Digital twinborn interaction experience optimization method based on generative artificial intelligence algorithm

Abstract

The invention discloses a digital twinning interaction experience optimization method based on a generative artificial intelligence algorithm. The method aims at solving the problem that traditional intelligent reflecting surface assisted digital twinning interaction lacks uplink and downlink signal optimization at the same time. Most of the users only pay attention to objective indexes in interaction and ignore important subjective experience of the users; the problem that the interaction experience of a user is affected due to the fact that a digital twinning model changes along with the change of an entity in the interaction process is mostly ignored is solved, and digital twinning interaction optimization with uplink and downlink combination and subjective and objective mixing is achieved. Based on a generative artificial intelligence algorithm, a phase offset matrix of an intelligent reflecting surface and a beam forming matrix received and transmitted by a digital twin server are jointly optimized, signal rendering resolution configuration is fed back, resource allocation is calculated, and the objective and subjective experience of a user in interaction is maximized. And generalization of the algorithm is realized to cope with random evolution of the digital twin model.

Description

Technical Field

[0001] The present invention belongs to a method for optimizing digital twin interaction experience, and particularly relates to an optimization algorithm based on generative artificial intelligence. Background Art

[0002] Digital Twin (DT) is regarded as a groundbreaking technology, whose main goal is to create high-fidelity and interactive virtual replicas of physical entities, called digital twin models. These models, as hyper-realistic and vivid virtual avatars of the real world, support real-time monitoring and predictive analysis, enabling people to have a deeper and more forward-looking understanding of the physical world. To fully unleash the potential of DT services, immersive interaction with DT models, which allows users to access DT models to obtain valuable feedback, is extremely important and indispensable. At the same time, DT services are shifting from cloud-based models to ubiquitous edge computing models, which urgently requires providing wireless network connections for mobile users to access DT models anytime and anywhere. Against this backdrop, Reconfigurable Intelligent Surface (RIS), with its powerful capabilities of flexibly guiding signals, mitigating interference, and enhancing coverage, has become a promising solution that can help mobile users achieve seamless connection during DT interaction in a dynamic wireless environment. However, despite the many benefits of RIS-assisted DT interaction, implementing this technology also faces a series of challenges:

[0003] First of all, in RIS-assisted DT interaction, the uplink interaction signal transmission of mobile users, the generation and transmission of the downlink feedback signals of DT models occur simultaneously. This not only shares limited communication and computing resources but also jointly affects the end-to-end objective and subjective performance. This indicates that all resource allocations related to RIS-assisted DT interaction, including the phase shift matrix, high-dimensional receive / transmit beamforming matrix, feedback signal rendering resolution, and computing resource configuration, must be jointly optimized. In addition, optimizing subjective (objective) performance usually requires increasing (decreasing) the feedback signal resolution, which in turn increases (decreases) the communication and computing burden and ultimately weakens the objective (subjective) performance. This inherent trade-off between objective and subjective performance makes it impossible to solve them separately, let alone different mobile users may have different weights for objective / subjective performance. These factors prompt us to design an efficient algorithm aimed at jointly allocating resources in RIS-assisted DT interaction while balancing objective and subjective performance, taking into account the personalized quality of experience requirements of each mobile user.

[0004] Secondly, the state of physical entities in the real world is constantly changing, sometimes beyond expectations, which drives the uncertain evolution of the corresponding DT models. Due to this evolution, the DT scenarios that mobile users participate in during the interaction may continue to change. Driven by this change and its impact on the quality of experience (QoE) of mobile users in different scenarios, the resource allocation for RIS-assisted DT interaction becomes scenario-specific rather than one-size-fits-all. This requires us to re-solve each scenario-specific problem by redesigning appropriate algorithms when the DT scenario changes. A potential way to circumvent this difficulty is to utilize the scenario-specific information in the corresponding DT scenario to guide the optimization algorithm to adapt to the newly changed scenario-specific problem. Generative artificial intelligence (GAI) and prompt-based learning may be very suitable for achieving this goal because it can train the algorithm to perform various tasks by simply modifying the prompts, with high adaptability and versatility. However, different from common GAI applications such as question answering and image generation, which usually adopt text-description prompts, the resource allocation problem for RIS-assisted DT interaction contains complex scenario-specific information that is not in text form. Therefore, we need to further explore the application of "decision trajectories" in prompt design to adapt to the problem under consideration. Such decision trajectories will record a series of past resource allocation decisions for RIS-assisted DT interaction.

[0005] Therefore, it is of great significance and quite challenging to design a resource allocation optimization algorithm for RIS-assisted DT interaction based on GAI, which has generalization to cope with the changing DT scenarios and aims to maximize the subjective and objective experiences of users in the interaction. Summary of the Invention

[0006] Object of the Invention: Aiming at the above RIS-assisted DT interaction, the present invention designs a resource allocation optimization algorithm based on GAI with generalization, aiming to jointly optimize the uplink and downlink resources while taking into account the optimization of subjective and objective performances.

[0007] Technical Solution: A digital twin interaction experience optimization method based on a generative artificial intelligence algorithm, which jointly optimizes the phase shift matrix of the intelligent reflecting surface, the receive and transmit beamforming matrices of the digital twin server, the feedback signal rendering resolution configuration, and the computing resource allocation based on the generative artificial intelligence algorithm, aiming to maximize the subjective and objective experiences of users in the interaction and achieve the generalization of the algorithm to cope with the random evolution of the digital twin model;

[0008] The method includes the following steps:

[0009] (1) Construct an uplink and downlink transmission model for intelligent surface-assisted digital twin interaction to characterize the uplink and downlink channels and the corresponding uplink and downlink transmission rates;

[0010] (2) Establish a user experience quality model, and its mathematical expression is:

[0011]

[0012] where is the preference weight of the user, t represents the time slot, is the mathematical function of the subjective experience model of the mobile user, represents the end-to-end delay of interaction for the objective experience of the mobile user, and represent the weights of the user for subjective and objective experiences respectively;

[0013]

[0014] In the formula, ε max represents the maximum subjective perception quality that each mobile user can achieve, represents the maximum interactive round-trip delay that each mobile user can tolerate;

[0015] (3) Construct an optimization problem for maximizing user experience, and its mathematical representation is:

[0016]

[0017] The constraint conditions are:

[0018]

[0019] where, Θ(i,t) is the phase shift matrix, V(i,t): = {v1(i,t), …, v K (i,t)} is the received beamforming matrix, W(i,t): = {w1(i,t), …, w K (i,t)} is the transmitted beamforming matrix, E k (i,t) is the feedback signal rendering resolution allocation, f k (i,t) represents the computing resource allocation;

[0020] (4) Construct the optimization problem described in step (3) into a Markov decision model, then implicitly represent the information of the task based on the decision trajectory, establish a prompt guidance mechanism based on the decision trajectory for different scenarios to capture new artificial intelligence scenario-specific information, and then use the causal Transformer algorithm to obtain decisions, including constructing a prompt guidance artificial intelligence combined with a zero-forcing optimization algorithm to help deduce high-dimensional decisions to maximize the user experience quality obtained by all users when interacting with any artificial intelligence scenario.

[0021] Furthermore, the modeling method in step (1) specifically includes:

[0022] (11) The direct uplink channel between each mobile user \(k\in K\) and the DT server \(a\) is modeled as:

[0023]

[0024] where \(\rho\) is the path loss coefficient, \(\alpha\) k,a is the path loss exponent, \(d\) k,a (i) is the distance between the mobile user \(k\) and the DT server, is the non-line-of-sight component of this link \(a\), and each element follows

[0025] (12) The non-direct uplink channel between each mobile user \(k\) and the intelligent reflecting surface \(r\) is modeled as:

[0026]

[0027] where \(\alpha\) k,r is the path loss exponent, \(d\) k,r (i) is the distance between the mobile user \(k\) and the intelligent reflecting surface, is the non-line-of-sight component of this link, and each element follows \(G\) k,r is the Rician factor, is the line-of-sight component of this link;

[0028] (13) The channel between the intelligent reflecting surface and the digital twin server is modeled as:

[0029]

[0030] where \(d\) r,a represents the distance between the intelligent reflecting surface and the digital twin server, \(G\) r,a is the Rician factor;

[0031] (14) The signal received by the digital twin server is modeled as:

[0032]

[0033] where is the uplink transmission power of the mobile user \(k\in K\), \(v\) k (i,t) is the beamforming vector, is the noise;

[0034] (15) The uplink transmission rate is: where \(b\) is the bandwidth, is the uplink signal-to-noise ratio;

[0035] (16) The downlink transmission rate is:

[0036] Furthermore, the construction of the user's subjective experience quality model in step (2) includes the following process:

[0037] (21) Establish the subjective experience of the mobile user, expressed as:

[0038]

[0039] where E min is the user's minimum resolution requirement, and E k (i,t) is the resolution assigned to the user;

[0040] (22) The transmission delay of the uplink interaction signal is expressed as:

[0041]

[0042] where is the data volume size of the uplink signal;

[0043] (23) The delay of processing the feedback signal is expressed as:

[0044]

[0045] (24) The downlink transmission delay of the feedback is expressed as:

[0046]

[0047] where is the data volume size of the feedback signal;

[0048] (25) The objective experience of the mobile user is expressed by the end-to-end delay of the interaction:

[0049]

[0050] (26) Combining the subjective and objective experiences, the expression of the experience quality model is obtained:

[0051]

[0052] where is the user's preference weight,

[0053] Furthermore, step (4) constructs a prompt-guided Decision Transformer combined with a zero-forcing-based optimization algorithm, specifically including:

[0054] (41) The problem is reconstructed as a Markov decision, mathematically expressed as where,

[0055] Status:

[0056] Action:

[0057] Reward:

[0058] State transition: Pr(i, s(i, t+1)|s(i, t), a(i, t)) ∈ [0, 1]

[0059] (42) Design hint mechanism: This formula implicitly represents the information of the task through a decision-making trajectory;

[0060] (43) Calculate the embedding layer of the input:

[0061] Adopt hint τ ★ (i), and the recent trajectory decision τ(i, t) of length L before the current time step as the historical trajectory;

[0062] Adopt a modality-specific trainable linear layer for embedding to process the three different token modalities of the return target, state, and decision in the hint;

[0063] Use a trainable linear layer to add position embeddings to the token embeddings within the same decision trajectory tuple. Similarly, apply a trainable linear layer to τ(i, t);

[0064] (44) Use a causal Transformer to obtain decisions:

[0065] The embedded input tokens are input into a causal transformer, which consists of stacked identical decoders; each decoder processes the tokens through a masked multi-head self-attention module to capture the dependencies between the tokens; a feed-forward layer enhances this representation through per-position transformation, while layer normalization stabilizes the training and promotes gradient updates; the causal transformer processes the tokens sequentially in stacked decoders and finally outputs the hidden states;

[0066] Then, input these hidden states into a trainable linear decision prediction layer to generate decisions;

[0067] (45) Design a zero-forcing optimization algorithm:

[0068] Given by the hint-guided Decision Transformer calculation:

[0069]

[0070] The problem of optimizing κ(i, t) = {V(i, t), W(i, t)} is constructed as:

[0071]

[0072] The constraint conditions are:

[0073]

[0074] (46) Obtained according to the zero - forcing algorithm Where

[0075] (47) Based on the obtained V(i, t), the problem is reconstructed as:

[0076]

[0077] Constraint conditions

[0078]

[0079] According to the zero - forcing algorithm,

[0080]

[0081] In the above formula, P is a diagonal matrix, where the k - th diagonal element is the received power of the mobile user at time step i, that is, p DL k(i, t);

[0082] Set the constraint conditions:

[0083]

[0084] Based on the constraint conditions, the optimization problem is further reconstructed as:

[0085]

[0086] Constraint conditions:

[0087]

[0088] For the above expression, use the water - filling algorithm to solve this problem and obtain a closed - form solution:

[0089]

[0090] where the k - th diagonal element is is a normalization factor used to ensure that In addition,

[0091] is the minimum received power limit for mobile users.

[0092] Beneficial effects: The method described in the present invention addresses the inefficient method of having to set a new resource allocation optimization algorithm for each DT scenario in the face of uncertainty-evolving digital twin interactions. It proposes an optimization method for digital twin interaction experiences based on generative artificial intelligence algorithms. This method designs a corresponding prompt based on "decision trajectories" for different DT scenarios to capture new DT scenario-specific information, and extends a traditional GAI algorithm, Decision Transformer, into a prompt-guided Decision Transformer with strong generalization ability. In addition, this method also integrates an optimization algorithm based on zero-forcing (ZF) to help derive high-dimensional decisions (i.e., receive / transmit beamforming matrices) to maximize the QoE obtained by all users when interacting with any DT scenario. At the same time, this method also achieves generalization, enabling each problem in the interaction process to be efficiently solved. Description of the Drawings

[0093] Figure 1 is a diagram of the intelligent reflecting surface-assisted digital twin interaction system in the present invention;

[0094] Figure 2 is the training convergence diagram of the method proposed in the present invention;

[0095] Figure 3 is a diagram comparing the overall system performance under different methods in the present invention. Detailed Embodiments

[0096] To elaborate in detail on the technical solutions disclosed in the present invention, the present invention will be further described below in conjunction with the drawings and specific embodiments.

[0097] First, the key problem to be solved by the method of the present invention is that in RIS (Reconfigurable Intelligent Surface)-assisted DT (Digital Twin) interaction, the uplink interaction signal transmission of mobile users, the downlink feedback signal generation and transmission of the DT model are carried out simultaneously. This not only shares limited communication and computing resources, but also jointly affects the end-to-end objective and subjective performance. This indicates that all resource allocations related to RIS-assisted DT interaction, including the phase shift matrix, high-dimensional receive / transmit beamforming matrix, feedback signal rendering resolution, and computing resource configuration, must be jointly optimized. In addition, the state of physical entities in the real world is constantly changing, sometimes beyond expectations, which drives the uncertain evolution of the corresponding DT model. Due to this evolution, the DT scenarios participated by mobile users during the interaction may continuously change. Driven by this change and its impact on the quality of experience (QoE) of mobile users in different scenarios, the resource allocation of RIS-assisted DT interaction becomes scenario-specific rather than one-size-fits-all. This requires us to re-solve each scenario-specific problem by re-designing appropriate algorithms when the DT scenario changes. To solve these problems, the present invention designs an optimization method for digital twin interaction experience based on generative artificial intelligence algorithms, aiming to jointly optimize uplink and downlink resources, while taking into account the optimization of subjective and objective performance, as well as the uncertain evolution of DT scenarios. The method provides the overall control process as Figure 1 shown.

[0098] An optimization method for digital twin interaction experience based on generative artificial intelligence algorithms, comprising the following steps:

[0099] Step 1: Construct an uplink and downlink transmission model for RIS-assisted DT interaction.

[0100] The method considers the uplink and downlink transmission of RIS-assisted DT interaction, including the modeling of uplink and downlink channels and the corresponding uplink and downlink transmission rates, specifically including the following steps:

[0101] (11) The direct uplink channel between each mobile user k∈K and the DT server can be modeled as:

[0102]

[0103] where ρ is the path loss coefficient, α k,a is the path loss exponent, is the distance between mobile user k and the DT server, is the non-line-of-sight (NLoS) component of this link, and each element follows

[0104] (12) The non-direct uplink channel between each mobile user k∈K and the RIS can be modeled as:

[0105]

[0106] where α k,r is the path loss exponent, is the distance between the mobile user k and the RIS, is the non-line-of-sight (NLoS) component of the link, and each element follows G k,r is the Rician factor. In addition, is the line-of-sight (LoS) component of the link, where is the cosine value of the angle of arrival (AoA) of the signal from the mobile user k to the RIS. In addition, d and λ represent the spacing of a single reflecting surface and the carrier wavelength, respectively.

[0107] (13) The channel between the RIS and the DT server can be modeled as:

[0108]

[0109] where is the distance between the RIS and the DT server, G r,a is the Rician factor.

[0110] (14) The signal received by the DT server can be modeled as:

[0111]

[0112] where is the uplink transmission power of the mobile user k ∈ K, is the beamforming vector, is the noise.

[0113] (15) The uplink transmission rate is: b is the bandwidth, is the uplink signal-to-noise ratio.

[0114] (16) Similarly, the downlink transmission rate is

[0115] Step 2: Quality of Experience (QoE) model analysis;

[0116] (21) The subjective experience of the mobile user is modeled as:

[0117]

[0118] where E min is the minimum resolution requirement of the user, and E k (i, t) is the resolution allocated to the user.

[0119] (22) The transmission delay of the uplink interaction signal is modeled as:

[0120]

[0121] Where is the data volume of the uplink signal.

[0122] (23) The delay of processing the feedback signal is modeled as:

[0123]

[0124] (24) The downlink transmission delay of the feedback is modeled as:

[0125]

[0126] Where is the data volume of the feedback signal.

[0127] (25) The objective experience of the mobile user is represented by the end-to-end delay of the interaction and can be modeled as:

[0128]

[0129] (26) Combining subjective and objective experiences, the quality of experience (QoE) is obtained:

[0130]

[0131] Where is the preference weight of the user, Step 3: Construct an optimization problem for maximizing the user experience;

[0132]

[0133] The constraint conditions are

[0134]

[0135] Among them, Θ(i,t) is the phase shift matrix, V(i,t):={v1(i,t),…,v K (i,t)} is the received beamforming matrix, W(i,t):={w1(i,t),…,w K (i,t)} is the transmit beamforming matrix, E k (i,t) is the feedback signal rendering resolution allocation, f k (i,t) is the computing resource allocation. Constraint 1) represents the control of the phase shift matrix, Constraint 2) represents the range of the feedback signal resolution, Constraint 3) represents the limit of the maximum allocable computing resources, Constraint 4) represents the power limit of the downlink transmission signal, and Constraint 5) represents that the end-to-end delay of the user needs to be within a threshold.

[0136] Step 4: Design a prompt to guide the Decision Transformer to combine with the zero-forcing (ZF)-based optimization algorithm;

[0137] (41) The problem is reconstructed as

[0138] Status:

[0139] Action:

[0140] Reward:

[0141] State transition: Pr(i, s(i, t+1)|s(i, t), a(i, t)) ∈ [0, 1]

[0142] (42) Design a prompt mechanism: Its essence is a decision-making trajectory, which implicitly represents the information of the task.

[0143] (43) Calculate the embedding layer of the input:

[0144] Adopt prompt τ ★ (i), and the recent trajectory decision τ(i, t) of length L before the current time step as the historical trajectory. To process the three different token modalities of the return target (RTG), state, and decision in the prompt, modality-specific trainable linear layers are used for embedding. In addition, a trainable linear layer is also used to add positional embeddings to the token embeddings within the same decision trajectory tuple. Similarly, a trainable linear layer is also applied to τ(i, t).

[0145] (44) Use a causal Transformer to obtain decisions:

[0146] The embedded input tokens are fed into a causal Transformer, which consists of stacked identical decoders. Each decoder processes the tokens through a masked multi-head self-attention module to capture the dependencies between the tokens. A feed-forward layer enhances these representations through per-position transformation, while layer normalization stabilizes the training and promotes gradient updates. The causal Transformer processes the tokens sequentially in stacked decoders and finally outputs the hidden states. Then, these hidden states are fed into a trainable linear decision prediction layer to generate decisions.

[0147] (45) Design the ZF optimization algorithm:

[0148] Given the

[0149]

[0150] The problem of optimizing κ(i,t) = {V(i,t), W(i,t)} can be formulated as

[0151]

[0152] The constraints are

[0153]

[0154] (46) Calculate V(i,t):

[0155] According to the properties of the ZF algorithm, we can obtain where

[0156] (47) Calculate W(i,t):

[0157] Based on the obtained V(i,t), the problem can be reformulated as:

[0158]

[0159] The constraints

[0160]

[0161] According to the ZF algorithm, we can obtain where

[0162] P is a diagonal matrix, where the k-th diagonal element is the received power of the moving user at time step , i.e., p DL k(i,t). And, in ZF, these 2 constraints must be satisfied: and Based on these 2 constraints, the problem can be further reformulated as:

[0163]

[0164] The constraints

[0165]

[0166] Using the water-filling algorithm to solve this problem, we obtain a closed-form solution where the k-th diagonal element is is a normalization factor to ensure that In addition, is the minimum received power limit for mobile users.

[0167] To comprehensively verify the digital twin interaction experience optimization method based on the generative artificial intelligence algorithm proposed in the present invention, the performance is evaluated through the following two metrics: training convergence experiment and total performance experiment.

[0168] Combined with Figure 2 as shown, the training convergence of the digital twin interaction experience optimization method based on the generative artificial intelligence algorithm is examined. From Figure 2 it can be seen that the mean squared error (MSE) of this method steadily decreases and converges to a small value. This indicates that the proposed method has learned a general solution strategy applicable to all historical scenario-specific problems In addition, the performance of the proposed method on unseen problems is further evaluated during the offline training process . The results show that when L MSE decreases and converges, increases and tends to be stable. This indicates that the method described in the present invention has good generalization ability on unseen problems, which is attributed to the prompts used in the proposed method, which enrich the representation of scenario-specific information and enhance the generalization ability.

[0169] Combined with Figure 3 , the overall performance of the digital twin interaction experience optimization method based on the generative artificial intelligence algorithm proposed in the present invention is better than that of the task-specific reinforcement learning algorithm and the DecisionTransformer without prompts in different DT scenarios. The reason is that the prompts of the proposed method provide rich scenario-specific information to help guide action generation, thus improving the generalization ability of the policy. In addition, the DecisionTransformer without prompts is better than the task-specific reinforcement learning algorithm because its reward mechanism partially captures the scenario-specific information to help it maximize the reward, thus guiding the generation of actions that contribute to achieving the expected reward .

[0170] The proposed DT-based RIS-assisted interactive system has significant application potential, especially in the field of distance education. By combining GAI with DT, such a system can provide a high-quality interactive learning experience for distance education. Specifically, in distance education applications, DT can reflect the status of physical entities in real time, such as the allocation of learning resources and the arrangement of teaching activities, turning distance education from a simple static video viewing into a dynamic and interactive learning process. This system can also intelligently adjust the presentation of learning content according to the needs and behaviors of students, providing a personalized learning experience. With the help of RIS, the data transmission between students and the virtual learning environment can be more efficient, reducing latency and improving the quality of interaction. In addition, by using GAI technology, the system can automatically optimize resource allocation according to the learning progress and feedback, thus maximizing the quality of the students' learning experience (QoE). This intelligent resource allocation and real-time feedback mechanism makes the effect of distance education closer to that of a face-to-face teaching environment.

Claims

1. A method for optimizing the digital twin interaction experience based on a generative artificial intelligence algorithm, characterized in that, This method is based on a generative artificial intelligence algorithm to jointly optimize the phase shift matrix of the intelligent reflecting surface, the receive and transmit beamforming matrices of the digital twin server, the feedback signal rendering resolution configuration, and the computing resource allocation, aiming to maximize the subjective and objective experience of users in the interaction and achieve the generalization of the algorithm to cope with the random evolution of the digital twin model; The steps are as follows: (1) Construct an uplink and downlink transmission model for intelligent surface-assisted digital twin interaction to characterize the uplink and downlink channels and the corresponding uplink and downlink transmission rates; (2) Establish a quality of user experience model, and the mathematical expression of this model is: where is the preference weight of the user, t represents the time slot, is a mathematical function of the subjective experience model of the mobile user, represents the objective experience of the mobile user with the interactive end-to-end delay, and represent the weights of the user for subjective and objective experiences respectively; where ε max represents the maximum subjective perceived quality that each mobile user can achieve, and represents the maximum interactive round-trip delay that each mobile user can tolerate; (3) Construct an optimization problem for maximizing the user experience, and its mathematical representation is: The constraint conditions are: Among them, Θ(i,t) is the phase shift matrix, V(i,t):={v1(i,t),…,v K (i,t)} is the receive beamforming matrix, W(i,t):={w1(i,t),…,w K (i,t)} is the transmit beamforming matrix, E k (i,t) is the feedback signal rendering resolution allocation, f k (i,t) represents the computing resource allocation; (4) Construct the optimization problem described in step (3) into a Markov decision model, then implicitly represent the information of the task based on the decision trajectory, establish a prompt guidance mechanism based on the decision trajectory for different scenarios to capture new artificial intelligence scenario-specific information, and then use the causal Transformer algorithm to obtain decisions, including constructing a prompt-guided artificial intelligence combined with a zero-forcing optimization algorithm to help derive high-dimensional decisions to maximize the quality of user experience obtained by all users when interacting with any artificial intelligence scenario.

2. The digital twin interaction experience optimization method based on the generative artificial intelligence algorithm according to claim 1, wherein The specific modeling method of step (1) includes: (11) The direct uplink channel between each mobile user k∈K and the DT server a is modeled as: where ρ is the path loss coefficient, α k,a is the path loss exponent, d k,a (i) is the distance between mobile user k and the DT server, is the non-line-of-sight component of link a, and each element follows (12) The non-direct uplink channel between each mobile user k and the intelligent hypersurface r is modeled as: where α k,r is the path loss exponent, d k,r (i) is the distance between the mobile user k and the IRS, is the NLOS component of the link, and each element follows G k,r is the Rician factor, is the LOS component of the link; (13) The channel between the intelligent hypersurface and the digital twin server is modeled as: where d r,a represents the distance between the intelligent metasurface and the digital twin server, G r,a is the Rician factor; (14) The signal received by the digital twin server is modeled as: where is the uplink transmission power of mobile user k∈K, and v k (i,t) is the beamforming vector, is the noise, represents the channel gain of other users, represents the signal received by the digital twin server from other users, and p m represents the transmission power of other users, and H is the conjugate transpose; (15) The uplink transmission rate is: where b is the bandwidth, and is the uplink signal-to-noise ratio; (16) Downlink transmission rate is:

3. The digital twin interaction experience optimization method based on the generative artificial intelligence algorithm according to claim 1, wherein Step (2) for the construction of the user subjective quality of experience model includes the following process: (21) Establish the subjective experience of the mobile user as: where E min is the minimum resolution requirement of the user, and E k (i, t) is the resolution assigned to the user; (22) The transmission delay of the uplink interaction signal is expressed as: Among them is the data volume size of the uplink signal; (23) The delay for processing the feedback signal is expressed as: (24) The downlink transmission delay of the feedback is expressed as: Among them is the data volume size of the feedback signal; (25) The objective experience of the mobile user is expressed by the end-to-end delay of the interaction: (26) Combine the subjective and objective experiences to obtain the expression of the quality of experience model: wherein is the preference weight of the user, 4. The digital twin interaction experience optimization method based on the generative artificial intelligence algorithm according to claim 1, characterized in that, Step (4) constructs a prompt-guided Decision Transformer combined with a zero-forcing optimization algorithm, specifically including: (41) The problem is reconstructed into a Markov decision, and the mathematical representation is where Status: Action: Reward: State transition: Pr(i, s(i, t+1)|s(i, t), a(i, t)) ∈ [0, 1 (42) Design prompt mechanism: This formula is based on the information of the implicit representation task of the decision-making trajectory; (43) Calculate the input embedding layer: Adopt prompt τ ＊ (i), and the most recent trajectory decision τ(i, t) of length L before the current time step as the historical trajectory; Use a modality-specific trainable linear layer for embedding to process the three different token modalities of the prompt reward target, state, and decision; Use a trainable linear layer to add positional embeddings to the tokens within the same decision trajectory tuple, and similarly, also apply a trainable linear layer to τ(i, t); (44) Obtain decisions using the causal Transformer: The embedded input tokens are fed into a causal transformer, which consists of identical decoders stacked; each decoder processes the tokens through a masked multi-head self-attention module to capture the dependencies between the tokens; a feed-forward layer enhances this representation through a per-position transformation, while layer normalization stabilizes the training and facilitates gradient updates; the causal transformer processes the tokens sequentially in stacked decoders, finally outputting the hidden states; Then, input these hidden states into a trainable linear decision prediction layer to generate decisions; (45) Design a zero-forcing optimization algorithm: Given by the prompt-guided Decision Transformer calculation: The problem of optimizing κ(i, t) = {V(i, t), W(i, t)} is constructed as follows: The constraint conditions are: (46) Obtained according to the zero-forcing algorithm Wherein (47) Based on the obtained V(i, t), the problem is reconstructed as: Constraints According to the zero forcing algorithm, In the above formula, P is a diagonal matrix, where the k-th diagonal element is the time step of the mobile user in the time step, i.e., the received power p DL k(i, t); Set the constraint conditions: Based on the constraint conditions, the optimization problem is further reconstructed as: Constraint conditions: For the above expression, use the water-filling algorithm to solve this problem and obtain a closed-form solution: where the k-th diagonal element is is a normalization factor used to ensure that In addition,[ is the minimum received power limit for mobile users.

Images