Variable configuration method and variable configuration device thereof
By receiving accident prediction notifications in the variable configuration device, using continuous time structural equation model and Bayesian optimization technology, PDU session strategy is adjusted in advance, and the problem that the radio access network cannot respond to service quality requirements in a timely manner is solved, and the user experience and Network efficiency.
Patent Information
- Application Number
- CN202310654828.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-05-18
- Filing Date
- 2023-06-05
- Publication Date
- 2024-11-22
AI Technical Summary
The existing third-generation partner program specifications for the radio access network cannot respond to service quality requirements in a timely manner, resulting in a decrease in user experience and interruption.
By receiving accident prediction notifications in the variable configuration device, deciding and outputting policy time causal variables in advance, in order to modify the policy of the PDU session before service quality violations, and avoiding quality degradation. The device includes processing circuits and storage circuits, and uses continuous time structural equation modeling and Bayesian optimization technologies to predict and adjust QoS parameters to ensure service quality.
It effectively avoids the occurrence of service quality violations, improves user experience and network efficiency, and ensures the stability of service quality through advance prediction and adjustment of strategies.
Smart Images

Figure CN119012227A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to a variable configuration method and a variable configuration device thereof, and in particular to a variable configuration method and a variable configuration device thereof that can improve efficiency and user experience. Background Art
[0002] The current 3rd Generation Partnership Project (3GPP) specification is that when the Radio Access Network (RAN) finds that the service quality requirements cannot be met, the RAN will (via the Access Mobility Management Function and the Session Management Function) send a Protocol Data Unit (PDU) session modification notification to the Policy Charging Function (PCF) and let the PCF decide to modify the PDU session policy. At this time, it is often too late and interruptions will occur, reducing the user experience. Summary of the invention
[0003] Therefore, the present invention mainly provides a variable configuration method and a variable configuration device thereof to improve efficiency and user experience.
[0004] An embodiment of the present invention discloses a variable configuration method for a variable configuration device, comprising receiving an accident prediction notification from a network at a first time point; determining at least one variable at a second time point in response to the accident prediction notification; and outputting the at least one variable at a third time point, wherein the accident prediction notification is used to instruct the variable configuration device to determine the at least one variable in response to a service quality violation prediction at a fourth time point, and the third time point is earlier than the fourth time point.
[0005] An embodiment of the present invention discloses a variable configuration device, comprising a storage circuit for storing an instruction, wherein the instruction comprises receiving an accident prediction notification from a network at a first time point; determining at least one variable at a second time point in response to the accident prediction notification; and outputting the at least one variable at a third time point, wherein the accident prediction notification is used to instruct the variable configuration device to determine the at least one variable in response to a service quality violation prediction at a fourth time point, and the third time point is earlier than the fourth time point; and a processing circuit coupled to the storage circuit for executing the instruction stored in the storage circuit. BRIEF DESCRIPTION OF THE DRAWINGS
[0006] Figure 1 A schematic diagram of a variable configuration device according to an embodiment of the present invention;
[0007] Figure 2 A schematic diagram of a variable configuration method according to a first embodiment of the present invention;
[0008] Figure 3 This is a sequence diagram of a variable configuration method according to an embodiment of the present invention;
[0009] Figure 4 A schematic diagram of the relationship between a loss function at a certain time point and a vector formed by a strategy time causal variable according to an embodiment of the present invention;
[0010] Figure 5 A schematic diagram of Bayesian optimization for a one-dimensional problem according to an embodiment of the present invention;
[0011] Figure 6 A schematic diagram of a cause-effect diagram of a part of an embodiment of the present invention;
[0012] Figure 7 It is a schematic diagram of basic data and a local cause-effect diagram of an embodiment of the present invention;
[0013] Figure 8 A schematic diagram of a causal variable and a sub-data according to an embodiment of the present invention;
[0014] Fig. 9 A schematic diagram of Bayesian optimization, causal Bayesian optimization, or dynamic causal Bayesian optimization;
[0015] Fig.10 FIG. 1 is a schematic diagram of a communication system according to an embodiment of the present invention.
[0016] Explanation of symbols
[0017] 10: Variable configuration device
[0018] 100: Processing circuit
[0019] 110: Storage circuit
[0020] 114: Program Code
[0021] 20,30: Variable configuration method
[0022] 3AMF: Access Mobility Management Function
[0023] 3PCF: Policy Charging Function
[0024] 3RAN: Radio Access Network
[0025] 3SMF: Session Management Function
[0026] CG1-CG2: Cause and Effect Diagram
[0027] cv (i-1) ,cv i ,cv (j-1) ,cv j ,cv y ,X1~X3,Y1~Y3,Z1~Z3: variables
[0028] f (i-1) ,fi ,f (j-1) ,f j :Observation function
[0029] gD: Basic data
[0030] L() : Loss Function
[0031] O1~O q: Independent Variable
[0032] p(s,v),p ⊥ (s,v),p ~ (s,v): prior
[0033] p(w x |s,v),p(cv y |s):Causal Mechanism
[0034] P1~P5,P × 5: Point
[0035] QP1~QP n ,Session-AMBR: Strategy Time Causal Variable
[0036] q(s,v|w x ):Inference Model
[0037] S200~S206, S301~S305: Steps
[0038] sc: semantic factors
[0039] t0~t5: time point
[0040] TCV1~TCV n :vector
[0041] v: Factors of change
[0042] w (i-1) ,w i ,w (j-1) ,w j ,w x :Subdata DETAILED DESCRIPTION
[0043] Figure 11 is a schematic diagram of a variable configuration device 10 according to an embodiment of the present invention. The variable configuration device 10 may be a network element (e.g., a Policy Charging Function (PCF)), but is not limited thereto. The variable configuration device 10 may include a processing circuit 100 and a storage circuit 110. The storage circuit 110 is used to store a program code 114, and the processing unit 100 may read and execute the program code 114 through the storage circuit 110.
[0044] Figure 2 FIG. 1 is a schematic diagram of a variable configuration method 20 according to an embodiment of the present invention. The variable configuration method 20 may be compiled into a program code 114 and used in the variable configuration device 10 , and may include the following steps:
[0045] Step S200: Start.
[0046] Step S202: Receive an accident prediction notification from a network (eg, a Radio Access Network (RAN)) at a first time point.
[0047] Step S204: In response to the accident prediction notification, determine at least one variable at a second time point later than the first time point.
[0048] Step S206: Outputting the determined at least one variable (e.g., a policy temporal causal variable) at a third time point later than the second time point, wherein the accident prediction notification is used to instruct, before the third time point, to determine the at least one variable in response to a Quality of Service (QoS) violation prediction at a fourth time point later than the third time point.
[0049] Step S206: End.
[0050] That is, instead of waiting until the network finds that the QoS requirement cannot be met, the network will first predict and send an accident prediction notification to the variable configuration device 10 in step S202 to instruct the modification of the Protocol Data Unit (PDU) session. In step S204, the variable configuration device 10 can determine how to set the policy (i.e., the at least one variable) of the PDU session. In step S206, the variable configuration device 10 transmits the determined policy so that the network can modify the QoS parameters or QoS characteristics of the QoS flows of the relevant PDU session accordingly to avoid the QoS violation predicted by the network. For example, when a video call cannot meet the requirements of the QoS parameters or QoS characteristics, directly reducing the speed (reducing the bandwidth of the video or audio) may affect the user experience. Therefore, in the present invention, the network can make a prediction that the video call cannot meet the QoS requirement before the video call fails to meet the QoS requirement, and the variable configuration device 10 can correspondingly pre-determine how to modify the policy of the PDU session, thereby avoiding the accident that the video call cannot meet the requirements of the QoS parameters or QoS characteristics.
[0051] Figure 3 It is a sequence diagram of a variable configuration method 30 according to an embodiment of the present invention.
[0052] The variable configuration method 30 may include the following steps:
[0053] Step S301: A radio access network 3RAN predicts at a time point t0 that a QoS violation may occur at a time point t5 (eg, a fourth time point) (eg, a video call cannot meet the QoS requirement).
[0054] In one embodiment, the radio access network 3RAN may use a recurrent neural network (RNN), a long short-term memory (LSTM), a transformer or other similar discrete time models in step S301 to perform prediction analysis of time causal variables. However, the discrete time model can only process a certain time step (e.g., time step t, t+Δt, t+2Δt) instead of a certain time instant (e.g., time point t, ) of time, so it can only predict what is most likely to happen in the future but cannot accurately predict when it will happen. In addition, since the time scale of the discrete time model (e.g., the length of the time period △t) should match the time scale of the dynamic random process being modeled, it is difficult to select an appropriate time scale for the discrete time model.
[0055] In another embodiment, the radio access network 3RAN may use a continuous time structural equation model (CTSEM) to perform prediction analysis of time causal variables in step S301 to predict the time point (e.g., time point t5) when the QoS requirement cannot be met, rather than the time period of the discrete time model. Since CTSEM may involve stochastic differential equations, which are continuous time operations, and it extracts the causal relationship in time to make accurate predictions, it can better understand and decide on the modification of the PDU session.
[0056] In CTSEM, the vector η of temporal causal variables h (t) is a function of time and satisfies or dη h (t) = (Aη h (t)+ξ h +Bz h +M∑ u x h,u δ(tu))dt+GdW h (t) (Equation 2). The matrix A can be used to characterize the vector η using the self-effects on the diagonal and the cross-effects on the off-diagonal. h (t) time relationship. Matrix I is the identity matrix. Random vector ξ h The vector η can be determined h (t) and can satisfy ξ h ~N(κ,φ ξ ), the vector κ can represent the continuous time intercept, and the matrix φ ξ It can be the covariance (e.g., the covariance across QoS flows). The matrix B can represent the (fixed) time-independent prediction vector z h Vector η h (t), and the number of rows may not be equal to the number of columns. The time-independent prediction vector z h The prediction vector x is usually different for different QoS flows, but can be constant for different QoS flows within the evaluation time range. h,u It can be observed at time u and only affects the vector η at time u h (t), x h,u The pulse pair vector η composed of δ(tu) h The influence of (t) can be expressed as a matrix M. The vector W h (s) can be an independent random walk in continuous time (for example, a Wiener process), dW h (s) can be a random error term. The lower triangular matrix G represents the vector ηh (t) changes. Satisfy Q = GG T The matrix Q represents the variance-covariance matrix of the diffusion process in continuous time. The matrices A, B, M, G, and the vector ξ can be obtained by fitting equation 1 or 2 to the data received, sent, or stored by the radio access network 3RAN. h 、z h 、x h,u , W h (s), time point u, or the value of the initial time point t0.
[0057] The continuous-time structural equation model can provide full information maximum likelihood estimation for continuous-time multivariate autoregressive models for time series or panel data. In one embodiment, the dynamic random process modeled is one process (i.e., time series and a single device) rather than multiple processes (i.e., panel data and multiple devices). Accordingly, the vector η h (t) may be equal to 1, and there is only one vector η1(t). In another embodiment, the vector η h (t) can correspond to different QoS flows respectively, so the n QoS flows will correspond to vectors η1(t)~η n (t), wherein h can be 1 to n, and h and n are positive integers.
[0058] In one embodiment, the time-causal variable may be, for example, the lowest flow bit rate (LFBR) of the uplink (UL) or downlink (UL) of a guaranteed bit rate (GBR) QoS flow, the highest flow bit rate (HFBR) of the UL or DL of a GBR QoS flow, the packet loss rate (PLR) of the UL or DL of a GBR QoS flow, the session-aggregate bit rate (Session-ABR) (at the PDU session level) of a non-guaranteed bit rate (non-GBR) QoS flow, the user equipment-aggregate bit rate (UE-ABR) of a non-GBR QoS flow, the packet delay time (PDT), the packet error rate (PER), or the burst volume (BurstVolume) (only for delay-critical GBR resource types). For the hth QoS flow of a PDU session, the vector of temporal causal variables η h (t) can be a multidimensional vector consisting of LFBR, HFBR, PLR, Session-ABR, UE-ABR, PDT, and PER, and can be expressed as (LFBR h ,HFBR h ,PLR h ,Session-ABR,UE-ABR,PDT h ,PER h ).
[0059] In one embodiment, the radio access network 3RAN may compare the relationship between the time causal variable at a certain time point and its corresponding preset threshold value (e.g., the policy time causal variable) in step S301 (e.g., the time causal variable is greater than or less than the policy time causal variable) to monitor QoS violation. For example, when the LFBR at time point t5 is less than the guaranteed flow bit rate (Guaranteed Flow Bit Rate, GFBR), or when the HFBR at time point t5 is greater than the maximum flow bit rate (Maximum Flow Bit Rate, MFBR), or when the PLR at time point t5 is greater than the maximum packet loss rate (Maximum-PLR), or when the Session-ABR at time point t5 is greater than the session-aggregate maximum bit rate (Session-AMBR), or when the UE-ABR at time point t5 is greater than the user equipment-aggregate maximum bit rate (UE-Aggregate Maximum Bit Rate, UE-AMBR), or when the PDT at time point t5 is greater than the packet delay budget (Packet Delay Budget, PDB), or when the PER at time point t5 is greater than the maximum packet error rate (Maximum-PER), the radio access network 3RAN determines that a QoS violation occurs at time point t5.
[0060] In one embodiment, the radio access network 3RAN (eg Fig.10 A central unit (CU) or a resource management unit of a central unit performs CTSEM prediction in step S301 to predict whether a QoS violation may occur or at which point in time a QoS violation may occur.
[0061] Step S302a: The radio access network 3RAN transmits a notification about the PDU session modification (eg, an accident prediction notification) to a session management function 3SMF (Session Management Function, SMF) via an access and mobility management function (AMF) 3AMF.
[0062] Step S302b: The session management function 3SMF notifies a policy charging function 3PCF at a time point t1 (eg, the first time point) to determine the policy used to modify the PDU session.
[0063] The radio access network 3RAN may use the transmission accident prediction notification in steps S302a and S302b to inform the policy charging function 3PCF to determine the policy for modifying the PDU session before the predicted time point t5, and the policy is at least for the QoS violation prediction at the upcoming time point t5. In other words, the radio access network 3RAN may indicate that the policy for modifying the PDU session should be re-determined before the predicted time point t5 to ensure that the QoS requirement is met at least at the time point t5.
[0064] In one embodiment, the policy (e.g., at least one variable or at least one policy time-causal variable of at least one variable) may include, for example, a QoS profile, an alternative profile, or a session-AMBR. The QoS profile may include a 5th generation mobile networks (5G) QoS identifier (5QI), an allocation and retention priority (ARP), a UL or DL GFBR of a GBR QoS flow, a UL or DL MFBR of a GBR QoS flow, a UL or DL Maximum-PLR of a GBR QoS flow, a Delay Critical Resource Type of a GBR QoS flow, a Notification Control of a GBR QoS flow, a Reflective QoS Attribute (RQA) of a non-GBR QoS flow, a Session-AMBR (at a PDU session level) of a non-GBR QoS flow, a UE-AMBR (at a user level) of a non-GBR QoS flow, or other QoS parameters. The QoS characteristics of the 5QI derived from the QoS parameters may include resource type (e.g., GBR, delay-critical GBR, or non-GBR), priority level, PDB, Maximum-PER, averaging window (only for GBR and delay-critical GBR resource types), or maximum data burst size (only for delay-critical GBR resource types). A PDU session contains the 1st QoS flow to the nth QoS flow, so the policy can be expressed as (TCV1, TCV2, ..., TCV n ). For the hth QoS flow of a PDU session, the vector TCV hIt can contain m policy time causal variables (such as Session-AMBR), where m is a positive integer. The vector TCV of the policy time causal variables of the hth QoS flow h It can be expressed as (QP h )、(QP h ,Session-AMBR), or (GFBR h ,MFBR h ,Maximum-PLR h ,UE-AMBR,PDB h ,Maximum-PER h ,Session-AMBR),QP h Indicates the QoS profile of the hth QoS flow.
[0065] Step S303: The policy charging function 3PCF determines the policy at a time point t2 (eg, the second time point). In one embodiment, the policy charging function 3PCF may consider the inputs it receives and use an algorithm to determine the policy.
[0066] For example, in one embodiment, the policy charging function 3PCF may receive input from a network data analytics function (NWDAF) (e.g., various network data analyses), an input from a unified data repository (UDR) function (e.g., information about subscriptions (e.g., spending limits), Nudr_DM_Notify, or application-specific information), an input from an application function (e.g., Application / Service Info), an input from an application server, or an input from a video streaming server in step S303 or before step S303, to determine the policy based on these inputs. In one embodiment, the spending limit may be, for example, all-you-can-eat or a certain data volume limit (e.g., a 1GB data volume limit).
[0067] In one embodiment, the algorithm that the policy charging function 3PCF can use in step S303 is, for example, Bayesian optimization (BO), causal Bayesian optimization (CBO), or dynamic causal Bayesian optimization (DCBO). The policy charging function 3PCF can determine or calculate the optimal value of the policy (for example, at least one policy time causal variable as an independent variable) based on BO, CBO, or DCBO (or can apply BO, CBO, or DCBO to at least one data (for example, data extracted from a signal sent, received, or stored by the policy charging function 3PCF)) to achieve the minimum value of the loss function at any time point. That is, the policy charging function 3PCF can determine the optimal policy time causal variable (for example, vector TCV) at a certain time point in step S303. h ). The loss function L() (which can be used as the objective function) can be related to the difference between the time causal variable at a certain point in time and the strategy time causal variable, for example, it can satisfy
[0068] The following uses BO as an example. BO is a black-box optimization algorithm that can be used to solve the extreme value problem of functions with unknown expressions. For example, L(TCV1,…,TCV n )=uef(TCV1,…,TCV n ), where uef() represents the unknown functional relationship and L() represents the loss function. For example, Figure 4 is a vector TCV1-TCV consisting of the loss function L() and the strategy time causal variable at a certain time point in the embodiment of the present invention. n A schematic diagram of the relationship between . Figure 4 It is just for illustration to make it easier to understand the loss function L() and the vectors TCV1~TCV n There is a correlation between them, however, the loss function L() and the vector TCV1~TCV formed by the strategy time causal variable n The functional relationship between uef() and uef() is unknown in most cases. Figure 4 A point on the horizontal axis can correspond to vectors TCV1~TCV n A point in the multidimensional space of strategic time causal variables.
[0069] Since the expression of the function relationship uef() is unknown, BO can roughly fit the function relationship uef() through partial / limited sampling points, and can use the information of previous sampling points to determine the next sampling point to find the extreme point. For example, Figure 5This is a schematic diagram of BO for a one-dimensional problem according to an embodiment of the present invention, wherein the thick solid line represents the estimated loss function L(), the solid black dots P1 to P5 represent the searched sampling points, the area surrounded by the two dotted lines represents the fluctuation range of the loss function L() corresponding to each point (centered on the mean and proportional to the standard deviation), and the thin solid line represents the extraction function. In BO, an initial set of candidate solutions can be generated first (for example, the loss function L() corresponding to the solid black dot P1, vectors TCV1 to TCV n ) of the strategy time causal variables); then, according to the sampling points of the initial candidate solution set (for example, the mean and variance of the real loss function can be estimated according to the function values of the already searched sampling points (for example, the loss function L() corresponding to the solid black point P1)), search for the next (possibly extreme) sampling point (for example, the solid black point P2); repeatedly search for the next (possibly extreme) sampling point (for example, the solid black points P3~P5) until the iteration terminates, and add all the searched sampling points (for example, the solid black points P1~P5) to the candidate solution set; finally, find the extreme point from the sampling points of the candidate solution set (searched from the beginning to now) as the extreme value of the loss function L(), thereby providing the best strategy time causal variables Session-AMBR, QP1~QP with the minimum loss function n (For example, it contains a vector TCV h The strategic time causal variable GFBR h MFBR h 、Maximum-PLR h 、PDB h 、Maximum-PER h ) as the solution to the problem (e.g., the loss function corresponding to the solid black point P5, the vectors TCV1~TCV n The strategy time causal variable) is used to determine the strategy at a certain point in time in step S303.
[0070] According to the results of Gaussian process regression, an extraction function can be constructed to solve the extreme value of the extraction function to determine the next sampling point of the loss function. Figure 5 The extraction function represented by the thin solid line can be established based on the mean and variance. The extraction function estimates whether a point (such as one of the solid black points P1 to P5) is the extreme point of the loss function, which reflects the degree to which this point is worth searching. The relative extreme point of the extraction function can correspond to the next sampling point of the loss function: for example Figure 5 The point P represented by the rectangular box × 5 is the maximum point of the extraction function, and (according to the corresponding vectors TCV1~TCV n ) may correspond to the next sampling point of the loss function (ie, the solid black dot P5). The extraction function may be, for example, knowledge gradient, entropy search, or predicted entropy search.
[0071] The algorithm of the present invention can use Gaussian process regression to predict the probability distribution of the function value of any point based on the function value of the loss function L() at a set of sampling points. Gaussian process regression can be extended to observations of independent normally distributed noise with known variance. The variance may be unknown, so it can be assumed that the noise has a common variance and the noise includes the variance as a hyperparameter. The present invention uses the posterior mean of a Gaussian process including noise (which is a drift value rather than the noise of the signal-to-interference-plus-noise ratio). In one embodiment, environmental factors such as temperature and humidity or measurement errors (such as measurement errors of LFBR) may cause the loss function to correspond to drift values of certain strategic time causal variables. In other words, the selected sampling points (such as the solid black dot P5) may not select / correspond to the extreme value of the expected functional relationship uef(), but may select / correspond to the relatively optimized extreme value close to the expected functional relationship uef() (i.e., the extreme value of the functional relationship uef() after being disturbed by noise).
[0072] In one embodiment, the algorithm of the present invention may need to consider many independent variables (or called independent variables) (not just the strategic time causal variables QP1-QP n , Session-AMBR, independent variables O1~O q ). As the spatial dimension increases, the performance of BO may decrease exponentially. Therefore, in step S303, the algorithm of the present invention can be extended to CBO. In other words, the loss function L() is related to the vectors TCV1-TCV n In the case of other independent variables, the present invention can find the loss function L() and the vectors TCV1-TCV n (and / or other independent variables), while CBO only optimizes the causal variables that are directly related to the loss function L().
[0073] For example, Figure 6 FIG. 1 is a schematic diagram of a partial causal graph CG1 of an embodiment of the present invention, wherein (a) and (b) respectively illustrate two possibilities of the causal graph CG1. q (may not exist), vector TCV1~TCV n The strategic time causal variables and the loss function L() constitute the causal graph CG1 used by the causal model, where q is a positive integer. Figure 6 As shown, the causal graph CG1 gives the causal relationship between the causal variables, and since the independent variables O1~O q And vector TCV1~TCV nThe strategic time causal variables directly point to / directly affect the loss function L(), so the causal intrinsic dimensionality discovered by CBO may be n×m+q (or n×m), and n×m+q (or n×m) independent variables are input into the loss function L() of CBO for optimization, so that the values of the n×m+q (or n×m) causal variables can be calculated to minimize the loss function L(). In other words, the causal graph CG1 can significantly improve the ability to infer the best decision, thereby reducing the optimization cost and avoiding suboptimal solutions. In addition, the causal intrinsic dimension of CBO is composed of the independent variables O1~O2 that are directly the cause of the loss function L(). q 、Vector TCV1~TCV n The number of strategic time causal variables (e.g., n×m or n×m+q) is used to determine the strategy, rather than as a vector TCV1~TCV n and independent variables O1~O q The number of causal variables (not shown) of the cause can therefore improve the ability to infer the best decision strategy.
[0074] In one embodiment, the optimized causal model can be selected based on maximum a posterior (MAP) and point estimate to obtain the loss function, vectors TCV1-TCV n The causal graph of the strategic time causal variables and other independent variables is composed of the causal graph of the causal model. According to this, the causal variables (such as which causal variables are selected, the number of causal variables, which attributes a causal variable has, or the number of attributes of a causal variable) and the causal structure (such as the connection between causal variables or their attributes) of the causal graph of the causal model are determined simultaneously / in parallel / together. Therefore, the causal variables and the causal structure can influence and constrain each other, and can avoid the problems caused by determining the causal variables first and then determining the causal structure (for example, the causal model can distinguish Figure 7 (a) and (b) to avoid the problem of being unable to distinguish them).
[0075] For example, Figure 7 Schematic diagram of a basic data gD and a local causal graph CG2 according to an embodiment of the present invention, wherein (a) and (b) respectively illustrate two possibilities of the basic data gD and the causal graph CG2, and the causal graph CG2 can be used as the causal graph CG1. Figure 7 , the causal structure of the causal graph CG2 can show the causal variable cv (i-1) 、cv i 、cv (j-1) 、cv j The relationship between the observation function f (i-1) 、f i 、f (j-1)、f j Then the sub-data w of the basic data gD can be (i-1) 、w i 、w (j-1) 、w j Mapped to the causal variable cv (i-1) 、cv i 、cv (j-1) 、cv j The causal variable cv (i-1) 、cv i 、cv (j-1) 、cv j Sub-data w of basic data gD (i-1) 、w i 、w (j-1) 、w j The relationship between, where i and j are positive integers. The mapping is based on the corresponding sub-data (such as w (i-1) 、w i 、w (j-1) 、w j )( Figure 7 The box area) rather than the entire underlying data gD. For example, the causal variable cv (i-1) Corresponding to the strategic time causal variable of vector TCV1, the sub-data w (i-1) With the causal variable cv (i-1) In other words, using a Bayesian network, the Bayesian probability mechanism can combine causal variables (such as cv (i-1) 、cv i 、cv (j-1) 、cv j ), causal structures, or observation functions (e.g., f (i-1) 、f i 、f (j-1) 、f j ) and obtain relevant joint inferences to explain the basic data gD, thereby generating a causal graph CG2.
[0076] In one embodiment, the sub-data w of the basic data gD may be i To observe the function f i and the posterior probability P(f i ,C|w i ) is maximized, so that the sub-data w of the basic data gD i Get / select / judge the causal structure C and its causal variables cv i (i.e. extracting vectors TCV1~TCV from the basic data gD n The strategic time causal variables and loss function L() or independent variables O1~O q ).
[0077] In one embodiment, the posterior probability P(f i ,C|w i ,Int) can satisfy P(f i ,C|w i ,Int)∝P(f i ,C)P(w i |f i ,C,Int), where f i can represent the observation function, C can represent the causal structure, and w i may represent part of the basic data gD, and Int may represent intervention (which may be omitted). In one embodiment, the posterior probability P(f i ,C|w i ) is proportional to P(f i ,C)P(w i |f i ,C) or where w i,t It can represent a causal variable cv corresponding to a time point t i Sub-data, s t-1 may represent the state (of all causal variables) at a time point t-1, T may represent the current time point, and γ may be a real number such as but not limited to 0 or 0.5. In one embodiment, P(w i |f i ,C) can be In one embodiment, P(w i,t |s t-1 ,C,f i ) can be or where s i,t Can represent the causal variable cv i At time t, Ncv represents the state (e.g., its value) of all causal variables (e.g., cv (i-1) 、cv i 、cv (j-1) 、cv j ), Ncv is a positive integer, Can represent the causal variable cv in the basic data gD i Status i Compatible data w i In one embodiment, the amount of data can be minimized. To select the causal variable cv i , so that the more commonly used data in the basic data gD (such as w i ) into smaller parts than less frequently used data.
[0078] In one embodiment, the basic data gD may be obtained from the space of all observable samples, and thus may also be referred to as observed data. In one embodiment, the basic data gD may include or be related to data extracted from signals sent or received by the policy charging function 3PCF at all times and in all ways (e.g., input from a network data analysis function).
[0079] like Figure 7 As shown, each causal variable (e.g. cv i ) will correspond to an observation function (for example, the observation function f i In one embodiment, a causal semantic generation model can be used to obtain an observation function (e.g., f i ), thereby predicting low-dimensional state attributes (such as causal variables cv i If the causal variable (for example, the vector TCV1~TCV n The strategic time causal variables and independent variables O1~O q or causal variable cv (i-1) 、cv i 、cv (j-1) 、cv j ) is manually defined (e.g., defined by domain experts (not automatically but one by one) or using a program that describes rules by domain experts to automatically define causal variables), each causal variable (e.g., cv i ) has a dedicated causal semantics to generate observation functions so that the causal variables are based on the corresponding sub-data (e.g. w i )( Figure 7 ), which means the sub-data ( Figure 7 The sub-data w corresponding to the frame area (i-1) 、w i 、w (j-1) 、w j ) is determined based on the definition of a specific causal variable by domain experts. In addition, the causal semantic generation model can avoid judging the variable factors as causal variables (such as cv i ) and can correctly judge semantic factors as causal variables (such as cv i ). In one embodiment, the variation factor and the semantic factor may belong to the observation data. In one embodiment, the causal semantic generation model is mainly based on the causal invariance criterion and involves variational Bayes.
[0080] In one embodiment, the observation function f i Can satisfy i,t =f i (w i,t In another embodiment, the observation function f i This can be achieved using a multivariate Gaussian distribution, for example or Or, observe that the function f i Can be related to or Where z is the causal variable cv in the basic data gD i No contributing data, mean μ v Fixed to the zero vector, Σ can be parameterized by Cholesky decomposition and can satisfy Σ = LL T The matrix L may be a lower triangular matrix with positive diagonal elements and may be parameterized, for example, to satisfy matrix L zz Can be a smaller lower triangular matrix, the matrix Can be any matrix. L zz It can be parameterized using the sum of the positive diagonal elements (identified by the exponential map) and the lower triangular matrix (which has no positive diagonal elements).
[0081] In one embodiment, the causal variable (e.g., cv i ) and subdata (such as w i ) is unknown, but the causal variables can be predicted from the sub-data using the causal semantic generation model. For example, Figure 8 is the causal variable cv in the first embodiment of the present invention y and a sub-data w x Schematic diagram of the causal semantic generation model, where (a), (b), (c), and (d) respectively depict possible structures of the causal semantic generation model, sc can represent semantic factors, v can represent change factors, and the solid arrow can represent the causal mechanism p(w x |sc,v) and p(cv y |sc), the dotted arrows represent the inference model q(sc,v|w x ).exist Figure 8 In (a), the solid undirected line segment between the semantic factor sc and the variation factor v can represent the domain-specific prior p(sc,v). Figure 8 (a) the undirected line segment between the semantic factor sc and the variable factor v, Figure 8 (b) introduces an independent prior p ⊥ (sc,v): = p(sc)p(v) to reflect the intervention to improve out-of-distribution generalization performance, Figure 8 (c) According to the causal invariance principle, the priori p between the semantic factor sc and the variable factor v is introduced ~ (sc,v) to reflect the intervention to utilize unsupervised data. In one embodiment, the causal semantic generation model p:= <p(sc,v),p(wx |sc,v),p(cv y |sc)> is fitted to the sub-data and variational inference and evidence lower bounds can be computed, and then, after applying the reparameterization technique, Monte Carlo estimation of the expectation can be used.
[0082] Vector TCV1~TCV n The strategy time causal variable can be a function of time (i.e., the value is different at different time points). CBO converts the output causal variable (such as the loss function L()) and the input causal variable (such as the vector TCV1~TCV n The strategic time causal variables and independent variables O1~O q ) is regarded as an unchanging independent variable, without considering whether the output causal variables and the input causal variables have time evolution and can change over time, breaking the time-dependent structure between the causal variables. Although ignoring time can significantly simplify the problem, ignoring time will hinder the identification of the best intervention at each time point, and (in non-steady-state scenarios) lead to suboptimal solutions instead of providing the current optimal solution at any moment in time. Therefore, the present invention can be extended to DCBO in step S303, which is beneficial to the scenario where the causal effect in the causal graph changes over time.
[0083] For example, Fig. 9 is a schematic diagram of BO, CBO or DCBO, X1~X3 represent a causal variable at three different time points, Y1~Y3 represent another causal variable at three different time points, Z1~Z3 represent another causal variable at three different time points, but the present invention is not limited to this and can be extended to multiple different time points or causal variables. DCBO combines BO and CBO to explain the causal relationship between causal variables and the causal relationship can evolve over time. For example, in DCBO, if Fig. 9As shown, the causal variable Y1 at the first time point is a function of the causal variable Z1 at the first time point, so the extreme value of the causal variable Y1 (which can represent the loss function L()) can be searched only for the causal variable Z1 directly related to the causal variable Y1, and the causal intrinsic dimension is 1. The causal variable Y2 at the second time point is a function of the causal variable Z2 at the second time point and the causal variable Y1 at the first time point, so the extreme value of the causal variable Y2 (which can represent the loss function L()) can be searched only for the causal variables Z2 and Y1 directly related to the causal variable Y2, and the causal intrinsic dimension is 2. The causal variable Y3 at the third time point is a function of the causal variable Z3 at the third time point and the causal variable Y2 at the second time point, so the extreme value of the causal variable Y3 (which can represent the loss function L()) can be searched only for the causal variables Z3 and Y2 directly related to the causal variable Y3, and the causal intrinsic dimension is 2. In other words, the causal variable that serves as the dependent variable at a certain point in time is a function of the causal variable (that serves as the dependent variable or independent variable) at the previous point in time or at the previous few points in time. Therefore, the extreme value of the former (that is, the causal variable that serves as the dependent variable at a certain point in time) can be searched only for the directly related latter (that is, the causal variable that serves as the dependent variable or independent variable at the previous point in time or at the previous few points in time).
[0084] In step S303, the best alternative of the QoS parameters or QoS characteristics (such as GFBR, MFBR, Maximum-PLR, PDB, Maximum-PER) calculated by the algorithm (such as DCBO) can be used for the strategic time causal variable QP h Step S303 can avoid waiting until the network finds that the QoS requirement cannot be met before taking remedial measures, so the QoS parameters or QoS characteristics (such as GFBR, PDB, Maximum-PER) used for the alternative profile can be ignored.
[0085] Step S304a: The policy charging function 3PCF transmits the determined policy to the session management function 3SMF at a time point t3 (eg, the third time point) (eg, transmits Npcf_SMPolicyControl_UpdateNotify request).
[0086] Step S304b: The session management function 3SMF correspondingly transmits a message (eg Namf_Communication_N1N2MessageTransfer) to the access mobility management function 3AMF.
[0087] Step S304c: The access mobility management function 3AMF correspondingly sends a message (such as N2Message or N2Session Request) to the radio access network 3RAN to modify the QoS parameters or QoS characteristics of the QoS flow of the related PDU session.
[0088] Step S305: The radio access network 3RAN determines the strategy (eg, strategy time causal variables QP1-QP n , Session-AMBR) to modify or set the QoS parameters (such as QoS parameters of the QoS profile) or QoS characteristics (such as PDB, Maximum-PER or other QoS characteristics) of the QoS flow of the relevant protocol data unit PDU session, so as to control the flow or speed of the data flow and avoid the predicted QoS violation at time point t5.
[0089] In one embodiment, the storage circuit used to store image data or instructions may be a read-only memory, a flash memory or a random access memory, a hard disk, a non-volatile storage device, or a non-transitory computer readable medium, but not limited thereto. The processing circuit used to execute instructions may be a central processing unit, a microprocessor, or a specific application integrated circuit, but not limited thereto.
[0090] To summarize, the network will first predict and send an accident prediction notification to the variable configuration device before the predicted QoS violation occurs, so as to instruct the variable configuration device to determine the strategy (i.e., how the QoS parameters or QoS characteristics of the QoS flow should be set), so that the network can set or modify the QoS parameters or QoS characteristics of the QoS flow according to the strategy.
[0091] The above descriptions are only preferred embodiments of the present invention. All equivalent changes and modifications made according to the claims of the present invention should fall within the scope of the present invention.
Claims
1. A variable configuration method, used in a variable configuration device, comprising: Receive accident prediction notification from the network at the first point in time; In response to the accident prediction notification, determining at least one variable at a second time point; and The at least one variable is output at a third time point, wherein, The accident prediction notification is used to instruct the variable configuration device to determine the at least one variable in response to a service quality violation prediction at a fourth time point, and the third time point is earlier than the fourth time point.
2. The variable configuration method according to claim 1, wherein: The network predicts that the service quality requirement cannot be met at the fourth time point according to the continuous-time structural equation model and outputs the accident prediction notice.
3. The variable configuration method of claim 1 , wherein the network compares at least one temporal causal variable at the fourth time point with the at least one variable at the fourth time point to perform the service quality violation prediction, and the vector η formed by the at least one temporal causal variable h (t) Satisfaction A represents the qualitative matrix, t0 represents the initial time point, I represents the identity matrix, ξ h represents a random vector, B represents a transformation matrix, z h represents the time-independent prediction vector, M represents the coefficient matrix, x h,u represents the time-related prediction vector, u represents the pulse time point, W h (s) represents the random walk vector, and G represents the lower triangular matrix.
4. The variable configuration method as described in claim 1, wherein at least one time-causal variable comprises a minimum stream bit rate, a maximum stream bit rate, a packet loss rate, a session-aggregate bit rate, a user equipment-aggregate bit rate, a packet delay time, or a packet error rate, wherein the at least one variable comprises a guaranteed stream bit rate, a maximum stream bit rate, a maximum packet loss rate, a session-aggregate maximum bit rate, a user equipment-aggregate maximum bit rate, a packet delay budget, or a maximum packet error rate.
5. The variable configuration method according to claim 1, wherein: The at least one variable is determined at the second time point according to Bayesian optimization, causal Bayesian optimization, or dynamic causal Bayesian optimization.
6. The variable configuration method as claimed in claim 1, wherein the network modifies at least one QoS parameter or at least one QoS characteristic corresponding to the at least one variable at a fifth time point earlier than the fourth time point, and the at least one variable is mapped to the minimum value of the loss function at the fifth time point.
7. The variable configuration method according to claim 1, wherein: The step of determining the at least one variable at the second time point comprises: The causal graph is generated by maximizing the multiple posterior probabilities of dividing the multiple sub-data of the basic data into multiple observation functions and causal structures, one of which is proportional to where s t-1 represents the state of multiple causal variables at time point t-1, T represents the fifth time point, C represents the causal structure, and f i represents the observation function of the i-th causal variable in the multiple observation functions corresponding to the multiple causal variables, w i,t Represents the sub-data corresponding to the i-th causal variable among the multiple sub-data at time point t.
8. The variable configuration method according to claim 1, wherein: The causal structure of the causal graph between the loss function and the plurality of causal variables is determined together with the plurality of causal variables of the causal graph, and the plurality of causal variables including the at least one variable are searched out from the plurality of independent variables.
9. The variable configuration method according to claim 1, wherein: Multiple observation functions are obtained based on the causal semantic generation model.
10. The variable configuration method according to claim 1, wherein: The loss function at the fifth time point is a function of the multiple causal variables at the fifth time point, the multiple causal variables at at least one time point earlier than the fifth time point, or the loss function at the at least one time point.
11. A variable configuration device, comprising: A storage circuit is used to store instructions, which include: Receive accident prediction notification from the network at the first point in time; In response to the accident prediction notification, determining at least one variable at a second time point; and The at least one variable is output at a third time point, wherein, The accident prediction notification is used to instruct the variable configuration device to determine the at least one variable to respond to the service quality violation prediction at a fourth time point, and the third time point is earlier than the fourth time point; as well as The processing circuit is coupled to the storage circuit and is used for executing the instruction stored in the storage circuit.
12. The variable configuration device according to claim 11, wherein: The network predicts that the service quality requirement cannot be met at the fourth time point according to the continuous-time structural equation model and outputs the accident prediction notice.
13. The variable configuration device of claim 11 , wherein the network compares at least one temporal causal variable at the fourth time point with the at least one variable at the fourth time point to perform the service quality violation prediction, the vector η formed by the at least one temporal causal variable h (t) Satisfaction A represents the qualitative matrix, t0 represents the initial time point, I represents the identity matrix, ξ h represents a random vector, B represents a transformation matrix, z h represents the time-independent prediction vector, M represents the coefficient matrix, x h,u represents the time-related prediction vector, u represents the pulse time point, W h (s) represents the random walk vector, and G represents the lower triangular matrix.
14. The variable configuration device of claim 13, wherein the at least one time-causal variable comprises a minimum stream bit rate, a maximum stream bit rate, a packet loss rate, a session-aggregate bit rate, a user equipment-aggregate bit rate, a packet delay time, or a packet error rate, wherein the at least one variable comprises a guaranteed stream bit rate, a maximum stream bit rate, a maximum packet loss rate, a session-aggregate maximum bit rate, a user equipment-aggregate maximum bit rate, a packet delay budget, or a maximum packet error rate.
15. The variable configuration device according to claim 11, wherein: The at least one variable is determined at the second time point according to Bayesian optimization, causal Bayesian optimization, or dynamic causal Bayesian optimization. 16 . The variable configuration device as claimed in claim 11 , wherein the network modifies at least one QoS parameter or at least one QoS characteristic corresponding to the at least one variable at a fifth time point earlier than the fourth time point, and the at least one variable is mapped to a minimum value of a loss function at the fifth time point.
17. The variable configuration device according to claim 11, wherein: The step of determining the at least one variable at the second time point comprises: The causal graph is generated by maximizing the multiple posterior probabilities of dividing the multiple sub-data of the basic data into multiple observation functions and causal structures, one of which is proportional to where s t-1 represents the state of multiple causal variables at a time point t-1, T represents the fifth time point, C represents the causal structure, and f i represents the observation function of the i-th causal variable in the multiple observation functions corresponding to the multiple causal variables, w i,t Represents the sub-data corresponding to the i-th causal variable among the multiple sub-data at a time point t.
18. The variable configuration device according to claim 17, wherein: The causal structure of the causal graph between the loss function and the multiple causal variables is determined together with the multiple causal variables of the causal graph, and the multiple causal variables including the at least one variable are searched out from multiple independent variables.
19. The variable configuration device according to claim 17, wherein: The multiple observation functions are obtained based on a causal semantic generation model.
20. The variable configuration device according to claim 11, wherein: The loss function at the fifth time point is a function of the multiple causal variables at the fifth time point, the multiple causal variables at a sixth time point earlier than the fifth time point, or the loss function at the sixth time point.
