UE associated with the soft bit computation, at least one of: a pathloss for the UE; an Inter-Cell Interference (ICI) or a Signal to Interference plus Noise Ratio (SINR) for the UE; a speed of the UE; and a channel delay profile for the UE.
  • the set of candidate models is determined periodically, and once it is determined, the determined set of candidate models is used during a corresponding period of time.
  • the set of candidate models is determined in response to a trigger event, and once it is determined, the determined set of candidate models is used until another trigger event is detected.
  • the step of determining the mode comprises at least one of: determining, with a first probability, the random mode to be used for model selection; and determining, with a second probability, the greedy mode to be used for model selection.
  • a sum of the first probability and the second probability is 1.
  • the first probability has a value that is continuously decreased in response to an increasing number of invocations of the feature since the last determination of the set of candidate models.
  • the step of determining the mode comprises at least one of: generating a random number between 0 and 1; determining whether the random number is less than a probability threshold or not; and determining the random mode to be used for model selection in response to determining that the random number is less than the probability threshold and/or determining the greedy mode to be used for model selection in response to determining that the random number is greater than or equal to the probability threshold.
  • an initial performance value is assigned to each of candidate models in the set once the set is determined.
  • the method further comprises at least one of: increasing the performance value of the determined model by a first value in response to a first outcome resulted from performing the feature by using the determined model; and decreasing the performance value of the determined model by a second value in response to a second outcome resulted from performing the feature by using the determined model.
  • the first value is different from the second value.
  • the method further comprises: removing a candidate model from the set of candidate models in response to determining that the candidate model has a lower performance value than a threshold and/or than all other candidate models for a period of time and/or for specific times.
  • a communication device comprises: a processor; a memory storing instructions which, when executed by the processor, cause the processor to perform the method of any of the first aspect.
  • a communication device comprising: a determining module configured to determine one of multiple modes to be used for model selection; and a selecting module configured to select, in the determined mode, a model from a set of candidate models to be used by a feature.
  • the communication device may comprise one or more further modules configured to perform any of the method of the first aspect.
  • a computer program comprising instructions.
  • the instructions when executed by at least one processor, cause the at least one processor to carry out the method of any of the first aspect.
  • a carrier containing the computer program of the fourth aspect is provided.
  • the carrier is one of an electronic signal, optical signal, radio signal, or computer readable storage medium.
  • Fig. 1 is a diagram illustrating an exemplary telecommunications network in which UEs and gNB may be operated according to an embodiment of the present disclosure.
  • Fig. 2A and Fig. 2B are diagrams illustrating exemplary constellations for de-mapping from symbols to bits according to an embodiment of the present disclosure.
  • Fig. 3 is a diagram illustrating an exemplary convolutional neural network (CNN) based on which an ML model for de-mapping may be trained and used according to an embodiment of the present disclosure.
  • CNN convolutional neural network
  • Fig. 4 is a diagram illustrating comparison between performance of a legacy de-mapper and an ML assisted de-mapper according to an embodiment of the present disclosure.
  • Fig. 5 is a flow chart illustrating an exemplary method for creating a set of candidate models according to an embodiment of the present disclosure.
  • Fig. 6 is a flow chart illustrating an exemplary method for selecting a model from a set of candidate models for de-mapping according to an embodiment of the present disclosure.
  • Fig. 7 is a flow chart illustrating an exemplary method at a communication device for model selection according to an embodiment of the present disclosure.
  • Fig. 8 schematically shows an embodiment of an arrangement which may be used in a UE, a network node, or a communication device according to an embodiment of the present disclosure.
  • Fig. 9 is a block diagram of an exemplary communication device according to an embodiment of the present disclosure.
  • Fig. 10 schematically illustrates a telecommunication network connected via an intermediate network to a host computer according to an embodiment of the present disclosure.
  • Fig. 11 is a generalized block diagram of a host computer communicating via a base station with a user equipment over a partially wireless connection according to an embodiment of the present disclosure.
  • Fig. 12 to Fig. 15 are flowcharts illustrating methods implemented in a communication system including a host computer, a base station, and a user equipment according to an embodiment of the present disclosure.
  • ′′exemplary′′ is used herein to mean ′′illustrative, ′′ or ′′serving as an example, ′′ and is not intended to imply that a particular embodiment is preferred over another or that a particular feature is essential.
  • the terms ′′first′′ , ′′second′′ , ′′third′′ , ′′fourth, ′′ and similar terms are used simply to distinguish one particular instance of an item or feature from another, and do not indicate a particular order or arrangement, unless the context clearly indicates otherwise.
  • the term ′′step, ′′ as used herein is meant to be synonymous with ′′operation′′ or ′′action. ′′ Any description herein of a sequence of steps does not imply that these operations must be carried out in a particular order, or even that these operations are carried out in any order at all, unless the context or the details of the described operation clearly indicates otherwise.
  • the term ′′or′′ is used in its inclusive sense (and not in its exclusive sense) so that when used, for example, to connect a list of elements, the term ′′or′′ means one, some, or all of the elements in the list.
  • the term ′′each, ′′ as used herein, in addition to having its ordinary meaning, can mean any subset of a set of elements to which the term ′′each′′ is applied.
  • processing circuits may in some embodiments be embodied in one or more application-specific integrated circuits (ASICs) .
  • these processing circuits may comprise one or more microprocessors, microcontrollers, and/or digital signal processors programmed with appropriate software and/or firmware to carry out one or more of the operations described above, or variants thereof.
  • these processing circuits may comprise customized hardware to carry out one or more of the functions described above. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive.
  • the inventive concept of the present disclosure may be applicable to any appropriate communication architecture, for example, to Global System for Mobile Communications (GSM) /General Packet Radio Service (GPRS) , Enhanced Data Rates for GSM Evolution (EDGE) , Code Division Multiple Access (CDMA) , Wideband CDMA (WCDMA) , Time Division -Synchronous CDMA (TD-SCDMA) , CDMA2000, Worldwide Interoperability for Microwave Access (WiMAX) , Wireless Fidelity (Wi-Fi) , 4th Generation Long Term Evolution (LTE) , LTE-Advance (LTE-A) , or 5G NR, etc.
  • GSM Global System for Mobile Communications
  • GPRS General Packet Radio Service
  • EDGE Enhanced Data Rates for GSM Evolution
  • CDMA Code Division Multiple Access
  • WCDMA Wideband CDMA
  • TD-SCDMA Time Division -Synchronous CDMA
  • CDMA2000 Code Division -Synchronous CDMA
  • the terms used herein may also refer to their equivalents in any other infrastructure.
  • the term ′′User Equipment′′ or ′′UE′′ used herein may refer to a terminal device, a mobile device, a mobile terminal, a mobile station, a user device, a user terminal, a wireless device, a wireless terminal, or any other equivalents.
  • the term ′′gNB′′ used herein may refer to a network node, a base station, a base transceiver station, an access point, a hot spot, a NodeB, an Evolved NodeB, a network element, or any other equivalents.
  • the term ′′communication device′′ may refer to any of those listed above or any device that can communicate with another device.
  • 3GPP TS 38.211 V17.0.0 (2021-12) Technical Specification, 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; NR; Physical channels and modulation (Release 17) .
  • Fig. 1 is a diagram illustrating an exemplary telecommunications network 10 in which UE #1 100-1, UE #2 100-2, and gNB 105 may be operated according to an embodiment of the present disclosure.
  • the telecommunications network 10 is a network defined in the context of 5G NR, the present disclosure is not limited thereto.
  • the network 10 may comprise one or more UEs 100-1 and 100-2 (collectively, UE (s) 100) and a RAN node 105, which could be a base station, a Node B, an evolved NodeB (eNB) , a gNB, or an AN node which provides the UEs 100 with access to the network. Further, the network 10 may comprise its core network portion that is not shown in Fig. 1.
  • the network 10 may comprise additional nodes, less nodes, or some variants of the existing nodes shown in Fig. 1.
  • the entities e.g., an eNB
  • the gNB 105 may be different from those shown in Fig. 1.
  • some of the entities may be same as those shown in Fig. 1, and others may be different.
  • two UEs 100 and one gNB 105 are shown in Fig. 1, the present disclosure is not limited thereto. In some other embodiments, any number of UEs and/or any number of gNBs may be comprised in the network 10.
  • the UEs 100 may be communicatively connected to the gNB 105 which in turn may be communicatively connected to a corresponding Core Network (CN) and then the Internet, such that the UEs 100 may finally communicate its user plane data with other devices outside the network 10, for example, via the gNB 105.
  • CN Core Network
  • the 5G Evolution/6G is envisioned to enable transmission at higher carrier frequencies and to provide opportunities with a larger spectrum to be allocated for transmissions.
  • transmission at these frequencies compared to the already existing frequency bands, is even more challenging due to the signal distortions induced by hardware impairments e.g., oscillator′s phase noise, power amplifier′s nonlinearities, and digital-to-analog converter′s quantization noise.
  • phase noise compensation approaches in 5G systems try to estimate phase noise from transmitted phase tracking reference signals (PTRS) and more demodulation reference signals (DMRS) , and compensate for the common phase error which is the average phase noise across the symbol.
  • PTRS transmitted phase tracking reference signals
  • DMRS demodulation reference signals
  • a constellation diagram is useful for QAM (e.g., the constellation for 16QAM shown in Fig. 2A and Fig. 2B) .
  • Fig. 2A and Fig. 2B are diagrams illustrating exemplary constellations for de-mapping from symbols to bits according to an embodiment of the present disclosure.
  • Fig. 2A shows an ideal constellation and each point thereof could be an ideal symbol detected by a receiver
  • Fig. 2B shows an actual symbol 200 detected by a receiver when compared and overlapped with the ideal constellation shown in Fig. 2A.
  • ideal constellation points are usually arranged in a square grid with equal vertical and horizontal spacing, although other configurations are possible (e.g., a hexagonal or triangular grid) .
  • the data is usually binary, so the number of points in the grid is typically a power of 2 (2, 4, 8, ... ) , corresponding to the number of bits per symbol.
  • the simplest and most commonly used QAM constellations consist of points arranged in a square, i.e. 16-QAM, 64-QAM and 256-QAM (even powers of two) .
  • Non-square constellations, such as Cross-QAM can offer greater efficiency but are rarely used because of the cost of increased modem complexity.
  • the receiver may determine that the symbol shall be de-mapped to ′′1100′′ since the symbol has I/Q components that are perfectly matched with the ideal constellation point corresponding to ′′1100′′ .
  • a symbol that was mapped to ′′1100′′ by the transmitter may be detected by the receiver as a symbol indicated by the indicator 200, which is located among and close to one or more ideal constellation points depending on its detected I/Q components.
  • the symbol 200 may be de-mapped to soft bits (instead of hard bits) , and finally decoded to ′′1100′′ , ′′1110′′ , ′′1101′′ or the like.
  • soft bits can be computed in a simplified way as follows. Assuming a received coded sequence is:
  • y is the received coded sequence
  • c is the modulated coded sequence taking values in the alphabet:
  • ⁇ 16QAM ⁇ 1 ⁇ 1j, ⁇ 1 ⁇ 3j, ⁇ 3 ⁇ 1j, ⁇ 3 ⁇ 3j ⁇ as also shown in Fig. 2A and Fig. 2B
  • n is the Additive White Gaussian Noise (AWGN) following the probability distribution function:
  • MAP Maximum a posteriori probability
  • bit mapping for the bit b 0 with 16QAM coded mapping is shown in Fig. 2A and Fig. 2B.
  • b 0 toggles from 0 to 1, only the real part of the constellation is affected.
  • a likelihood ratio can be defined if:
  • the soft bit for bit b 0 is:
  • bit mapping for the bit b 2 with 16QAM coded mapping is also shown in Fig. 2A and Fig. 2B. Similar to b 0 , when b 2 toggles from 0 to 1, only the real part of the constellation is affected.
  • a likelihood ratio can be defined if:
  • the soft bit for bit b 2 is:
  • the soft bit for bit b 2 can be further simplified to:
  • bits b 1 and b 3 are identical to softbits for b 0 and b 2 , respectively, except that the decisions are based on the imaginary component y im of the received symbol.
  • the soft bit for bit b 1 is:
  • the soft bit for bit b 3 is:
  • the symbol 200 may have a real part of -1.5 and an imaginary part of -1, then the corresponding soft bits are:
  • the symbol 200 may be de-mapped as soft bits (1.5, 1, -0.5, -1) , which may then be converted to hard bits.
  • a simple method to determine hard bits from the soft bits is to find the minimum Euclidean distance between the soft bits and the ideal constellation points (e.g., (1, 1, 0, 0) , (1, 1, 1, 0) ) in the vector space.
  • the ideal constellation point that is closest to the soft bits of the symbol 200 is the constellation point corresponding to the hard bits (1, 1, 0, 0) .
  • the present disclosure is not limited thereto. In some other embodiments, different methods for soft bit calculation and/or soft-bit to hard-bit conversion may be applicable.
  • this method heavily relies on an assumption that the noise introduced into the received symbol is AWGN.
  • some residual ICI due to the phase noise fluctuations between modulated symbols also remains, which influences the performance of the receiver.
  • full compensation of the phase noise is not typically feasible or requires a very high density of the reference signals, it is likely that even after phase noise compensation, some residual ICI remains.
  • the residual ICI in the DFT-s-OFDM or CP-OFDM reception leads to the deviation of the distribution of the noise in the equalized signal at the receiver side from uncorrelated, which is typically assumed in conventional receivers as mentioned above. This mismatch causes degraded soft bit computation and leads to sub-optimum performance of the soft decoding algorithms for Low Density Parity Check (LDPC) or Turbo codes.
  • LDPC Low Density Parity Check
  • machine learning techniques may also be applicable to UEs or other nodes than the base stations.
  • Some Machine learning based de-mapper technologies has been studied and implemented in PoC (Prototype of Concept) in Q4 2021.
  • Fig. 3 is a diagram illustrating an exemplary CNN 300 based on which an ML model for de-mapping may be trained and used according to an embodiment of the present disclosure.
  • CNN convolutional neural network
  • ConvNet convolutional neural network
  • ANN artificial neural network
  • the CNN 300 may comprise one or more layers, such as one or more convolutional layers (e.g., convolutional layers 315 and 325) , one or more pooling layers (e.g., pooling layers 320 and 330) , a flatten layer (e.g., a flatten layer 335) , and/or a fully connected layer (e.g., a full connected layer 340) .
  • the CNN 300 may comprise one or more other types of layers, such as batch normalization layers, dilated layers, loss layers, etc.
  • the present disclosure is not limited thereto.
  • other ML-based networks/models/algorithms or even other none-ML based technologies may also be applicable for model selection.
  • Convolutional networks were inspired by biological processes in that the connectivity pattern between neurons resembles the organization of the animal visual cortex. Individual cortical neurons respond to stimuli only in a restricted region of the visual field known as the receptive field (e.g., the receptive field 360 shown in Fig. 3) . The receptive fields of different neurons partially overlap such that they cover the entire visual field.
  • Convolutional neural network indicates that the network employs a mathematical operation called convolution.
  • Convolutional networks are a specialized type of neural networks that use convolution in place of general matrix multiplication in at least one of their layers.
  • the input is a tensor with a shape: (number of inputs) x (input height) x (input width) x (input channels) .
  • the input e.g., an image, or in the embodiment shown in Fig. 3, received symbols
  • a feature map also called an activation map
  • shape (number of inputs) x (feature map height) x (feature map width) x (feature map channels)
  • the output of the convolutional layer #1 315 may be a feature map 365 with shape: (1, 24, 28, 28) .
  • the output of the convolutional layer #2 325 may be a feature map 365 with shape: (1, 24, 10, 10) .
  • shape (1, 24, 10, 10) .
  • Fig. 3 is only provided for illustration purpose, and therefore the shapes of the feature maps may not be scaled properly.
  • Convolutional layers convolve the input and pass its result to the next layer. This is similar to the response of a neuron in the visual cortex to a specific stimulus. Each convolutional neuron processes data only for its receptive field.
  • feedforward neural networks can be used to learn features and classify data, this architecture is generally impractical for larger inputs. It would require a very high number of neurons, even in a shallow architecture. Instead, convolution reduces the number of free parameters, allowing the network to be deeper. For example, regardless of input size, using 5 x 5 kernels, each with the same shared weights, requires only 25 learnable parameters. Using regularized weights over fewer parameters avoids the vanishing gradients and exploding gradients problems seen during backpropagation in traditional neural networks. Furthermore, convolutional neural networks are ideal for data with a grid-like topology as spatial relations between separate features are taken into account during convolution and/or pooling.
  • Convolutional networks may also include local and/or global pooling layers (e.g., the pooling layers 320 and 330 shown in Fig. 3) along with traditional convolutional layers.
  • Pooling layers reduce the dimensions of data by combining the outputs of neuron clusters at one layer into a single neuron in the next layer.
  • Local pooling combines small clusters, tiling sizes such as 2 x 2 are commonly used.
  • Global pooling acts on all the neurons of the feature map. There are two common types of pooling in popular use: max and average. Max pooling uses the maximum value of each local cluster of neurons in the feature map, while average pooling takes the average value.
  • the pooling layers #1 320 and the pooling layer #2 330 may be max pooling layers.
  • a flatten layer may flatten a multi-dimension feature map into a one-dimension feature vector, for example, by concatenating vectors in the feature map one to another. As shown in Fig. 3, a flatten layer 335 may convert the output of the pooling layer #2 330 into a one-dimension vector that is input to the fully connected layer 340.
  • Fully connected layers or dense layers (e.g., the fully connected layers 340 shown in Fig. 3) connect every neuron in one layer to every neuron in another layer. It is the same as a traditional multilayer perceptron neural network (MLP) .
  • MLP multilayer perceptron neural network
  • the flattened vector goes through a fully connected layer to determine the output, e.g., soft bits or hard bits in the embodiment shown in Fig. 3.
  • each neuron receives input from some number of locations in the previous layer.
  • each neuron receives input from only a restricted area of the previous layer called the neuron′s receptive field (e.g., the receptive field 360) .
  • the area is a square (e.g., a 5 by 5 neuron) .
  • the receptive field is the entire previous layer (e.g., as shown by the fully connected layer 340) .
  • each neuron takes input from a larger area in the input than previous layers. This is due to applying the convolution over and over, which takes into account a value, as well as its surrounding values.
  • the number of pixels in the receptive field remains constant, but the field is more sparsely populated as its dimensions grow when combining the effect of several layers.
  • Each neuron in a neural network computes an output value by applying a specific function to the input values received from the receptive field in the previous layer.
  • the function that is applied to the input values is determined by a vector of weights and a bias. Learning consists of iteratively adjusting these biases and weights.
  • the vector of weights and the bias are called filters and represent particular features of the input (e.g., a particular pattern in the input) .
  • a distinguishing feature of CNNs is that many neurons can share the same filter. This reduces the memory footprint because a single bias and a single vector of weights may be used across all receptive fields that share that filter, as opposed to each receptive field having its own bias and vector weighting. However, in some embodiments, all or some of the receptive fields may have their own bias and vector weighting, respectively.
  • an input 310 (e.g., I/Q components of symbols detected at the receiver) may be processed by the CNN 300, which are trained based on training data sets for de-mapping (e.g., a set of collected data for gNB-UE communications in a practical telecommunications network) , and finally the demodulated soft/hard bits may be derived as the output 345.
  • training data sets for de-mapping e.g., a set of collected data for gNB-UE communications in a practical telecommunications network
  • Fig. 4 is a diagram illustrating comparison between performance of a legacy de-mapper and an ML assisted de-mapper according to an embodiment of the present disclosure.
  • a Prototype of Concept (PoC) implementation of ML assisted de-mapper has proven that:
  • - ML assisted de-mapper (410) outperforms the legacy commercial De-mapper (420) at high SINR area (cell center) due to less DMRS (i.e., less overhead) , and it has observed 8%peak throughput increase in lab test;
  • - ML assisted de-mapper (410) outperforms the legacy commercial De-mapper (420) at middle or low SINR area (cell middle or cell edge) due to better control of phase noise, and it has observed around 5dB gain in lab test.
  • ML assisted de-mapper Since ML assisted de-mapper has proven its high capability to suppress phase noise, it has attracted a lot of interest to study how to deploy it in commercial product. Different with lab tests or simulations which normally target for one specific scenario, a real product needs to consider various kinds of channel environments, different UE speed and SINR. Hence ML based de-mappers should be well-generalized to handle all these scenarios.
  • An ML based de-mapper whose neural network is specifically optimized (trained) for Clustered Delay Line -E (CDL-E) channel (LOS, Line-of-Sight channel) and 120 km/h speed, may have a total number of parameters: 21, 801, a number of trainable parameters: 21, 725, and a number of non-trainable parameters: 76.
  • CDL-E Clustered Delay Line -E
  • An ML based de-mapper whose neural network is specifically optimized (trained) for CDL-B channel (typical NLOS, non-line-of-sight channel) and pedestrian UE speed at 3 km/h, is a neural network with good performance. It could be trained with similar complexity with Case 1 (but different neural network weights) .
  • Case-2-optimized neural network performs about 2 dB worse than Case-1-optimized neural network.
  • This 2 dB gap means Case-2-optimized neural network is not well-generalized for all scenarios.
  • An ML based de-mapper whose neural network is one neural network to be optimized for both Case 1 and Case 2, tries to get similar performance with Case-1-optimized neural network in Case 1, and Case-2-optimized neural network in Case 2 separately.
  • This neural network is designed with higher complexity, i.e., more channels, more hidden layer and neurons, and its computation complexity is 27.6M Flops (compared with Case-1-optimized neural network 5.06M Flops)
  • Some embodiments of the present disclosure propose a method for automatically selecting a proper ML model from a candidate list, if a group of ′tiny′ ML models is designed.
  • some embodiments of the present disclosure propose a generalized ML model selection, i.e., ⁇ -greedy solution to automatically select the proper ML model.
  • a general solution according to some embodiments of the present disclosure may comprise steps as follows.
  • Step 1 With a certain period, based on internal measurement, one optimal ML model candidate sub-set ⁇ S ⁇ may be selected.
  • the measurement (s) could for example be at least one of:
  • SINR - Inter-cell interference
  • the term ′period′ means that a receiver thinks, in this period, scenario is coherent, e.g., for 3 second.
  • Step 2 For each model selected into the candidate sub-set, it will be assigned an initial (performance) value V, indicating how good performance this ML model is expected. (higher V, better performance) .
  • Step 3 For each receiving occasion, the receiver will select a model:
  • Step 4 If a model is selected, the receiver will use this model to process the data. Its final performance, for example, de-mapper result, could be correctly decoded or not will impact model′s (performance) value V. For example, if a model′s output could be successfully decoded, its value V will increase ⁇ , or if failed, its value V will decrease ⁇ ′. In some embodiments, ⁇ may be different from or same as ⁇ ′.
  • Step 5 if a model′s value V is too low, i.e., always perform worse than others, it could be removed from candidate sub-set ⁇ S ⁇ .
  • Step 6 ⁇ determines exploration possibility and it normally has higher value for a new candidate sub-set ⁇ S ⁇ , while ⁇ continuously decrease when models in candidate sub-set ⁇ S ⁇ has been tried for several times.
  • the advantage of the ⁇ -greedy is that it will spend more and more time exploring the interesting parts of the environment, as the performance value (V) estimates get better and better, while still spending some time visiting unknown performance of other models. It is quite common to set an initial high value for ⁇ (e.g., 50 %) while (e.g., 5%) for models in candidate sub-set ⁇ S ⁇ has been tried for several times.
  • some embodiments of the present disclosure allow a candidate sub-set ⁇ S ⁇ , in which each model is simple but only work for limited scenarios.
  • Advantages of some embodiments of the present disclosure may comprise (but not limited to) avoiding any super complex ML model and auto-selecting proper ML model from candidate list, which make receiver more robust.
  • Fig. 5 is a flow chart illustrating an exemplary method 500 for creating a set of candidate models according to an embodiment of the present disclosure.
  • the system needs a timer triggered function, to periodically create optimal ML candidate sub-set ⁇ S ⁇ .
  • This method 500 may be designed to reduce the number of candidate ML models to explore in the following ⁇ -greedy selection. In some embodiments, it is optional but recommended to keep it especially when there is a lot of candidate ML models for selection. However, the present disclosure is not limited thereto. In some other embodiments, this method 500 may be omitted. In other words, all the candidate models are comprised in the candidate set. Further, although the method 500 is shown as being triggered by a timer, the present disclosure is not limited thereto. In other embodiments, the method 500 may be triggered by an event, for example, upon detection of the availability of a peer for communication.
  • the method 500 will be triggered periodically.
  • the period may be 100 Transmission Time Intervals (TTIs) , 1000 TTIs, or any other value depending on the specific design.
  • TTIs Transmission Time Intervals
  • the channel environment or UE speed etc. could be regarded as a constant, and therefore only one candidate sub-set ⁇ S ⁇ is needed.
  • This selection of the candidate set tries to identify, for example, that the current UE is locating at what kind of channel, for example LOS (line-of-sight channel) or NLOS (non-line-of-sight channel) , whether UE speed is stationary, pedestrian, medium speed, or high speed.
  • This kind of coarse channel identification may remove some ML models easily from the candidate sub-set ⁇ S ⁇ (for example, if an ML model is trained to handle high speed, then this model will not be used for stationary UEs) .
  • measurements for identifying large-scale low variant channel characteristics could include at least one of, (but not limited to) :
  • - Path loss (for example, indicating whether a UE is at cell edge or cell center) ;
  • SINR for example, indicating whether UE is in high signal quality area or low signal quality area
  • the method 500 may begin with step S510 where a 100 TTI timer is expired.
  • ML candidates may be selected for an optimal (or candidate) subset. As mentioned above, the ML candidates may be selected at least based on the measurements and/or other factors.
  • an initial (performance) value may be assigned for the newly added models in the optimal subset. However, the present disclosure is not limited thereto. In some other embodiments, for each model in the subset, no matter whether it is newly added or previously presented, a same initial value will be assigned.
  • the timer may be restarted at step S540, and the method 500 may end at step S550 until the next expiration of the timer.
  • Fig. 6 is a flow chart illustrating an exemplary method for selecting a model from a set (or subset) of candidate models for de-mapping according to an embodiment of the present disclosure.
  • the method 600 may begin with step S610 where a new transmission is received.
  • the receiver may select, at step S620, S630, S640, and S650, a model:
  • one exemplary method for implementing probability ⁇ is to generate a pure random value P at step S620, equally distributed between value 0 and 1. Each time, the receiver may generate this random value P at step S620 and compare it with ⁇ at S630, and
  • the receiver may go to random mode. It will randomly select a model from candidate optimal ML sub-set at step S650.
  • the receiver will go to greedy mode. It will select the model who has the best performance in previous reception (i.e., highest value V) at step S640.
  • the receiver may use it to process the data at step S660. Its final performance (like decoding is correct or not) may impact its value V as shown by step S670. For example, if a model′s output could be successfully decoded, its value V will increase ⁇ , or if failed, its value V will decrease ⁇ ′. In some embodiments, ⁇ may be different from or same as ⁇ ′.
  • a model′s (performance) value V is too low, i.e., always perform poorly, it could be removed from the candidate subset, which may avoid unnecessary model exploration waste.
  • a super complex ML model may be avoided and a proper ML model may be automatically selected from the candidate list, which can make a receiver more robust.
  • Fig. 7 is a flow chart of an exemplary method 700 at a communication device for model selection according to an embodiment of the present disclosure.
  • the method 700 may be performed at a user equipment (e.g., the UE 100 or the gNB 105) .
  • the method 700 may comprise steps S710 and S720.
  • the present disclosure is not limited thereto.
  • the method 700 may comprise more steps, less steps, different steps, or any combination thereof.
  • the steps of the method 700 may be performed in a different order than that described herein.
  • a step in the method 700 may be split into multiple sub-steps and performed by different entities, and/or multiple steps in the method 700 may be combined into a single step.
  • the method 700 may begin at step S710 where one of multiple modes to be used for model selection may be determined.
  • a model may be selected, in the determined mode, from a set of candidate models to be used by a feature.
  • the multiple modes may comprise at least one of: a random mode in which a model is to be randomly selected; and a greedy mode in which a model with the highest performance value is to be selected.
  • the feature may comprise at least one of: soft bit computation.
  • the method 700 may further comprise: performing the feature by at least using the selected model.
  • the set of candidate models may comprise one or more trained ML models.
  • at least one of the step of determining the mode and the step of selecting the model may be performed in response to each invocation of the feature.
  • the method 700 may further comprise: determining the set of candidate models at least based on one or more measurements related to the feature.
  • the one or more measurements may comprise, for a UE associated with the soft bit computation, at least one of: a pathloss for the UE; an ICI or an SINR for the UE; a speed of the UE; and a channel delay profile for the UE.
  • the set of candidate models may be determined periodically, and once it is determined, the determined set of candidate models may be used during a corresponding period of time.
  • the set of candidate models may be determined in response to a trigger event, and once it is determined, the determined set of candidate models may be used until another trigger event is detected.
  • the step of determining the mode may comprise at least one of: determining, with a first probability, the random mode to be used for model selection; and determining, with a second probability, the greedy mode to be used for model selection.
  • a sum of the first probability and the second probability may be 1.
  • the first probability may have a value that is continuously decreased in response to an increasing number of invocations of the feature since the last determination of the set of candidate models.
  • the step of determining the mode may comprise at least one of: generating a random number between 0 and 1; determining whether the random number is less than a probability threshold or not; and determining the random mode to be used for model selection in response to determining that the random number is less than the probability threshold and/or determining the greedy mode to be used for model selection in response to determining that the random number is greater than or equal to the probability threshold.
  • an initial performance value may be assigned to each of candidate models in the set once the set is determined.
  • the method 700 may further comprise at least one of: increasing the performance value of the determined model by a first value in response to a first outcome resulted from performing the feature by using the determined model; and decreasing the performance value of the determined model by a second value in response to a second outcome resulted from performing the feature by using the determined model.
  • the first value is different from the second value.
  • the method 700 may further comprise: removing a candidate model from the set of candidate models in response to determining that the candidate model has a lower performance value than a threshold and/or than all other candidate models for a period of time and/or for specific times.
  • Fig. 8 schematically shows an embodiment of an arrangement 800 which may be used in a user equipment (e.g., the UE 100) , a network node (e.g., the gNB 105) , or a communication device according to an embodiment of the present disclosure.
  • a processing unit 806 e.g., with a Digital Signal Processor (DSP) or a Central Processing Unit (CPU) .
  • the processing unit 806 may be a single unit or a plurality of units to perform different actions of procedures described herein.
  • the arrangement 800 may also comprise an input unit 802 for receiving signals from other entities, and an output unit 804 for providing signal (s) to other entities.
  • the input unit 802 and the output unit 804 may be arranged as an integrated entity or as separate entities.
  • the arrangement 800 may comprise at least one computer program product 808 in the form of a non-volatile or volatile memory, e.g., an Electrically Erasable Programmable Read-Only Memory (EEPROM) , a flash memory and/or a hard drive.
  • the computer program product 808 comprises a computer program 810, which comprises code/computer readable instructions, which when executed by the processing unit 806 in the arrangement 800 causes the arrangement 800 and/or the UE/network node/communication device in which it is comprised to perform the actions, e.g., of the procedure described earlier in conjunction with Fig. 5 to Fig. 7 or any other variant.
  • EEPROM Electrically Erasable Programmable Read-Only Memory
  • the computer program 810 may be configured as a computer program code structured in computer program modules 810A and 810B.
  • the code in the computer program of the arrangement 800 includes: a module 810A configured to determine one of multiple modes to be used for model selection; and a module 810B configured to select, in the determined mode, a model from a set of candidate models to be used by a feature.
  • the computer program modules could essentially perform the actions of the flow illustrated in Fig. 4 to Fig. 6, to emulate the UE, the network node, and/or the communication device.
  • the different computer program modules when executed in the processing unit 806, they may correspond to different modules in the UE, the network node, and/or the communication device.
  • code means in the embodiments disclosed above in conjunction with Fig. 8 are implemented as computer program modules which when executed in the processing unit causes the arrangement to perform the actions described above in conjunction with the figures mentioned above, at least one of the code means may in alternative embodiments be implemented at least partly as hardware circuits.
  • the processor may be a single CPU (Central processing unit) , but could also comprise two or more processing units.
  • the processor may include general purpose microprocessors; instruction set processors and/or related chips sets and/or special purpose microprocessors such as Application Specific Integrated Circuit (ASICs) .
  • the processor may also comprise board memory for caching purposes.
  • the computer program may be carried by a computer program product connected to the processor.
  • the computer program product may comprise a computer readable medium on which the computer program is stored.
  • the computer program product may be a flash memory, a Random-access memory (RAM) , a Read-Only Memory (ROM) , or an EEPROM, and the computer program modules described above could in alternative embodiments be distributed on different computer program products in the form of memories within the UE, the network node, and/or the communication device.
  • RAM Random-access memory
  • ROM Read-Only Memory
  • EEPROM Electrically Erasable programmable read-only memory
  • the computer program modules described above could in alternative embodiments be distributed on different computer program products in the form of memories within the UE, the network node, and/or the communication device.
  • FIG. 9 is a block diagram of a communication device 900 according to an embodiment of the present disclosure.
  • the communication device 900 may be, e.g., the UE 100 or the gNB 105 in some embodiments.
  • the communication device 900 may be configured to perform the method 700 as described above in connection with Fig. 7. As shown in Fig. 9, the communication device 900 may comprise a determining module 910 configured to determine one of multiple modes to be used for model selection; and a selecting module 920 configured to select, in the determined mode, a model from a set of candidate models to be used by a feature.
  • a determining module 910 configured to determine one of multiple modes to be used for model selection
  • a selecting module 920 configured to select, in the determined mode, a model from a set of candidate models to be used by a feature.
  • the above modules 910 and/or 920 may be implemented as a pure hardware solution or as a combination of software and hardware, e.g., by one or more of: a processor or a micro-processor and adequate software and memory for storing of the software, a Programmable Logic Device (PLD) or other electronic component (s) or processing circuitry configured to perform the actions described above, and illustrated, e.g., in Fig. 7.
  • the communication device 900 may comprise one or more further modules, each of which may perform any of the steps of the method 700 described with reference to Fig. 7.
  • a communication system includes a telecommunication network 3210, such as a 3GPP-type cellular network, which comprises an access network 3211, such as a radio access network, and a core network 3214.
  • the access network 3211 comprises a plurality of base stations 3212a, 3212b, 3212c, such as NBs, eNBs, gNBs or other types of wireless access points, each defining a corresponding coverage area 3213a, 3213b, 3213c.
  • Each base station 3212a, 3212b, 3212c is connectable to the core network 3214 over a wired or wireless connection 3215.
  • a first UE 3291 located in coverage area 3213c is configured to wirelessly connect to, or be paged by, the corresponding base station 3212c.
  • a second UE 3292 in coverage area 3213a is wirelessly connectable to the corresponding base station 3212a. While a plurality of UEs 3291, 3292 are illustrated in this example, the disclosed embodiments are equally applicable to a situation where a sole UE is in the coverage area or where a sole UE is connecting to the corresponding base station 3212.
  • the telecommunication network 3210 is itself connected to a host computer 3230, which may be embodied in the hardware and/or software of a standalone server, a cloud-implemented server, a distributed server or as processing resources in a server farm.
  • the host computer 3230 may be under the ownership or control of a service provider, or may be operated by the service provider or on behalf of the service provider.
  • the connections 3221, 3222 between the telecommunication network 3210 and the host computer 3230 may extend directly from the core network 3214 to the host computer 3230 or may go via an optional intermediate network 3220.
  • the intermediate network 3220 may be one of, or a combination of more than one of, a public, private or hosted network; the intermediate network 3220, if any, may be a backbone network or the Internet; in particular, the intermediate network 3220 may comprise two or more sub-networks (not shown) .
  • the communication system of Fig. 10 as a whole enables connectivity between one of the connected UEs 3291, 3292 and the host computer 3230.
  • the connectivity may be described as an over-the-top (OTT) connection 3250.
  • the host computer 3230 and the connected UEs 3291, 3292 are configured to communicate data and/or signaling via the OTT connection 3250, using the access network 3211, the core network 3214, any intermediate network 3220 and possible further infrastructure (not shown) as intermediaries.
  • the OTT connection 3250 may be transparent in the sense that the participating communication devices through which the OTT connection 3250 passes are unaware of routing of uplink and downlink communications.
  • a base station 3212 may not or need not be informed about the past routing of an incoming downlink communication with data originating from a host computer 3230 to be forwarded (e.g., handed over) to a connected UE 3291. Similarly, the base station 3212 need not be aware of the future routing of an outgoing uplink communication originating from the UE 3291 towards the host computer 3230.
  • a host computer 3310 comprises hardware 3315 including a communication interface 3316 configured to set up and maintain a wired or wireless connection with an interface of a different communication device of the communication system 3300.
  • the host computer 3310 further comprises processing circuitry 3318, which may have storage and/or processing capabilities.
  • the processing circuitry 3318 may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions.
  • the host computer 3310 further comprises software 3311, which is stored in or accessible by the host computer 3310 and executable by the processing circuitry 3318.
  • the software 3311 includes a host application 3312.
  • the host application 3312 may be operable to provide a service to a remote user, such as a UE 3330 connecting via an OTT connection 3350 terminating at the UE 3330 and the host computer 3310. In providing the service to the remote user, the host application 3312 may provide user data which is transmitted using the OTT connection 3350.
  • the communication system 3300 further includes a base station 3320 provided in a telecommunication system and comprising hardware 3325 enabling it to communicate with the host computer 3310 and with the UE 3330.
  • the hardware 3325 may include a communication interface 3326 for setting up and maintaining a wired or wireless connection with an interface of a different communication device of the communication system 3300, as well as a radio interface 3327 for setting up and maintaining at least a wireless connection 3370 with a UE 3330 located in a coverage area (not shown in Fig. 11) served by the base station 3320.
  • the communication interface 3326 may be configured to facilitate a connection 3360 to the host computer 3310.
  • the connection 3360 may be direct or it may pass through a core network (not shown in Fig.
  • the hardware 3325 of the base station 3320 further includes processing circuitry 3328, which may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions.
  • the base station 3320 further has software 3321 stored internally or accessible via an external connection.
  • the communication system 3300 further includes the UE 3330 already referred to.
  • Its hardware 3335 may include a radio interface 3337 configured to set up and maintain a wireless connection 3370 with a base station serving a coverage area in which the UE 3330 is currently located.
  • the hardware 3335 of the UE 3330 further includes processing circuitry 3338, which may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions.
  • the UE 3330 further comprises software 3331, which is stored in or accessible by the UE 3330 and executable by the processing circuitry 3338.
  • the software 3331 includes a client application 3332.
  • the client application 3332 may be operable to provide a service to a human or non-human user via the UE 3330, with the support of the host computer 3310.
  • an executing host application 3312 may communicate with the executing client application 3332 via the OTT connection 3350 terminating at the UE 3330 and the host computer 3310.
  • the client application 3332 may receive request data from the host application 3312 and provide user data in response to the request data.
  • the OTT connection 3350 may transfer both the request data and the user data.
  • the client application 3332 may interact with the user to generate the user data that it provides.
  • the host computer 3310, base station 3320 and UE 3330 illustrated in Fig. 11 may be identical to the host computer 3230, one of the base stations 3212a, 3212b, 3212c and one of the UEs 3291, 3292 of Fig. 10, respectively.
  • the inner workings of these entities may be as shown in Fig. 11 and independently, the surrounding network topology may be that of Fig. 10.
  • the OTT connection 3350 has been drawn abstractly to illustrate the communication between the host computer 3310 and the use equipment 3330 via the base station 3320, without explicit reference to any intermediary devices and the precise routing of messages via these devices.
  • Network infrastructure may determine the routing, which it may be configured to hide from the UE 3330 or from the service provider operating the host computer 3310, or both. While the OTT connection 3350 is active, the network infrastructure may further take decisions by which it dynamically changes the routing (e.g., on the basis of load balancing consideration or reconfiguration of the network) .
  • the wireless connection 3370 between the UE 3330 and the base station 3320 is in accordance with the teachings of the embodiments described throughout this disclosure.
  • One or more of the various embodiments improve the performance of OTT services provided to the UE 3330 using the OTT connection 3350, in which the wireless connection 3370 forms the last segment. More precisely, the teachings of these embodiments may improve the latency and power consumption and thereby provide benefits such as reduced user waiting time, better responsiveness, extended battery lifetime.
  • a measurement procedure may be provided for the purpose of monitoring data rate, latency, and other factors on which the one or more embodiments improve.
  • the measurement procedure and/or the network functionality for reconfiguring the OTT connection 3350 may be implemented in the software 3311 of the host computer 3310 or in the software 3331 of the UE 3330, or both.
  • sensors (not shown) may be deployed in or in association with communication devices through which the OTT connection 3350 passes; the sensors may participate in the measurement procedure by supplying values of the monitored quantities exemplified above, or supplying values of other physical quantities from which software 3311, 3331 may compute or estimate the monitored quantities.
  • the reconfiguring of the OTT connection 3350 may include message format, retransmission settings, preferred routing etc.; the reconfiguring need not affect the base station 3320, and it may be unknown or imperceptible to the base station 3320. Such procedures and functionalities may be known and practiced in the art.
  • measurements may involve proprietary UE signaling facilitating the host computer′s 3310 measurements of throughput, propagation times, latency, and the like.
  • the measurements may be implemented in that the software 3311, 3331 causes messages to be transmitted, in particular empty or ′dummy′ messages, using the OTT connection 3350 while it monitors propagation times, errors etc.
  • Fig. 12 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment.
  • the communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 12 will be included in this section.
  • the host computer provides user data.
  • the host computer provides the user data by executing a host application.
  • the host computer initiates a transmission carrying the user data to the UE.
  • the base station transmits to the UE the user data which was carried in the transmission that the host computer initiated, in accordance with the teachings of the embodiments described throughout this disclosure.
  • the UE executes a client application associated with the host application executed by the host computer.
  • Fig. 13 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment.
  • the communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 13 will be included in this section.
  • the host computer provides user data.
  • the host computer provides the user data by executing a host application.
  • the host computer initiates a transmission carrying the user data to the UE. The transmission may pass via the base station, in accordance with the teachings of the embodiments described throughout this disclosure.
  • the UE receives the user data carried in the transmission.
  • Fig. 14 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment.
  • the communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 14 will be included in this section.
  • the UE receives input data provided by the host computer.
  • the UE provides user data.
  • the UE provides the user data by executing a client application.
  • the UE executes a client application which provides the user data in reaction to the received input data provided by the host computer.
  • the executed client application may further consider user input received from the user.
  • the UE initiates, in an optional third substep 3630, transmission of the user data to the host computer.
  • the host computer receives the user data transmitted from the UE, in accordance with the teachings of the embodiments described throughout this disclosure.
  • Fig. 15 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment.
  • the communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 15 will be included in this section.
  • the base station receives user data from the UE.
  • the base station initiates transmission of the received user data to the host computer.
  • the host computer receives the user data carried in the transmission initiated by the base station.

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Computational Linguistics (AREA)
  • Signal Processing (AREA)
  • Evolutionary Computation (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computer Networks & Wireless Communication (AREA)
  • Mobile Radio Communication Systems (AREA)

Abstract

A method at a communication device in a telecommunications network for model selection comprises: determining one of multiple modes to be used for model selection; and selecting, in the determined mode, a model from a set of candidate models to be used by a feature.

Description

AUTOMATIC MODEL SELECTION Technical Field
The present disclosure is related to the field of telecommunication, and in particular, to a communication device and a method for automatic model selection.
Background
With the development of the electronic and telecommunications technologies, mobile devices, such as a mobile phone, a smart phone, a laptop, a tablet, a vehicle mounted device, becomes an important part of our daily lives. To support a numerous number of mobile devices, a highly efficient Radio Access Network (RAN) , such as a fifth generation (5G) New Radio (NR) RAN, will be required.
Modulation mappers are one of the most important functional blocks that can significantly impact the performance of a 5G RAN. A variety of different modulation mappers are defined in 3rd Generation Partnership Project (3GPP) technical specifications (TS) for 5G RAN technology, such as, π/2 -Binary Phase Shift Keying (π/2-BPSK) , BPSK, Quadrature Phase Shift Keying (QPSK) , 16 Quadrature Amplitude Modulation (16QAM) , 64QAM, 256 QAM, 1024QAM, etc.
Phase Shift Keying (PSK) : 5G technology implements BPSK and QPSK as the lower order modulation formats. Although they will provide the slower data throughput, they will also provide the most robust link and as such it can be used when signal levels are low or when interference is high. Another form of PSK called π/2-BPSK is used in conjunction with Discrete Fourier Transform spread Orthogonal Frequency Division Multiplexing (DFT-s-OFDM) on the uplink.
Quadrature Amplitude Modulation (QAM) : QAM enables the data throughput to be increased. Formats used within 5G mobile communications system include 16QAM, 64QAM and 256QAM. Further, with the evolution of 5G technology, other QAM formats may also be used, for example 1024QAM. The higher the order of modulation, the greater the throughput, although the penalty is the noise resilience. Therefore, 256QAM or higher order formats are only used when link quality is good, and it reduces to 64QAM and then 16QAM etc., as the link deteriorates. It is a balance between data throughput and resilience.
QAM is widely used for data transmission as it enables better levels of spectral efficiency than other formats of modulation. QAM uses two carriers on the same frequency shifted by 90° which are modulated by two data streams: an I or In-phase component and a Q or Quadrature component.
Although 3GPP defines some modulation mappers in Subclause 5.1 of 3GPP TS 38.211, definitions of corresponding demodulation de-mappers are not provided by 3GPP and therefore they are vendor specific.
Summary
According to a first aspect of the present disclosure, a method at a communication device in a telecommunications network for model selection is provided. The method comprises: determining one of multiple modes to be used for model selection; and selecting, in the determined mode, a model from a set of candidate models to be used by a feature.
In some embodiments, the multiple modes comprise at least one of: a random mode in which a model is to be randomly selected; and a greedy mode in which a model with the highest performance value is to be selected. In some embodiments, the feature comprises at least one of: soft bit computation. In some embodiments, the method further comprises: performing the feature by at least using the selected model. In some embodiments, the set of candidate models comprise one or more trained Machine Learning (ML) models. In some embodiments, at least one of the step of determining the mode and the step of selecting the model is performed in response to each invocation of the feature.
In some embodiments, before the step of selecting the model, the method further comprises: determining the set of candidate models at least based on one or more measurements related to the feature. In some embodiments, when the feature comprises the soft bit computation, the one or more measurements comprise, for a UE associated with the soft bit computation, at least one of: a pathloss for the UE; an Inter-Cell Interference (ICI) or a Signal to Interference plus Noise Ratio (SINR) for the UE; a speed of the UE; and a channel delay profile for the UE. In some embodiments, the set of candidate models is determined periodically, and once it is determined, the determined set of candidate models is used during a corresponding period of time.
In some embodiments, the set of candidate models is determined in response to a trigger event, and once it is determined, the determined set of candidate models is used until another trigger event is detected. In some embodiments, the step of determining the mode comprises at least one of: determining, with a first probability, the random mode to be used for model selection; and determining, with a second probability, the greedy mode to be used for model selection. In some embodiments, a sum of the first probability and the second probability is 1. In some embodiments, the first probability has a value that is continuously decreased in response to an increasing number of invocations of the feature since the last determination of the set of candidate models. In some embodiments, the step of determining the mode comprises at least one of: generating a random number between 0 and 1; determining whether the random number is less than a probability threshold or not; and determining the random mode to be used for model selection in response to determining that the random number is less than the probability threshold and/or determining the greedy mode to be used for model selection in response to determining that the random number is greater than or equal to the probability threshold.
In some embodiments, an initial performance value is assigned to each of candidate models in the set once the set is determined. In some embodiments, the method further comprises at least one of: increasing the performance value of the determined model by a first value in response to a first outcome resulted from performing the feature by using the determined model; and decreasing the performance value of the determined model by a second value in response to a second outcome resulted from performing the feature by using the determined model. In some embodiments, the first value is different from the second value. In some embodiments, the method further comprises: removing a candidate model from the set of candidate models in response to determining that the candidate model has a lower performance value than a threshold and/or than all other candidate models for a period of time and/or for specific times.
According to a second aspect of the present disclosure, a communication device is provided. The communication device comprises: a processor; a memory storing instructions which, when executed by the processor, cause the processor to perform the method of any of the first aspect.
According to a third aspect of the present disclosure, a communication device is provided. The communication device comprises: a determining module configured to determine one of multiple modes to be used for model selection; and a selecting module configured to select, in the determined mode, a model from a set of candidate models to be used by a feature. In some embodiments, the communication device may comprise one or more further modules configured to perform any of the method of the first aspect.
According to a fourth aspect of the present disclosure, a computer program comprising instructions is provided. The instructions, when executed by at least one processor, cause the at least one processor to carry out the method of any of the first aspect.
According to a fifth aspect of the present disclosure, a carrier containing the computer program of the fourth aspect is provided. In some embodiments, the carrier is one of an electronic signal, optical signal, radio signal, or computer readable storage medium.
Brief Description of the Drawings
Fig. 1 is a diagram illustrating an exemplary telecommunications network in which UEs and gNB may be operated according to an embodiment of the present disclosure.
Fig. 2A and Fig. 2B are diagrams illustrating exemplary constellations for de-mapping from symbols to bits according to an embodiment of the present disclosure.
Fig. 3 is a diagram illustrating an exemplary convolutional neural network (CNN) based on which an ML model for de-mapping may be trained and used according to an embodiment of the present disclosure.
Fig. 4 is a diagram illustrating comparison between performance of a legacy de-mapper and an ML assisted de-mapper according to an embodiment of the present disclosure.
Fig. 5 is a flow chart illustrating an exemplary method for creating a set of candidate models according to an embodiment of the present disclosure.
Fig. 6 is a flow chart illustrating an exemplary method for selecting a model from a set of candidate models for de-mapping according to an embodiment of the present disclosure.
Fig. 7 is a flow chart illustrating an exemplary method at a communication device for model selection according to an embodiment of the present disclosure.
Fig. 8 schematically shows an embodiment of an arrangement which may be used in a UE, a network node, or a communication device according to an embodiment of the present disclosure.
Fig. 9 is a block diagram of an exemplary communication device according to an embodiment of the present disclosure.
Fig. 10 schematically illustrates a telecommunication network connected via an intermediate network to a host computer according to an embodiment of the present disclosure.
Fig. 11 is a generalized block diagram of a host computer communicating via a base station with a user equipment over a partially wireless connection according to an embodiment of the present disclosure.
Fig. 12 to Fig. 15 are flowcharts illustrating methods implemented in a communication system including a host computer, a base station, and a user equipment according to an embodiment of the present disclosure.
Detailed Description
Hereinafter, the present disclosure is described with reference to embodiments shown in the attached drawings. However, it is to be understood that those descriptions are just provided for illustrative purpose, rather than limiting the present disclosure. Further, in the following, descriptions of known structures and techniques are omitted so as not to unnecessarily obscure the concept of the present disclosure.
Those skilled in the art will appreciate that the term ″exemplary″ is used herein to mean ″illustrative, ″ or ″serving as an example, ″ and is not intended to imply that a particular embodiment is preferred over another or that a particular feature is essential. Likewise, the terms ″first″ , ″second″ , ″third″ , ″fourth, ″ and similar terms, are used simply to distinguish one particular instance of an item or feature from another, and do not indicate a particular order or arrangement, unless the context clearly indicates otherwise. Further, the term ″step, ″ as used herein, is meant to be synonymous with ″operation″ or ″action. ″ Any description herein of a sequence of steps does not imply that these operations must be carried out in a particular order, or even that these  operations are carried out in any order at all, unless the context or the details of the described operation clearly indicates otherwise.
Conditional language used herein, such as ″can, ″ ″might, ″ ″may, ″ ″e.g., ″ and the like, unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments include, while other embodiments do not include, certain features, elements and/or states. Thus, such conditional language is not generally intended to imply that features, elements and/or states are in any way required for one or more embodiments or that one or more embodiments necessarily include logic for deciding, with or without author input or prompting, whether these features, elements and/or states are included or are to be performed in any particular embodiment. Also, the term ″or″ is used in its inclusive sense (and not in its exclusive sense) so that when used, for example, to connect a list of elements, the term ″or″ means one, some, or all of the elements in the list. Further, the term ″each, ″ as used herein, in addition to having its ordinary meaning, can mean any subset of a set of elements to which the term ″each″ is applied.
The term ″based on″ is to be read as ″based at least in part on. ″ The term ″one embodiment″ and ″an embodiment″ are to be read as ″at least one embodiment. ″ The term ″another embodiment″ is to be read as ″at least one other embodiment. ″ Other definitions, explicit and implicit, may be included below. In addition, language such as the phrase ″at least one of X, Y and Z, ″ unless specifically stated otherwise, is to be understood with the context as used in general to convey that an item, term, etc. may be either X, Y, or Z, or a combination thereof.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limitation of example embodiments. As used herein, the singular forms ″a″ , ″an″ , and ″the″ are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms ″comprises″ , ″comprising″ , ″has″ , ″having″ , ″includes″ and/or ″including″ , when used herein, specify the presence of stated features, elements, and/or components etc., but do not preclude the presence or addition of one or more other features, elements, components and/or combinations thereof. It will be also understood that the terms ″connect (s) , ″ ″connecting″ , ″connected″ , etc. when used herein, just mean that there is an electrical or communicative connection between two elements  and they can be connected either directly or indirectly, unless explicitly stated to the contrary.
Of course, the present disclosure may be carried out in other specific ways than those set forth herein without departing from the scope and essential characteristics of the disclosure. One or more of the specific processes discussed below may be carried out in any electronic device comprising one or more appropriately configured processing circuits, which may in some embodiments be embodied in one or more application-specific integrated circuits (ASICs) . In some embodiments, these processing circuits may comprise one or more microprocessors, microcontrollers, and/or digital signal processors programmed with appropriate software and/or firmware to carry out one or more of the operations described above, or variants thereof. In some embodiments, these processing circuits may comprise customized hardware to carry out one or more of the functions described above. The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive.
Although multiple embodiments of the present disclosure will be illustrated in the accompanying Drawings and described in the following Detailed Description, it should be understood that the disclosure is not limited to the disclosed embodiments, but instead is also capable of numerous rearrangements, modifications, and substitutions without departing from the present disclosure that as will be set forth and defined within the claims.
Further, please note that although the following description of some embodiments of the present disclosure is given in the context of 5G NR, the present disclosure is not limited thereto. In fact, as long as automatic model selection is involved, the inventive concept of the present disclosure may be applicable to any appropriate communication architecture, for example, to Global System for Mobile Communications (GSM) /General Packet Radio Service (GPRS) , Enhanced Data Rates for GSM Evolution (EDGE) , Code Division Multiple Access (CDMA) , Wideband CDMA (WCDMA) , Time Division -Synchronous CDMA (TD-SCDMA) , CDMA2000, Worldwide Interoperability for Microwave Access (WiMAX) , Wireless Fidelity (Wi-Fi) , 4th Generation Long Term Evolution (LTE) , LTE-Advance (LTE-A) , or 5G NR, etc. Therefore, one skilled in the arts could readily understand that the terms used herein may also refer to their equivalents in any other infrastructure. For example, the term ″User Equipment″ or ″UE″ used herein may refer to a terminal device, a mobile device, a mobile terminal, a mobile  station, a user device, a user terminal, a wireless device, a wireless terminal, or any other equivalents. For another example, the term ″gNB″ used herein may refer to a network node, a base station, a base transceiver station, an access point, a hot spot, a NodeB, an Evolved NodeB, a network element, or any other equivalents. Further, the term ″communication device″ may refer to any of those listed above or any device that can communicate with another device.
Further, the following 3GPP document is incorporated herein by reference in its entirety:
- 3GPP TS 38.211 V17.0.0 (2021-12) , Technical Specification, 3rd Generation Partnership Project; Technical Specification Group Radio Access Network; NR; Physical channels and modulation (Release 17) .
Fig. 1 is a diagram illustrating an exemplary telecommunications network 10 in which UE #1 100-1, UE #2 100-2, and gNB 105 may be operated according to an embodiment of the present disclosure. Although the telecommunications network 10 is a network defined in the context of 5G NR, the present disclosure is not limited thereto.
As shown in Fig. 1, the network 10 may comprise one or more UEs 100-1 and 100-2 (collectively, UE (s) 100) and a RAN node 105, which could be a base station, a Node B, an evolved NodeB (eNB) , a gNB, or an AN node which provides the UEs 100 with access to the network. Further, the network 10 may comprise its core network portion that is not shown in Fig. 1.
However, the present disclosure is not limited thereto. In some other embodiments, the network 10 may comprise additional nodes, less nodes, or some variants of the existing nodes shown in Fig. 1. For example, in a network with the 4G architecture, the entities (e.g., an eNB) which perform these functions may be different from those (e.g., the gNB 105) shown in Fig. 1. For another example, in a network with a mixed 4G/5G architecture, some of the entities may be same as those shown in Fig. 1, and others may be different. Further, although two UEs 100 and one gNB 105 are shown in Fig. 1, the present disclosure is not limited thereto. In some other embodiments, any number of UEs and/or any number of gNBs may be comprised in the network 10.
As shown in Fig. 1, the UEs 100 may be communicatively connected to the gNB 105 which in turn may be communicatively connected to a corresponding Core Network  (CN) and then the Internet, such that the UEs 100 may finally communicate its user plane data with other devices outside the network 10, for example, via the gNB 105.
The 5G Evolution/6G is envisioned to enable transmission at higher carrier frequencies and to provide opportunities with a larger spectrum to be allocated for transmissions. However, transmission at these frequencies, compared to the already existing frequency bands, is even more challenging due to the signal distortions induced by hardware impairments e.g., oscillator′s phase noise, power amplifier′s nonlinearities, and digital-to-analog converter′s quantization noise.
This distortion violates the orthogonal reception of modulated signals (e.g., DFT-s-OFDMl modulated signals or Cyclic Prefix -OFDM (CP-OFDM) modulated signals) , and hence, causes inter-carrier interference (ICI) in the received signal. Typical phase noise compensation approaches in 5G systems try to estimate phase noise from transmitted phase tracking reference signals (PTRS) and more demodulation reference signals (DMRS) , and compensate for the common phase error which is the average phase noise across the symbol.
As in many digital modulation schemes, a constellation diagram is useful for QAM (e.g., the constellation for 16QAM shown in Fig. 2A and Fig. 2B) . Fig. 2A and Fig. 2B are diagrams illustrating exemplary constellations for de-mapping from symbols to bits according to an embodiment of the present disclosure. To be specific, Fig. 2A shows an ideal constellation and each point thereof could be an ideal symbol detected by a receiver, and Fig. 2B shows an actual symbol 200 detected by a receiver when compared and overlapped with the ideal constellation shown in Fig. 2A.
As shown in Fig. 2A and Fig. 2B, in QAM, ideal constellation points are usually arranged in a square grid with equal vertical and horizontal spacing, although other configurations are possible (e.g., a hexagonal or triangular grid) . In digital telecommunications, the data is usually binary, so the number of points in the grid is typically a power of 2 (2, 4, 8, ... ) , corresponding to the number of bits per symbol. The simplest and most commonly used QAM constellations consist of points arranged in a square, i.e. 16-QAM, 64-QAM and 256-QAM (even powers of two) . Non-square constellations, such as Cross-QAM, can offer greater efficiency but are rarely used because of the cost of increased modem complexity.
By moving to a higher-order constellation, it is possible to transmit more bits per symbol. However, if the mean energy of the constellation is to remain the same (by way  of making a fair comparison) , the points must be closer together and are thus more susceptible to noise and other corruption; this results in a higher bit error rate and so higher-order QAM can deliver more data less reliably than lower-order QAM, for constant mean constellation energy. Using higher-order QAM without increasing the bit error rate requires a higher signal-to-noise ratio (SNR) by increasing signal energy, reducing noise, or both.
In theory, when a symbol transmitted by a transmitter is detected by the receiver, for example, a symbol mapped to a sequence of bits ″1100″  (b 0b 1b 2b 3) as shown in Fig. 2A, the receiver may determine that the symbol shall be de-mapped to ″1100″ since the symbol has I/Q components that are perfectly matched with the ideal constellation point corresponding to ″1100″ .
However, in practice, such ideal symbols are typically not possible due to interference and noise. As shown in Fig. 2B, a symbol that was mapped to ″1100″ by the transmitter may be detected by the receiver as a symbol indicated by the indicator 200, which is located among and close to one or more ideal constellation points depending on its detected I/Q components. In such a case, the symbol 200 may be de-mapped to soft bits (instead of hard bits) , and finally decoded to ″1100″ , ″1110″ , ″1101″ or the like.
In some embodiments, soft bits can be computed in a simplified way as follows. Assuming a received coded sequence is:
y=c+n
where y is the received coded sequence, c is the modulated coded sequence taking values in the alphabet: α 16QAM = {±1 ± 1j, ±1 ± 3j, ±3 ± 1j, ±3 ± 3j} as also shown in Fig. 2A and Fig. 2B, and n is the Additive White Gaussian Noise (AWGN) following the probability distribution function:
Figure PCTCN2022088430-appb-000001
with mean μ = 0 and variance
Figure PCTCN2022088430-appb-000002
the value
Figure PCTCN2022088430-appb-000003
is the double sided power spectral density (PSD) of the noise.
For demodulation, a probability that the bit b m (m = 0, 1, 2, or 3) was transmitted given the received y, i.e. P (b m|y) , needs to be maximized. This criterion is called Maximum a posteriori probability (MAP) .
Using Bayes rule,
Figure PCTCN2022088430-appb-000004
Note: The probability that all constellation points occur are equally likely, so maximizing P (b m|y) is equivalent to maximizing P (y|b m) .
Taking b 0 for example, the bit mapping for the bit b 0 with 16QAM coded mapping is shown in Fig. 2A and Fig. 2B. When b 0 toggles from 0 to 1, only the real part of the constellation is affected.
When the b 0 is 1, the real part of the QAM constellation takes values -3 or -1. The conditional probability of the received signal given b 0 is 1 is:
Figure PCTCN2022088430-appb-000005
When the b 0 is 0, the real part of the QAM constellation takes values +1 or +3. The conditional probability given b 0 is 0 is:
Figure PCTCN2022088430-appb-000006
A likelihood ratio can be defined if:
b 0 = 1, 
Figure PCTCN2022088430-appb-000007
The likelihood ratio for b 0 is:
Figure PCTCN2022088430-appb-000008
This equation can be simplified by segmentation.
When y re < -2, then it can be assumed that relative contribution by constellation -1 in the numerator and +3 in the denominator is less and can be ignored. Therefore, the likelihood ratio reduces to:
Figure PCTCN2022088430-appb-000009
Taking logarithm on both sides:
Figure PCTCN2022088430-appb-000010
Figure PCTCN2022088430-appb-000011
Similarly, when -2 ≤ y re < 0 or 0 ≤ y re < 2, then it can be assumed that relative contribution by constellation -3 in the numerator and +3 in the denominator is less and can be ignored. Therefore, the likelihood ratio reduces to:
Figure PCTCN2022088430-appb-000012
Taking logarithm on both sides:
Figure PCTCN2022088430-appb-000013
Similarly, when 2 ≤ y re, then it can be assumed that relative contribution by constellation -3 in the numerator and +1 in the denominator is less and can be ignored. Therefore, the likelihood ratio reduces to:
Figure PCTCN2022088430-appb-000014
Taking logarithm on both sides:
Figure PCTCN2022088430-appb-000015
Please note that as the factor
Figure PCTCN2022088430-appb-000016
is common to all the terms, and it can be removed. Therefore, the soft bit for bit b 0 is:
Figure PCTCN2022088430-appb-000017
The bit mapping for the bit b 2 with 16QAM coded mapping is also shown in Fig. 2A and Fig. 2B. Similar to b 0, when b 2 toggles from 0 to 1, only the real part of the constellation is affected.
When the b 2 is 1, the real part of the QAM constellation takes values -3 or +3. The conditional probability of the received signal given b 2 is 1 is:
Figure PCTCN2022088430-appb-000018
When the b 2 is 0, the real part of the QAM constellation takes values -1 or +1. The conditional probability given b 2 is 0 is:
Figure PCTCN2022088430-appb-000019
A likelihood ratio can be defined if:
b 2=1, 
Figure PCTCN2022088430-appb-000020
The likelihood ratio for b 2 is:
Figure PCTCN2022088430-appb-000021
When y re < -2 or -2 ≤ y re < 0, then it can be assumed that relative contribution by constellation +3 in the numerator and +1 in the denominator is less and can be ignored. Therefore, the likelihood ratio reduces to:
Figure PCTCN2022088430-appb-000022
Taking logarithm on both sides:
Figure PCTCN2022088430-appb-000023
Figure PCTCN2022088430-appb-000024
Similarly, when 0 ≤ y re < 2 or 2 ≤ y re, then it can be assumed that relative contribution by constellation -3 in the numerator and -1 in the denominator is less and can be ignored. Therefore, the likelihood ratio reduces to:
Figure PCTCN2022088430-appb-000025
Taking logarithm on both sides:
Figure PCTCN2022088430-appb-000026
Please note that as the factor
Figure PCTCN2022088430-appb-000027
is common to all the terms, it can be removed. Therefore, the soft bit for bit b 2 is:
Figure PCTCN2022088430-appb-000028
The soft bit for bit b 2 can be further simplified to:
sb (b 2) = |y re | -2, for all y re
It is easy to observe that the soft bits for bits b 1 and b 3 are identical to softbits for b 0 and b 2, respectively, except that the decisions are based on the imaginary component y im of the received symbol.
Therefore, the soft bit for bit b 1 is:
Figure PCTCN2022088430-appb-000029
The soft bit for bit b 3 is:
sb (b 3) = |y im| -2, for all y im
In this way, soft bits can be approximately calculated in a very quick manner. For example, referring back to Fig. 2B, the symbol 200 may have a real part of -1.5 and an imaginary part of -1, then the corresponding soft bits are:
Figure PCTCN2022088430-appb-000030
Figure PCTCN2022088430-appb-000031
Figure PCTCN2022088430-appb-000032
Figure PCTCN2022088430-appb-000033
That is, the symbol 200 may be de-mapped as soft bits (1.5, 1, -0.5, -1) , which may then be converted to hard bits. A simple method to determine hard bits from the soft bits is to find the minimum Euclidean distance between the soft bits and the ideal constellation points (e.g., (1, 1, 0, 0) , (1, 1, 1, 0) ) in the vector space. In the case shown in Fig. 2B, the ideal constellation point that is closest to the soft bits of the symbol 200 is the constellation point corresponding to the hard bits (1, 1, 0, 0) . However, the present disclosure is not limited thereto. In some other embodiments, different methods for soft bit calculation and/or soft-bit to hard-bit conversion may be applicable.
As can be seen from the above description, this method heavily relies on an assumption that the noise introduced into the received symbol is AWGN. However, in practice, some residual ICI due to the phase noise fluctuations between modulated symbols also remains, which influences the performance of the receiver. Considering that full compensation of the phase noise is not typically feasible or requires a very high density of the reference signals, it is likely that even after phase noise compensation, some residual ICI remains. The residual ICI in the DFT-s-OFDM or CP-OFDM reception leads to the deviation of the distribution of the noise in the equalized signal at the receiver side from uncorrelated, which is typically assumed in conventional receivers as mentioned above. This mismatch causes degraded soft bit computation and leads to sub-optimum performance of the soft decoding algorithms for Low Density Parity Check (LDPC) or Turbo codes.
Considering that base stations have more computational resources compared to user equipment, receivers in uplink transmission scenarios would be suitable for the operation of machine learning techniques. However, the present disclosure is not limited thereto. In some embodiments, the machine learning techniques may also be applicable to UEs or other nodes than the base stations. Some Machine learning based de-mapper technologies has been studied and implemented in PoC (Prototype of Concept) in Q4 2021.
Fig. 3 is a diagram illustrating an exemplary CNN 300 based on which an ML model for de-mapping may be trained and used according to an embodiment of the present disclosure.
In deep learning, a convolutional neural network (CNN, or ConvNet) is a class of artificial neural network (ANN) , most commonly applied to analyze visual imagery. However, it may also be used for de-mapping, which will be described below in details.
As shown in Fig. 3, the CNN 300 may comprise one or more layers, such as one or more convolutional layers (e.g., convolutional layers 315 and 325) , one or more pooling layers (e.g., pooling layers 320 and 330) , a flatten layer (e.g., a flatten layer 335) , and/or a fully connected layer (e.g., a full connected layer 340) . Further, the CNN 300 may comprise one or more other types of layers, such as batch normalization layers, dilated layers, loss layers, etc. However, the present disclosure is not limited thereto. In some other embodiments, other ML-based networks/models/algorithms or even other none-ML based technologies may also be applicable for model selection.
Convolutional networks were inspired by biological processes in that the connectivity pattern between neurons resembles the organization of the animal visual cortex. Individual cortical neurons respond to stimuli only in a restricted region of the visual field known as the receptive field (e.g., the receptive field 360 shown in Fig. 3) . The receptive fields of different neurons partially overlap such that they cover the entire visual field.
The name ″convolutional neural network″ indicates that the network employs a mathematical operation called convolution. Convolutional networks are a specialized type of neural networks that use convolution in place of general matrix multiplication in at least one of their layers.
In a CNN (e.g., the CNN 300) , the input is a tensor with a shape: (number of inputs) x (input height) x (input width) x (input channels) . After passing through a  convolutional layer, the input (e.g., an image, or in the embodiment shown in Fig. 3, received symbols) becomes abstracted to a feature map, also called an activation map, with shape: (number of inputs) x (feature map height) x (feature map width) x (feature map channels) . For example, as shown in Fig. 3, the output of the convolutional layer #1 315 may be a feature map 365 with shape: (1, 24, 28, 28) . For another example, as also shown in Fig. 3, the output of the convolutional layer #2 325 may be a feature map 365 with shape: (1, 24, 10, 10) . Please note that Fig. 3 is only provided for illustration purpose, and therefore the shapes of the feature maps may not be scaled properly.
Convolutional layers convolve the input and pass its result to the next layer. This is similar to the response of a neuron in the visual cortex to a specific stimulus. Each convolutional neuron processes data only for its receptive field. Although fully connected feedforward neural networks can be used to learn features and classify data, this architecture is generally impractical for larger inputs. It would require a very high number of neurons, even in a shallow architecture. Instead, convolution reduces the number of free parameters, allowing the network to be deeper. For example, regardless of input size, using 5 x 5 kernels, each with the same shared weights, requires only 25 learnable parameters. Using regularized weights over fewer parameters avoids the vanishing gradients and exploding gradients problems seen during backpropagation in traditional neural networks. Furthermore, convolutional neural networks are ideal for data with a grid-like topology as spatial relations between separate features are taken into account during convolution and/or pooling.
Convolutional networks may also include local and/or global pooling layers (e.g., the pooling layers 320 and 330 shown in Fig. 3) along with traditional convolutional layers. Pooling layers reduce the dimensions of data by combining the outputs of neuron clusters at one layer into a single neuron in the next layer. Local pooling combines small clusters, tiling sizes such as 2 x 2 are commonly used. Global pooling acts on all the neurons of the feature map. There are two common types of pooling in popular use: max and average. Max pooling uses the maximum value of each local cluster of neurons in the feature map, while average pooling takes the average value. In some embodiments, the pooling layers #1 320 and the pooling layer #2 330 may be max pooling layers.
A flatten layer may flatten a multi-dimension feature map into a one-dimension feature vector, for example, by concatenating vectors in the feature map one to another.  As shown in Fig. 3, a flatten layer 335 may convert the output of the pooling layer #2 330 into a one-dimension vector that is input to the fully connected layer 340.
Fully connected layers or dense layers (e.g., the fully connected layers 340 shown in Fig. 3) connect every neuron in one layer to every neuron in another layer. It is the same as a traditional multilayer perceptron neural network (MLP) . The flattened vector goes through a fully connected layer to determine the output, e.g., soft bits or hard bits in the embodiment shown in Fig. 3.
In neural networks, each neuron receives input from some number of locations in the previous layer. In a convolutional layer, each neuron receives input from only a restricted area of the previous layer called the neuron′s receptive field (e.g., the receptive field 360) . Typically, the area is a square (e.g., a 5 by 5 neuron) . Whereas, in a fully connected layer, the receptive field is the entire previous layer (e.g., as shown by the fully connected layer 340) . Thus, in each convolutional layer, each neuron takes input from a larger area in the input than previous layers. This is due to applying the convolution over and over, which takes into account a value, as well as its surrounding values. When using dilated layers, the number of pixels in the receptive field remains constant, but the field is more sparsely populated as its dimensions grow when combining the effect of several layers.
Each neuron in a neural network computes an output value by applying a specific function to the input values received from the receptive field in the previous layer. The function that is applied to the input values is determined by a vector of weights and a bias. Learning consists of iteratively adjusting these biases and weights.
The vector of weights and the bias are called filters and represent particular features of the input (e.g., a particular pattern in the input) . A distinguishing feature of CNNs is that many neurons can share the same filter. This reduces the memory footprint because a single bias and a single vector of weights may be used across all receptive fields that share that filter, as opposed to each receptive field having its own bias and vector weighting. However, in some embodiments, all or some of the receptive fields may have their own bias and vector weighting, respectively.
As shown in Fig. 3, an input 310 (e.g., I/Q components of symbols detected at the receiver) may be processed by the CNN 300, which are trained based on training data sets for de-mapping (e.g., a set of collected data for gNB-UE communications in a  practical telecommunications network) , and finally the demodulated soft/hard bits may be derived as the output 345.
Fig. 4 is a diagram illustrating comparison between performance of a legacy de-mapper and an ML assisted de-mapper according to an embodiment of the present disclosure. As clearly shown in Fig. 4, a Prototype of Concept (PoC) implementation of ML assisted de-mapper has proven that:
- ML assisted de-mapper (410) outperforms the legacy commercial De-mapper (420) at high SINR area (cell center) due to less DMRS (i.e., less overhead) , and it has observed 8%peak throughput increase in lab test;
- ML assisted de-mapper (410) outperforms the legacy commercial De-mapper (420) at middle or low SINR area (cell middle or cell edge) due to better control of phase noise, and it has observed around 5dB gain in lab test.
Since ML assisted de-mapper has proven its high capability to suppress phase noise, it has attracted a lot of interest to study how to deploy it in commercial product. Different with lab tests or simulations which normally target for one specific scenario, a real product needs to consider various kinds of channel environments, different UE speed and SINR. Hence ML based de-mappers should be well-generalized to handle all these scenarios.
Unfortunately, generalization issue is not easy to handle, here 3 cases are provided to show the problem.
Case 1: Neural_network_case_1
An ML based de-mapper, whose neural network is specifically optimized (trained) for Clustered Delay Line -E (CDL-E) channel (LOS, Line-of-Sight channel) and 120 km/h speed, may have a total number of parameters: 21, 801, a number of trainable parameters: 21, 725, and a number of non-trainable parameters: 76.
Complexity of case 1 neural network (counted on floating point operation per second (FLOPS) :
- Convolutional layer: 5MFlops
- Batch Normalization layer: 8KFlops
- Max Pooling: 54KFlops
- Fully connected layer: 1kFlops
- In total: 5.06M Flops
Case 2: Neural_network_case_2
An ML based de-mapper, whose neural network is specifically optimized (trained) for CDL-B channel (typical NLOS, non-line-of-sight channel) and pedestrian UE speed at 3 km/h, is a neural network with good performance. It could be trained with similar complexity with Case 1 (but different neural network weights) .
The problem is, in Case 1 scenario, Case-2-optimized neural network performs about 2 dB worse than Case-1-optimized neural network. This 2 dB gap means Case-2-optimized neural network is not well-generalized for all scenarios.
Case 3: Neural_network_case_1&2
An ML based de-mapper, whose neural network is one neural network to be optimized for both Case 1 and Case 2, tries to get similar performance with Case-1-optimized neural network in Case 1, and Case-2-optimized neural network in Case 2 separately.
This neural network is designed with higher complexity, i.e., more channels, more hidden layer and neurons, and its computation complexity is 27.6M Flops (compared with Case-1-optimized neural network 5.06M Flops)
Below is experiment result, 3 cases running on one X86 server with python code:
Case 1 and Case 2: 2.77 ms ± 84.5 μs per loop (mean ± standard deviation of 7 runs, 100 loops each) ;
- Case 3: 7.96 ms ± 71.2 μs per loop (mean ± standard deviation of 7 runs, 100 loops each) .
It indicates that if a more powerful neural network is used to fit diverse scenarios, more complexity will be introduced. Sometime this complexity increase is not tiny, but dramatically. Therefore, there is a problem of how to solve complexity increase issue introduced by generalization.
There are two main approaches to handle generalization issue:
- Design a group of ′tiny′ ML models: Each model could outperform others in limited scenarios. Once UE works in a specific scenario, gNB will select a proper ML model from candidate list. Good thing is that each ML model complexity is limited and well optimized, while the challenge is that it is hard to correctly auto select the best ML model.
- Introducing a super complex ML model to have good performance in all scenarios. Good thing is that no need to select ML model, while the challenge is that too high complexity means a high burden to implementation cost.
Some embodiments of the present disclosure propose a method for automatically selecting a proper ML model from a candidate list, if a group of ′tiny′ ML models is designed. In other words, considering the complexity restrictions, some embodiments of the present disclosure propose a generalized ML model selection, i.e., ε-greedy solution to automatically select the proper ML model.
A general solution according to some embodiments of the present disclosure may comprise steps as follows.
Step 1: With a certain period, based on internal measurement, one optimal ML model candidate sub-set {S} may be selected. In some embodiments, the measurement (s) could for example be at least one of:
- Pathloss;
- Inter-cell interference or SINR;
- UE speed; and
- Channel delay profile (LOS or NLOS) .
In some embodiments, the term ′period′ means that a receiver thinks, in this period, scenario is coherent, e.g., for 3 second.
Step 2: For each model selected into the candidate sub-set, it will be assigned an initial (performance) value V, indicating how good performance this ML model is expected. (higher V, better performance) .
Step 3: For each receiving occasion, the receiver will select a model:
- randomly from candidate sub-set {S} with probability ε; or
- or greedily (choosing the model with the highest value V) with probability 1-ε.
Step 4: If a model is selected, the receiver will use this model to process the data. Its final performance, for example, de-mapper result, could be correctly decoded or not will impact model′s (performance) value V. For example, if a model′s output could be successfully decoded, its value V will increase Δ, or if failed, its value V will decrease Δ′. In some embodiments, Δ may be different from or same as Δ′.
Step 5: if a model′s value V is too low, i.e., always perform worse than others, it could be removed from candidate sub-set {S} .
Step 6: ε determines exploration possibility and it normally has higher value for a new candidate sub-set {S} , while ε continuously decrease when models in candidate sub-set {S} has been tried for several times.
The advantage of the ε-greedy (compared to a completely random) is that it will spend more and more time exploring the interesting parts of the environment, as the performance value (V) estimates get better and better, while still spending some time visiting unknown performance of other models. It is quite common to set an initial high value for ε (e.g., 50 %) while (e.g., 5%) for models in candidate sub-set {S} has been tried for several times.
In this way, instead of introducing a super complex ML model, some embodiments of the present disclosure allow a candidate sub-set {S} , in which each model is simple but only work for limited scenarios.
To correctly select the proper model from candidate sub-set {S} , 3 key features are proposed:
- Physical layer Measurement-based model candidate sub-set {S} creation;
- ε-greedy selection from sub-set; and
- Final performance-based model evaluation.
Advantages of some embodiments of the present disclosure may comprise (but not limited to) avoiding any super complex ML model and auto-selecting proper ML model from candidate list, which make receiver more robust.
Two components in the system, timer based periodical candidate sub-set creation and ML model selection for each reception of data, will be described in detail with reference to Fig. 5 and Fig. 6, respectively.
Fig. 5 is a flow chart illustrating an exemplary method 500 for creating a set of candidate models according to an embodiment of the present disclosure. In the embodiment shown in Fig. 5, the system needs a timer triggered function, to periodically create optimal ML candidate sub-set {S} . This method 500 may be designed to reduce the number of candidate ML models to explore in the following ε-greedy selection. In some embodiments, it is optional but recommended to keep it especially when there is a lot of candidate ML models for selection. However, the present disclosure is not limited thereto. In some other embodiments, this method 500 may be omitted. In other words, all the candidate models are comprised in the candidate set. Further, although the method 500 is shown as being triggered by a timer, the present  disclosure is not limited thereto. In other embodiments, the method 500 may be triggered by an event, for example, upon detection of the availability of a peer for communication.
As shown in Fig. 5, the method 500 will be triggered periodically. For example, the period may be 100 Transmission Time Intervals (TTIs) , 1000 TTIs, or any other value depending on the specific design. In this period, the channel environment or UE speed etc. could be regarded as a constant, and therefore only one candidate sub-set {S} is needed.
This selection of the candidate set tries to identify, for example, that the current UE is locating at what kind of channel, for example LOS (line-of-sight channel) or NLOS (non-line-of-sight channel) , whether UE speed is stationary, pedestrian, medium speed, or high speed. This kind of coarse channel identification may remove some ML models easily from the candidate sub-set {S} (for example, if an ML model is trained to handle high speed, then this model will not be used for stationary UEs) .
For example, measurements for identifying large-scale low variant channel characteristics could include at least one of, (but not limited to) :
- Path loss (for example, indicating whether a UE is at cell edge or cell center) ;
- SINR (for example, indicating whether UE is in high signal quality area or low signal quality area) ;
- UL received doppler frequency shift or channel coherence (for example, indicating UE speed) ; and
- Channel delay profile, i.e., delay spread (LOS or NLOS) .
All above mentioned measurements can be obtained from legacy solution (e.g., channel estimation) and will not explained in detail.
As shown in Fig. 5, the method 500 may begin with step S510 where a 100 TTI timer is expired. At step S520, ML candidates may be selected for an optimal (or candidate) subset. As mentioned above, the ML candidates may be selected at least based on the measurements and/or other factors. At step S530, an initial (performance) value may be assigned for the newly added models in the optimal subset. However, the present disclosure is not limited thereto. In some other embodiments, for each model in the subset, no matter whether it is newly added or previously presented, a same initial value will be assigned. Once the subset and performance values are determined, the  timer may be restarted at step S540, and the method 500 may end at step S550 until the next expiration of the timer.
Fig. 6 is a flow chart illustrating an exemplary method for selecting a model from a set (or subset) of candidate models for de-mapping according to an embodiment of the present disclosure.
The method 600 may begin with step S610 where a new transmission is received. For each receiving occasion, the receiver may select, at step S620, S630, S640, and S650, a model:
- randomly from models in candidate sub-set {S} with probability ε, for example, shown as steps S620, S630, and S650; or
- greedily (choosing the model with the highest value V) with probability 1-ε, for example, shown as steps S620, S630, and S640.
As shown in Fig. 6, one exemplary method for implementing probability ε is to generate a pure random value P at step S620, equally distributed between  value  0 and 1. Each time, the receiver may generate this random value P at step S620 and compare it with ε at S630, and
- if P< ε (whose possibility is ε) , the receiver may go to random mode. It will randomly select a model from candidate optimal ML sub-set at step S650.
- or (whose possibility is 1-ε) , the receiver will go to greedy mode. It will select the model who has the best performance in previous reception (i.e., highest value V) at step S640.
Once a model is selected, the receiver may use it to process the data at step S660. Its final performance (like decoding is correct or not) may impact its value V as shown by step S670. For example, if a model′s output could be successfully decoded, its value V will increase Δ, or if failed, its value V will decrease Δ′. In some embodiments, Δmay be different from or same as Δ′.
In some embodiments, if a model′s (performance) value V is too low, i.e., always perform poorly, it could be removed from the candidate subset, which may avoid unnecessary model exploration waste.
With the embodiments described above, a super complex ML model may be avoided and a proper ML model may be automatically selected from the candidate list, which can make a receiver more robust.
Fig. 7 is a flow chart of an exemplary method 700 at a communication device for model selection according to an embodiment of the present disclosure. The method 700 may be performed at a user equipment (e.g., the UE 100 or the gNB 105) . The method 700 may comprise steps S710 and S720. However, the present disclosure is not limited thereto. In some other embodiments, the method 700 may comprise more steps, less steps, different steps, or any combination thereof. Further the steps of the method 700 may be performed in a different order than that described herein. Further, in some embodiments, a step in the method 700 may be split into multiple sub-steps and performed by different entities, and/or multiple steps in the method 700 may be combined into a single step.
The method 700 may begin at step S710 where one of multiple modes to be used for model selection may be determined.
At step S720, a model may be selected, in the determined mode, from a set of candidate models to be used by a feature.
In some embodiments, the multiple modes may comprise at least one of: a random mode in which a model is to be randomly selected; and a greedy mode in which a model with the highest performance value is to be selected. In some embodiments, the feature may comprise at least one of: soft bit computation. In some embodiments, the method 700 may further comprise: performing the feature by at least using the selected model. In some embodiments, the set of candidate models may comprise one or more trained ML models. In some embodiments, at least one of the step of determining the mode and the step of selecting the model may be performed in response to each invocation of the feature.
In some embodiments, before the step of selecting the model, the method 700 may further comprise: determining the set of candidate models at least based on one or more measurements related to the feature. In some embodiments, when the feature comprises the soft bit computation, the one or more measurements may comprise, for a UE associated with the soft bit computation, at least one of: a pathloss for the UE; an ICI or an SINR for the UE; a speed of the UE; and a channel delay profile for the UE. In some embodiments, the set of candidate models may be determined periodically, and once it is determined, the determined set of candidate models may be used during a corresponding period of time.
In some embodiments, the set of candidate models may be determined in response to a trigger event, and once it is determined, the determined set of candidate models may be used until another trigger event is detected. In some embodiments, the step of determining the mode may comprise at least one of: determining, with a first probability, the random mode to be used for model selection; and determining, with a second probability, the greedy mode to be used for model selection. In some embodiments, a sum of the first probability and the second probability may be 1. In some embodiments, the first probability may have a value that is continuously decreased in response to an increasing number of invocations of the feature since the last determination of the set of candidate models. In some embodiments, the step of determining the mode may comprise at least one of: generating a random number between 0 and 1; determining whether the random number is less than a probability threshold or not; and determining the random mode to be used for model selection in response to determining that the random number is less than the probability threshold and/or determining the greedy mode to be used for model selection in response to determining that the random number is greater than or equal to the probability threshold.
In some embodiments, an initial performance value may be assigned to each of candidate models in the set once the set is determined. In some embodiments, the method 700 may further comprise at least one of: increasing the performance value of the determined model by a first value in response to a first outcome resulted from performing the feature by using the determined model; and decreasing the performance value of the determined model by a second value in response to a second outcome resulted from performing the feature by using the determined model. In some embodiments, the first value is different from the second value. In some embodiments, the method 700 may further comprise: removing a candidate model from the set of candidate models in response to determining that the candidate model has a lower performance value than a threshold and/or than all other candidate models for a period of time and/or for specific times.
Fig. 8 schematically shows an embodiment of an arrangement 800 which may be used in a user equipment (e.g., the UE 100) , a network node (e.g., the gNB 105) , or a communication device according to an embodiment of the present disclosure. Comprised in the arrangement 800 are a processing unit 806, e.g., with a Digital Signal  Processor (DSP) or a Central Processing Unit (CPU) . The processing unit 806 may be a single unit or a plurality of units to perform different actions of procedures described herein. The arrangement 800 may also comprise an input unit 802 for receiving signals from other entities, and an output unit 804 for providing signal (s) to other entities. The input unit 802 and the output unit 804 may be arranged as an integrated entity or as separate entities.
Furthermore, the arrangement 800 may comprise at least one computer program product 808 in the form of a non-volatile or volatile memory, e.g., an Electrically Erasable Programmable Read-Only Memory (EEPROM) , a flash memory and/or a hard drive. The computer program product 808 comprises a computer program 810, which comprises code/computer readable instructions, which when executed by the processing unit 806 in the arrangement 800 causes the arrangement 800 and/or the UE/network node/communication device in which it is comprised to perform the actions, e.g., of the procedure described earlier in conjunction with Fig. 5 to Fig. 7 or any other variant.
The computer program 810 may be configured as a computer program code structured in  computer program modules  810A and 810B. Hence, in an exemplifying embodiment when the arrangement 800 is used in a UE, a network node, or a communication device, the code in the computer program of the arrangement 800 includes: a module 810A configured to determine one of multiple modes to be used for model selection; and a module 810B configured to select, in the determined mode, a model from a set of candidate models to be used by a feature.
The computer program modules could essentially perform the actions of the flow illustrated in Fig. 4 to Fig. 6, to emulate the UE, the network node, and/or the communication device. In other words, when the different computer program modules are executed in the processing unit 806, they may correspond to different modules in the UE, the network node, and/or the communication device.
Although the code means in the embodiments disclosed above in conjunction with Fig. 8 are implemented as computer program modules which when executed in the processing unit causes the arrangement to perform the actions described above in conjunction with the figures mentioned above, at least one of the code means may in alternative embodiments be implemented at least partly as hardware circuits.
The processor may be a single CPU (Central processing unit) , but could also comprise two or more processing units. For example, the processor may include general  purpose microprocessors; instruction set processors and/or related chips sets and/or special purpose microprocessors such as Application Specific Integrated Circuit (ASICs) . The processor may also comprise board memory for caching purposes. The computer program may be carried by a computer program product connected to the processor. The computer program product may comprise a computer readable medium on which the computer program is stored. For example, the computer program product may be a flash memory, a Random-access memory (RAM) , a Read-Only Memory (ROM) , or an EEPROM, and the computer program modules described above could in alternative embodiments be distributed on different computer program products in the form of memories within the UE, the network node, and/or the communication device.
Correspondingly to the method 700 as described above, an exemplary communication device is provided. Fig. 9 is a block diagram of a communication device 900 according to an embodiment of the present disclosure. The communication device 900 may be, e.g., the UE 100 or the gNB 105 in some embodiments.
The communication device 900 may be configured to perform the method 700 as described above in connection with Fig. 7. As shown in Fig. 9, the communication device 900 may comprise a determining module 910 configured to determine one of multiple modes to be used for model selection; and a selecting module 920 configured to select, in the determined mode, a model from a set of candidate models to be used by a feature.
The above modules 910 and/or 920 may be implemented as a pure hardware solution or as a combination of software and hardware, e.g., by one or more of: a processor or a micro-processor and adequate software and memory for storing of the software, a Programmable Logic Device (PLD) or other electronic component (s) or processing circuitry configured to perform the actions described above, and illustrated, e.g., in Fig. 7. Further, the communication device 900 may comprise one or more further modules, each of which may perform any of the steps of the method 700 described with reference to Fig. 7.
With reference to Fig. 10, in accordance with an embodiment, a communication system includes a telecommunication network 3210, such as a 3GPP-type cellular network, which comprises an access network 3211, such as a radio access network, and a core network 3214. The access network 3211 comprises a plurality of  base stations  3212a, 3212b, 3212c, such as NBs, eNBs, gNBs or other types of wireless access points,  each defining a  corresponding coverage area  3213a, 3213b, 3213c. Each  base station  3212a, 3212b, 3212c is connectable to the core network 3214 over a wired or wireless connection 3215. A first UE 3291 located in coverage area 3213c is configured to wirelessly connect to, or be paged by, the corresponding base station 3212c. A second UE 3292 in coverage area 3213a is wirelessly connectable to the corresponding base station 3212a. While a plurality of  UEs  3291, 3292 are illustrated in this example, the disclosed embodiments are equally applicable to a situation where a sole UE is in the coverage area or where a sole UE is connecting to the corresponding base station 3212.
The telecommunication network 3210 is itself connected to a host computer 3230, which may be embodied in the hardware and/or software of a standalone server, a cloud-implemented server, a distributed server or as processing resources in a server farm. The host computer 3230 may be under the ownership or control of a service provider, or may be operated by the service provider or on behalf of the service provider. The  connections  3221, 3222 between the telecommunication network 3210 and the host computer 3230 may extend directly from the core network 3214 to the host computer 3230 or may go via an optional intermediate network 3220. The intermediate network 3220 may be one of, or a combination of more than one of, a public, private or hosted network; the intermediate network 3220, if any, may be a backbone network or the Internet; in particular, the intermediate network 3220 may comprise two or more sub-networks (not shown) .
The communication system of Fig. 10 as a whole enables connectivity between one of the connected  UEs  3291, 3292 and the host computer 3230. The connectivity may be described as an over-the-top (OTT) connection 3250. The host computer 3230 and the connected  UEs  3291, 3292 are configured to communicate data and/or signaling via the OTT connection 3250, using the access network 3211, the core network 3214, any intermediate network 3220 and possible further infrastructure (not shown) as intermediaries. The OTT connection 3250 may be transparent in the sense that the participating communication devices through which the OTT connection 3250 passes are unaware of routing of uplink and downlink communications. For example, a base station 3212 may not or need not be informed about the past routing of an incoming downlink communication with data originating from a host computer 3230 to be forwarded (e.g., handed over) to a connected UE 3291. Similarly, the base station  3212 need not be aware of the future routing of an outgoing uplink communication originating from the UE 3291 towards the host computer 3230.
Example implementations, in accordance with an embodiment, of the UE, base station and host computer discussed in the preceding paragraphs will now be described with reference to Fig. 11. In a communication system 3300, a host computer 3310 comprises hardware 3315 including a communication interface 3316 configured to set up and maintain a wired or wireless connection with an interface of a different communication device of the communication system 3300. The host computer 3310 further comprises processing circuitry 3318, which may have storage and/or processing capabilities. In particular, the processing circuitry 3318 may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions. The host computer 3310 further comprises software 3311, which is stored in or accessible by the host computer 3310 and executable by the processing circuitry 3318. The software 3311 includes a host application 3312. The host application 3312 may be operable to provide a service to a remote user, such as a UE 3330 connecting via an OTT connection 3350 terminating at the UE 3330 and the host computer 3310. In providing the service to the remote user, the host application 3312 may provide user data which is transmitted using the OTT connection 3350.
The communication system 3300 further includes a base station 3320 provided in a telecommunication system and comprising hardware 3325 enabling it to communicate with the host computer 3310 and with the UE 3330. The hardware 3325 may include a communication interface 3326 for setting up and maintaining a wired or wireless connection with an interface of a different communication device of the communication system 3300, as well as a radio interface 3327 for setting up and maintaining at least a wireless connection 3370 with a UE 3330 located in a coverage area (not shown in Fig. 11) served by the base station 3320. The communication interface 3326 may be configured to facilitate a connection 3360 to the host computer 3310. The connection 3360 may be direct or it may pass through a core network (not shown in Fig. 11) of the telecommunication system and/or through one or more intermediate networks outside the telecommunication system. In the embodiment shown, the hardware 3325 of the base station 3320 further includes processing circuitry 3328, which may comprise one or more programmable processors, application-specific integrated circuits, field  programmable gate arrays or combinations of these (not shown) adapted to execute instructions. The base station 3320 further has software 3321 stored internally or accessible via an external connection.
The communication system 3300 further includes the UE 3330 already referred to. Its hardware 3335 may include a radio interface 3337 configured to set up and maintain a wireless connection 3370 with a base station serving a coverage area in which the UE 3330 is currently located. The hardware 3335 of the UE 3330 further includes processing circuitry 3338, which may comprise one or more programmable processors, application-specific integrated circuits, field programmable gate arrays or combinations of these (not shown) adapted to execute instructions. The UE 3330 further comprises software 3331, which is stored in or accessible by the UE 3330 and executable by the processing circuitry 3338. The software 3331 includes a client application 3332. The client application 3332 may be operable to provide a service to a human or non-human user via the UE 3330, with the support of the host computer 3310. In the host computer 3310, an executing host application 3312 may communicate with the executing client application 3332 via the OTT connection 3350 terminating at the UE 3330 and the host computer 3310. In providing the service to the user, the client application 3332 may receive request data from the host application 3312 and provide user data in response to the request data. The OTT connection 3350 may transfer both the request data and the user data. The client application 3332 may interact with the user to generate the user data that it provides.
It is noted that the host computer 3310, base station 3320 and UE 3330 illustrated in Fig. 11 may be identical to the host computer 3230, one of the  base stations  3212a, 3212b, 3212c and one of the  UEs  3291, 3292 of Fig. 10, respectively. This is to say, the inner workings of these entities may be as shown in Fig. 11 and independently, the surrounding network topology may be that of Fig. 10.
In Fig. 11, the OTT connection 3350 has been drawn abstractly to illustrate the communication between the host computer 3310 and the use equipment 3330 via the base station 3320, without explicit reference to any intermediary devices and the precise routing of messages via these devices. Network infrastructure may determine the routing, which it may be configured to hide from the UE 3330 or from the service provider operating the host computer 3310, or both. While the OTT connection 3350 is active, the network infrastructure may further take decisions by which it dynamically  changes the routing (e.g., on the basis of load balancing consideration or reconfiguration of the network) .
The wireless connection 3370 between the UE 3330 and the base station 3320 is in accordance with the teachings of the embodiments described throughout this disclosure One or more of the various embodiments improve the performance of OTT services provided to the UE 3330 using the OTT connection 3350, in which the wireless connection 3370 forms the last segment. More precisely, the teachings of these embodiments may improve the latency and power consumption and thereby provide benefits such as reduced user waiting time, better responsiveness, extended battery lifetime.
A measurement procedure may be provided for the purpose of monitoring data rate, latency, and other factors on which the one or more embodiments improve. There may further be an optional network functionality for reconfiguring the OTT connection 3350 between the host computer 3310 and UE 3330, in response to variations in the measurement results. The measurement procedure and/or the network functionality for reconfiguring the OTT connection 3350 may be implemented in the software 3311 of the host computer 3310 or in the software 3331 of the UE 3330, or both. In embodiments, sensors (not shown) may be deployed in or in association with communication devices through which the OTT connection 3350 passes; the sensors may participate in the measurement procedure by supplying values of the monitored quantities exemplified above, or supplying values of other physical quantities from which  software  3311, 3331 may compute or estimate the monitored quantities. The reconfiguring of the OTT connection 3350 may include message format, retransmission settings, preferred routing etc.; the reconfiguring need not affect the base station 3320, and it may be unknown or imperceptible to the base station 3320. Such procedures and functionalities may be known and practiced in the art. In certain embodiments, measurements may involve proprietary UE signaling facilitating the host computer′s 3310 measurements of throughput, propagation times, latency, and the like. The measurements may be implemented in that the  software  3311, 3331 causes messages to be transmitted, in particular empty or ′dummy′ messages, using the OTT connection 3350 while it monitors propagation times, errors etc.
Fig. 12 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment. The communication system includes a  host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 12 will be included in this section. In a first step 3410 of the method, the host computer provides user data. In an optional substep 3411 of the first step 3410, the host computer provides the user data by executing a host application. In a second step 3420, the host computer initiates a transmission carrying the user data to the UE. In an optional third step 3430, the base station transmits to the UE the user data which was carried in the transmission that the host computer initiated, in accordance with the teachings of the embodiments described throughout this disclosure. In an optional fourth step 3440, the UE executes a client application associated with the host application executed by the host computer.
Fig. 13 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment. The communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 13 will be included in this section. In a first step 35110 of the method, the host computer provides user data. In an optional substep (not shown) the host computer provides the user data by executing a host application. In a second step 3520, the host computer initiates a transmission carrying the user data to the UE. The transmission may pass via the base station, in accordance with the teachings of the embodiments described throughout this disclosure. In an optional third step 3530, the UE receives the user data carried in the transmission.
Fig. 14 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment. The communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 14 will be included in this section. In an optional first step 3610 of the method, the UE receives input data provided by the host computer. Additionally or alternatively, in an optional second step 3620, the UE provides user data. In an optional substep 3621 of the second step 3620, the UE provides the user data by executing a client application. In a further optional substep 3611 of the first step 3610, the UE executes a client application which provides the user data in reaction to the received input data provided by the host computer. In providing the user data, the executed client application may  further consider user input received from the user. Regardless of the specific manner in which the user data was provided, the UE initiates, in an optional third substep 3630, transmission of the user data to the host computer. In a fourth step 3640 of the method, the host computer receives the user data transmitted from the UE, in accordance with the teachings of the embodiments described throughout this disclosure.
Fig. 15 is a flowchart illustrating a method implemented in a communication system, in accordance with one embodiment. The communication system includes a host computer, a base station and a UE which may be those described with reference to Fig. 10 and Fig. 11. For simplicity of the present disclosure, only drawing references to Fig. 15 will be included in this section. In an optional first step 3710 of the method, in accordance with the teachings of the embodiments described throughout this disclosure, the base station receives user data from the UE. In an optional second step 3720, the base station initiates transmission of the received user data to the host computer. In a third step 3730, the host computer receives the user data carried in the transmission initiated by the base station.
The present disclosure is described above with reference to the embodiments thereof. However, those embodiments are provided just for illustrative purpose, rather than limiting the present disclosure. The scope of the disclosure is defined by the attached claims as well as equivalents thereof. Those skilled in the art can make various alternations and modifications without departing from the scope of the disclosure, which all fall into the scope of the disclosure.

Claims (21)

  1. A method (700) at a communication device (100, 105, 800, 900) in a telecommunications network (10) for model selection, the method (700) comprising:
    determining (S620, S630, S710) one of multiple modes to be used for model selection; and
    selecting (S640, S650, S720) , in the determined mode, a model from a set of candidate models to be used by a feature.
  2. The method (700) of claim 1, wherein the multiple modes comprise at least one of:
    - a random mode in which a model is to be randomly selected; and
    - a greedy mode in which a model with the highest performance value is to be selected.
  3. The method (700) of claim 1 or 2, wherein the feature comprises at least one of:
    - soft bit computation.
  4. The method (700) of any of claims 1 to 3, further comprising:
    performing the feature by at least using the selected model.
  5. The method (700) of any of claims 1 to 4, wherein the set of candidate models comprise one or more trained Machine Learning (ML) models.
  6. The method (700) of any of claims 1 to 5, wherein at least one of the step of determining (S620, S630, S710) the mode and the step of selecting (S640, S650, S720) the model is performed in response to each invocation of the feature (S610) .
  7. The method (700) of any of claims 1 to 6, wherein before the step of selecting (S640, S650, S720) the model, the method (700) further comprises:
    determining (S520, S710) the set of candidate models at least based on one or more measurements related to the feature.
  8. The method (700) of claim 7, wherein when the feature comprises the soft bit computation, the one or more measurements comprise, for a User Equipment (UE) (100) associated with the soft bit computation, at least one of:
    - a pathloss for the UE (100) ;
    - an Inter-Cell Interference (ICI) or a Signal to Interference plus Noise Ratio (SINR) for the UE (100) ;
    - a speed of the UE (100) ; and
    - a channel delay profile for the UE (100) .
  9. The method (700) of claim 7 or 8, wherein the set of candidate models is determined periodically, and once it is determined, the determined set of candidate models is used during a corresponding period of time.
  10. The method (700) of claim 7 or 8, wherein the set of candidate models is determined in response to a trigger event, and once it is determined, the determined set of candidate models is used until another trigger event is detected.
  11. The method (700) of any of claims 3 to 10, wherein the step of determining (S620, S630, S710) the mode comprises at least one of:
    determining (S620, S630) , with a first probability, the random mode to be used for model selection; and
    determining (S620, S630) , with a second probability, the greedy mode to be used for model selection.
  12. The method (700) of claim 11, wherein a sum of the first probability and the second probability is 1.
  13. The method (700) of claim 11 or 12, wherein the first probability has a value that is continuously decreased in response to an increasing number of invocations of the feature since the last determination of the set of candidate models.
  14. The method (700) of any of claims 3 to 13, wherein the step of determining (S620, S630, S710) the mode comprises at least one of:
    generating (S620) a random number between 0 and 1;
    determining (S630) whether the random number is less than a probability threshold or not; and
    determining (S630: No) the random mode to be used for model selection in response to determining that the random number is less than the probability threshold and/or determining (S630: Yes) the greedy mode to be used for model selection in response to determining that the random number is greater than or equal to the probability threshold.
  15. The method (700) of any of claims 1 to 14, wherein an initial performance value is assigned (S530) to each of candidate models in the set once the set is determined.
  16. The method (700) of claim 15, further comprising at least one of:
    increasing the performance value of the determined model by a first value in response to a first outcome resulted from performing the feature by using the determined model; and
    decreasing the performance value of the determined model by a second value in response to a second outcome resulted from performing the feature by using the determined model.
  17. The method (700) of claim 16, wherein the first value is different from the second value.
  18. The method (700) of any of claims 1 to 17, further comprising:
    removing a candidate model from the set of candidate models in response to determining that the candidate model has a lower performance value than a threshold and/or than all other candidate models for a period of time and/or for specific times.
  19. A communication device (100, 105, 800, 900) , comprising:
    a processor (806) ;
    a memory (808) storing instructions which, when executed by the processor (806) , cause the processor (806) to perform the method (700) of any of claims 1 to 18.
  20. A computer program (810) comprising instructions which, when executed by at least one processor (806) , cause the at least one processor (806) to carry out the method (700) of any of claims 1 to 18.
  21. A carrier (808) containing the computer program (810) of claim 20, wherein the carrier is one of an electronic signal, optical signal, radio signal, or computer readable storage medium.
PCT/CN2022/088430 2022-04-22 2022-04-22 Automatic model selection WO2023201695A1 (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/088430 WO2023201695A1 (en) 2022-04-22 2022-04-22 Automatic model selection

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/CN2022/088430 WO2023201695A1 (en) 2022-04-22 2022-04-22 Automatic model selection

Publications (1)

Publication Number Publication Date
WO2023201695A1 true WO2023201695A1 (en) 2023-10-26

Family

ID=88418865

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/088430 WO2023201695A1 (en) 2022-04-22 2022-04-22 Automatic model selection

Country Status (1)

Country Link
WO (1) WO2023201695A1 (en)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021079206A1 (en) * 2019-10-24 2021-04-29 Telefonaktiebolaget Lm Ericsson (Publ) System, method and associated computer readable media for facilitating machine learning engine selection in a network environment
US20210328630A1 (en) * 2020-04-16 2021-10-21 Qualcomm Incorporated Machine learning model selection in beamformed communications
US20210326701A1 (en) * 2020-04-16 2021-10-21 Qualcomm Incorporated Architecture for machine learning (ml) assisted communications networks
US20220036123A1 (en) * 2021-10-20 2022-02-03 Intel Corporation Machine learning model scaling system with energy efficient network data transfer for power aware hardware
US20220116764A1 (en) * 2020-10-09 2022-04-14 Qualcomm Incorporated User equipment (ue) capability report for machine learning applications

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2021079206A1 (en) * 2019-10-24 2021-04-29 Telefonaktiebolaget Lm Ericsson (Publ) System, method and associated computer readable media for facilitating machine learning engine selection in a network environment
US20210328630A1 (en) * 2020-04-16 2021-10-21 Qualcomm Incorporated Machine learning model selection in beamformed communications
US20210326701A1 (en) * 2020-04-16 2021-10-21 Qualcomm Incorporated Architecture for machine learning (ml) assisted communications networks
US20220116764A1 (en) * 2020-10-09 2022-04-14 Qualcomm Incorporated User equipment (ue) capability report for machine learning applications
US20220036123A1 (en) * 2021-10-20 2022-02-03 Intel Corporation Machine learning model scaling system with energy efficient network data transfer for power aware hardware

Similar Documents

Publication Publication Date Title
US11296817B2 (en) Method and network node, for handling link adaption of a channel
EP2115893B1 (en) Mitigation of co-channel interference in a wireless communication system
US20220386151A1 (en) Techniques for machine learning based peak to average power ratio reduction
US9602242B2 (en) Coherent reception with noisy channel state information
US20240187133A1 (en) Decoder success predictor signaling for adjusting mirs scheduling policy
Chen et al. AMC with a BP-ANN scheme for 5G enhanced mobile broadband
WO2023201695A1 (en) Automatic model selection
WO2023004541A1 (en) Methods and apparatuses for adaptation of communication parameter
US9516607B2 (en) Method for determining transmission power
US20240284201A1 (en) Ml model category grouping configuration
WO2023015430A1 (en) The combined ml structure parameters configuration
US20230068633A1 (en) Ue indication of null tone placement for demodulation
EP4377836A1 (en) Online optimization for joint computation and communication in edge learning
Kim et al. Optimum MCS for high-throughput long-range ambient backscatter communication networks
US11438038B1 (en) Neural network based nonlinear MU-MIMO precoding
US20230084883A1 (en) Group-common reference signal for over-the-air aggregation in federated learning
US20240080165A1 (en) Ack coalescing performance through dynamic stream selection
US12047788B2 (en) Transfer/federated learning approaches to mitigate blockage in millimeter wave systems
US20240314798A1 (en) Dci-based indication to trigger the combined ml model
WO2022036642A1 (en) Method and apparatus for beamforming
WO2023060417A1 (en) Ue clustering in fl model update reporting
US20240063961A1 (en) Null tones adaptation using reinforcement learning
EP4340307A1 (en) A wireless device and method for transmitting a wake-up signal
WO2023236175A1 (en) Techniques for backscatter and backscatter-aided advanced communication
WO2024092743A1 (en) Precoded reference signal for model monitoring for ml-based csi feedback

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22937931

Country of ref document: EP

Kind code of ref document: A1