Port communication network dynamic relay selection method based on multi-arm bandit and coalition game

Abstract

The invention discloses a port communication network dynamic relay selection method based on a multi-arm bandit and a coalition game, which belongs to the field of port communication, and comprises the following steps: a mobile terminal periodically collects context information and historical communication feedback of available relay nodes; constructing a context confrontation multi-arm bandit model, estimating a relay transmission rate by using a ridge regression method, and estimating queuing delay in a weighted moving average mode; introducing a joint learning mechanism based on a coalition game, dividing terminal nodes with similar communication conditions into a coalition, and sharing a relay context and feedback information to accelerate model convergence; dynamically updating a model parameter attenuation factor and a time sensitive weight according to the current communication performance deviation; a utility function fusing multiple communication indexes is constructed, throughput and delay are comprehensively considered, and optimal relay selection is realized in multiple scenes. According to the invention, the method can effectively cope with the highly dynamic change of the channel state in the port and relay selection under the multi-service heterogeneous requirements, and remarkably improves the communication efficiency and stability of the system.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of port communication, and particularly relates to a port communication network dynamic relay selection method based on a multi-armed bandit and coalition game. BACKGROUND

[0002] With the rapid development of Internet of Things technology, seaports all over the world are transforming towards intelligentization and automation operation. Intelligent terminals such as automated guided vehicles, intelligent cranes and unmanned inspection equipment put forward higher real-time and stability requirements for wireless communication systems. The port communication system needs to realize multi-service coordination, high reliability and low delay data transmission under complex conditions such as high interference, multipath fading and high-speed node movement, which brings significant challenges to traditional relay selection and resource scheduling methods, mainly including:

[0003] (1) The port communication condition has high dynamicity and non-idealness. Influenced by harsh physical factors such as salt spray and dust and high mobility of terminal nodes, channel state information (CSI) is difficult to accurately obtain and changes frequently. In the unknown or unpredictable CSI scene, how to design a relay selection algorithm with adaptive ability to cope with complex network state is a key technical problem currently faced.

[0004] (2) The port service type is highly heterogeneous, and different services have significant differences in communication performance requirements. For example, control type feedback data is extremely sensitive to communication delay, while video stream transmission and multi-sensor fusion data require higher bandwidth and throughput. Therefore, how to balance between multiple performance indicators, optimize relay selection to adapt to multi-scene services while meeting the needs of low delay and high throughput is another important problem to be solved.

[0005] To cope with the above problems, the multi-armed bandit model is gradually applied to the relay selection problem because it has online learning and dynamic decision-making capabilities. However, existing researches mostly ignore the context information between terminal nodes, resulting in slow learning convergence speed and weak strategy generalization ability in high dynamic scenarios. At the same time, existing relay selection methods usually optimize based on a single performance indicator (such as channel capacity or transmission rate), lack comprehensive consideration of multiple communication performance indicators such as communication delay and channel capacity, and are difficult to meet the complex needs of multiple services running simultaneously in the port system. In addition, in the face of problems such as dense distribution of terminal nodes and incomplete environmental perception, it is difficult to obtain effective strategies by relying only on single-node local learning.

[0006] Therefore, how to construct a relay selection method that supports context awareness, coalition game, dynamic update and is oriented to the optimization of multiple communication performance indicators has become one of the key problems to be solved in intelligent port communication systems. SUMMARY

[0007] The purpose of the present application is to provide a port communication network dynamic relay selection method based on multi-armed bandit and coalition game, which overcomes the problems of strong communication channel dynamics, high business demand diversity, insufficient use of context information, slow strategy convergence speed and the like in the prior art, and realizes adaptive optimization of communication performance in a multi-service scenario.

[0008] To solve the above technical problems, the technical solution adopted by the present application is:

[0009] A port communication network dynamic relay selection method based on multi-armed bandit and coalition game, comprising the following steps,

[0010] S1, a mobile terminal periodically collects context information and historical communication feedback of available relay nodes for subsequent relay performance estimation and strategy optimization;

[0011] S2, a context-adversarial multi-armed bandit model is constructed, the context and historical feedback of the relay are modeled using the ridge regression method, the instantaneous transmission rate of the relay is estimated, and the weighted moving average method is used to estimate the relay queuing delay;

[0012] S3, a joint learning mechanism based on coalition game is introduced, mobile terminals with similar communication conditions are divided into communication alliances according to the context similarity, relay context and feedback information are shared within the alliance, and parameters are trained together to speed up model convergence;

[0013] S4, according to the current communication performance estimation deviation, the decay factor and time sensitivity weight of the model parameters are dynamically updated to improve the adaptability to communication state changes and enhance the online learning ability of the model;

[0014] S5, a relay selection utility function integrating multiple communication indicators is constructed, considering at least key performance indicators including transmission rate and communication delay, and dynamically selecting the optimal relay to adapt to performance requirements in a multi-service scenario.

[0015] The further improvement of the technical solution of the present application is that in S1, the context information collected by the mobile terminal includes but is not limited to signal strength, terminal speed, motion direction, system capacity history record, constituting a six-dimensional feature vector x t,n , and synchronously collecting transmission feedback of the relay node in the past communication round to construct a relay performance estimation basic data set.

[0016] The further improvement of the technical solution of the present application is that in S2, the estimation of the relay transmission rate uses the ridge regression method, and the regression parameters t,n , c t,n are obtained according to the historical context-feedback (G The calculation formula is:

[0017]

[0018] where λ is a regularization parameter; G t,n denotes the historical context matrix; c t,n denotes the historical transmission rate feedback;

[0019] According to the estimated parameters, the transmission rate of the relay node n at the current time is predicted as:

[0020]

[0021] To estimate the queuing delay, a weighted moving average model is used:

[0022]

[0023] where ω(t',b) is a time window weight function, τ t',n is the historical queuing delay.

[0024] Further improvement of the technical scheme of the present application is that the communication alliance in S3 is formed through context similarity, specifically, at each round of time t, the terminal node i compares the context feature quantization difference with other nodes j, if the difference does not exceed the preset threshold ε, the communication alliance is formed:

[0025]

[0026] Further improvement of the technical scheme of the present application is that the terminal nodes in the alliance share context information and feedback information, and jointly train the process parameters of the ridge regression process in a transfer learning manner, the members in the alliance C k use the process parameters when performing transmission rate estimation is expressed as:

[0027]

[0028] The process parameters used by the members in the alliance C k when performing queuing delay estimation are expressed as:

[0029]

[0030] Through sharing information and collaborative modeling, the convergence speed is improved.

[0031] Further improvement of the technical scheme of the present application is that the model parameter decay factor γ t,n is dynamically adjusted according to the error between the relay estimation performance and the actual feedback, and the adjustment formula is as follows:

[0032]

[0033] wherein, gamma max , gamma min is a set range boundary, k is a scaling factor.

[0034] Further improvement of the technical scheme of the present application is that the time-sensitive coefficient is calculated in a time window sliding manner in S4, the current time step is T, the window length is T w , and the weighting function of the bth window segment is:

[0035] omega(t',b) = mu (B-b) + phi[t'-(b-1)T w -T]

[0036] wherein, mu is an attenuation base in (0, 1), phi is a position correction function in the window, and B represents the total number of window segments.

[0037] Further improvement of the technical scheme of the present application is that the relay selection in S5 is based on the principle of maximizing the utility function, and the utility function is:

[0038]

[0039] wherein, alpha i + beta i = 1, k 1,i , k 2,i are non-negative weight coefficients for balancing the relative influence of throughput rate and delay, and the optimal relay selection is:

[0040]

[0041] Due to the adoption of the above technical scheme, the present application has achieved the following technical progress:

[0042] 1. The present application estimates the relay transmission performance by using context information and historical communication feedback, and dynamically updates the process parameters to enhance adaptability, thereby realizing accurate evaluation of relay performance. After introducing the coalition game theory, terminals in similar context states can form a coalition to share information and cooperate in training, significantly improving learning efficiency and enabling the algorithm to quickly adapt to changes in the communication network.

[0043] 2. The present application can also dynamically adjust the relay selection strategy according to the specific communication needs of different data types, ensuring optimal performance in various port business scenarios. Simulation results show that in highly dynamic and complex communication demand environments, the present application has better performance compared to existing relay selection strategies.

[0044] 3、The application is superior to the existing relay selection method in communication environment adaptability, model convergence speed, multi-terminal cooperation capability and multi-index optimization performance, has wide practical value and popularization potential, and is particularly suitable for complex, dynamic and heterogeneous port communication network environments. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed to be used in the embodiments or prior art description will be briefly introduced. Obviously, the drawings in the following description are some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0046] Figure 1 is a structural block diagram of a port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the embodiment of the present application;

[0047] Figure 2 is a data transmission process schematic diagram based on a coalition mechanism in a port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the embodiment of the present application;

[0048] Figure 3 is a time sensitivity coefficient schematic diagram in a port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the embodiment of the present application;

[0049] Figure 4 is a simulation topology graph used in the embodiment of the present application;

[0050] Figure 5 is a simulation graph for high-throughput data relay optimal selection rate in the embodiment of the present application;

[0051] Figure 6 is a simulation graph for high-throughput data cumulative regret value in the embodiment of the present application;

[0052] Figure 7 is a simulation graph for high-throughput data cumulative network throughput in the embodiment of the present application;

[0053] Figure 8 is a simulation graph for time delay sensitive data relay optimal selection rate in the embodiment of the present application;

[0054] Figure 9 is a simulation graph for time delay sensitive data cumulative regret value in the embodiment of the present application;

[0055] Figure 10 is a simulation graph for time delay sensitive data cumulative time delay in the embodiment of the present application. DETAILED DESCRIPTION

[0056] It should be noted that the terms "comprising" and "having" and any variations thereof in the specification and in the claims and the above drawings are intended to cover a non-exclusive inclusion, for example, a process, method, system, product or apparatus that includes a list of steps or units not necessarily limited to those clearly listed, but can include other steps or units not clearly listed or inherent to such processes, methods, products or apparatuses.

[0057] The technical problems to be solved by the present application include:

[0058] (1) Under the condition that channel state information is unpredictable or difficult to obtain, how to realize accurate estimation of relay transmission performance through context modeling and feedback learning;

[0059] (2) Under the premise that terminal nodes are widely distributed and have limited sensing ability, how to improve model learning efficiency and strategy stability through a cooperative mechanism;

[0060] (3) How to realize the optimal relay selection strategy under multi-index fusion while meeting the heterogeneous communication needs of low latency and high throughput.

[0061] The present application will be further described in detail below in conjunction with the drawings and embodiments:

[0062] As shown in Figure 1 , a port communication network dynamic relay selection method based on multi-arm bandit and coalition game, comprising the following steps:

[0063] S1, the mobile terminal periodically collects the context information and historical communication feedback of the available relay nodes, which are used for subsequent relay performance estimation and strategy optimization;

[0064] Specifically, each mobile terminal i will perceive the candidate relay set N t within the communication range in each time slot t, and sample the context feature vector x t,n ∈R d of the relay node, including but not limited to signal strength, terminal speed, motion direction, system capacity history, constituting a six-dimensional feature vector x t,n ; and synchronously collect the transmission feedback of the relay node in the past communication round to build a relay performance estimation basis dataset. At the same time, the terminal records the historical interaction feedback c t,n such as the transmission rate R t,n , queuing delay τ t,n , etc. returned by the relay, as the environmental reward input for strategy optimization.

[0065] S2, construct the context against multi-arm bandit model, utilize the ridge regression method to model the context and historical feedback of relay, estimate the instant transmission rate of relay, and adopt the weighted moving average method to estimate the queuing delay of relay;

[0066] Specifically, the application introduces a context multi-arm bandit structure in policy selection. To estimate the relay transmission rate, a parameter estimation method in the form of ridge regression is introduced:

[0067]

[0068] Wherein, lambda is a regularization parameter, G t,n represents the historical context matrix, c t,n represents the historical transmission rate feedback.

[0069] According to the estimated parameters, the transmission rate of relay node n at the current time is predicted as:

[0070]

[0071] In addition, to estimate the queuing delay, a weighted moving average model is used:

[0072]

[0073] Wherein, omega(t',b) is a time window weight function, tau t',n is the historical queuing delay.

[0074] S3, propose a joint learning mechanism based on coalition game, according to the context similarity, the mobile terminals with similar communication conditions are divided into communication alliances, the relay context and feedback information are shared in the alliance, and the parameters are trained together to speed up the convergence speed of the context against multi-arm bandit model;

[0075] Specifically, in order to efficiently capture the change of relay performance in dynamic environment, the alliance game framework is introduced for relay cooperative learning in port communication network, as shown in Figure 2 By constructing a communication alliance, each member can share the context information, feedback data and model training parameters of the relay, so as to speed up the learning process and improve the convergence speed of the algorithm.

[0076] S3.1, alliance division strategy

[0077] The communication alliance mechanism is introduced, and the alliance is dynamically constructed through context similarity. If the context features of terminal i and other terminal j satisfy:

[0078]

[0079] The terminal nodes accelerate the learning process by sharing data and improve the convergence speed of each node in the alliance. The alliance utility function is defined as:

[0080] U gain =ρ·ln(1+D share ) (5)

[0081] At the same time, data sharing among alliance members brings certain energy consumption costs. The energy consumption of the terminal node is determined by its transmission power and energy consumption coefficient:

[0082] U cost =p i ·ξ i (6)

[0083] Therefore, the net coalition utility of the terminal node is:

[0084] U coa,i =ρ·ln(1+D share )-p i ·ξ i (7)

[0085] At the same time, after new members join, their utility should be improved without reducing the utility of existing members, that is:

[0086]

[0087] The terminal node can exit the current alliance at any time and choose to join another alliance. The present invention designs a distributed alliance formation algorithm to achieve a Nash-stable alliance structure.

[0088] S3.2. Collaborative Training Based on Transfer Learning

[0089] The terminals in the alliance share historical observation data and jointly update the modeling parameters. k The process parameters used by members in the transmission rate estimation and It can be expressed as:

[0090]

[0091] The final estimated coalition parameters are:

[0092]

[0093] League C k The process parameters used by members in the queue delay estimation and It can be expressed as:

[0094]

[0095] The cooperative learning mechanism significantly improves modeling accuracy and strategy stability, reduces exploration cost in the learning process, and is suitable for deployment in a multi-terminal cooperative scenario in a port environment.

[0096] S4, according to the estimation deviation of the current communication performance, dynamically updating the attenuation factor and time-sensitive weight of the model parameter, improving the adaptability to the change of communication state, and enhancing the online learning ability of the model;

[0097] Specifically, in order to enhance the adaptability of the system to the dynamic communication environment, the application constructs a parameter updating mechanism to improve the timeliness of data. The mechanism dynamically adjusts the attenuation factor according to the estimation deviation of the relay performance, and simultaneously introduces a time-sensitive coefficient to prefer recent observations, so as to reduce the error caused by outdated data.

[0098] S4.1, adaptive online attenuation factor mechanism of dynamic transmission rate estimation

[0099] In order to effectively adapt to the dynamic communication environment, an online parameter updating mechanism with a dynamic attenuation factor is introduced in the transmission rate estimation. The mechanism dynamically adjusts the attenuation factor according to the difference between the actual value and the estimated value, so as to more accurately perceive the change of the communication condition. When the channel state fluctuates, the relay transmission performance changes accordingly. In order to capture the influence of the change of the communication condition over time on the relay performance, the following loss function is defined to measure the deviation between the actual transmission rate and the estimated value:

[0100]

[0101] The loss function reflects the effectiveness of the process parameter under the current communication environment. When the loss is large, it means that the process parameter has become invalid, and the model parameter needs to be updated in time. At this time, the attenuation factor should be reduced, and the weight of outdated data should be reduced, so that the model can better respond to the current communication environment:

[0102]

[0103] The regression matrix and vector are updated by using the attenuation factor:

[0104]

[0105] Through the above steps, the model can continuously adapt to the dynamic change and improve the estimation accuracy of the transmission rate.

[0106] S4.2, time-sensitive online updating mechanism of queuing delay estimation

[0107] In port communication networks, the connection state between terminals and relays is constantly changing due to the mobility of terminal devices, while the set of available relays for each terminal is also evolving, resulting in dynamic changes in the load state of relays. To this end, a time-sensitive parameter online updating mechanism is proposed, which enables the algorithm to better adapt to the dynamic changes of the communication system and gradually approach the true value of the actual queuing delay. As shown in Figure 3 , the time domain T is divided into multiple time periods, and the time window length is T w , the number of complete time periods at the current time t=T is denoted as B:

[0108]

[0109] The discount factor ω reflects the time decay effect of historical data. For the t'th time slot and the b'th time period, the weight function is defined as:

[0110]

[0111] where μ(B-b) is the base weight of the time period, is used to adjust the relative weight of different time slots within the time period. This structure ensures that recent data is given a higher weight and more accurately reflects the current system state. At the current time t, the selected relay node i processes all the backhaul information in the current time period with weighting, and updates the process parameters of the queuing delay:

[0112]

[0113] By fusing these weighted observations, this mechanism enables the queuing delay estimation to dynamically reflect recent data and reduce the impact of outdated information, thereby better adapting to changes in the communication system state and achieving more accurate and reliable relay selection.

[0114] S5, a relay selection utility function is constructed by fusing multiple communication indicators, considering at least key performance indicators including transmission rate and communication delay, and dynamically selecting the optimal relay to adapt to performance requirements in multiple service scenarios.

[0115] Specifically, in order to achieve more refined relay selection decisions in heterogeneous service scenarios, a relay utility function is constructed by fusing multiple communication performance indicators, aiming to consider the trade-off between throughput and communication delay. The utility function is defined as follows:

[0116]

[0117] where α i +β i =1, k 1,i , k 2,iare non-negative weight coefficients, used to balance the relative influence of throughput and delay, the optimal relay selection is:

[0118]

[0119] That is, in each time step, the terminal selects the candidate relay node with the highest utility value for communication, thereby dynamically adapting to environmental changes and improving overall system performance.

[0120] Performance evaluation of the present application:

[0121] Specifically, a port heterogeneous communication network simulation environment based on the real topological structure of Qinhuangdao Port is adopted, as shown in Figure 4 The simulation takes an autonomous surface ship as the test object, considering the influence of its mobility on the dynamic changes of the communication topology, thereby exacerbating the dynamics of the candidate relay set. The communication data is generated by MATLAB R2023b, covering the following influencing factors: path loss, environmental attenuation (salt spray, dust, typhoon), multipath fading, shadow effect, etc. Each relay node adopts an M / M / 1 queuing model, and the queuing delay under different terminal densities is set to simulate the influence of network congestion on communication performance.

[0122] Firstly, the performance of each algorithm under high throughput demand data is compared, covering three key performance indicators: optimal selection rate, cumulative regret value, and cumulative throughput. Figure 5 represents the optimal selection rate, reflecting the efficiency of each algorithm in selecting the optimal relay under dynamic communication conditions; Figure 6 represents the cumulative regret value, reflecting the cumulative gap between the rewards obtained by the algorithm and the theoretical maximum reward; Figure 7 The cumulative network throughput is shown, which reflects the overall data transmission performance. The port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided in the present application combines the coalition game learning mechanism and the dynamic parameter updating rule, and performs best in all indicators. The coalition game learning mechanism enables the algorithm to quickly adapt to changes in communication conditions, while the dynamic parameter updating rule improves the accuracy of perception, thereby achieving a higher optimal selection rate and a lower cumulative regret value in a dynamic environment, with a significantly faster convergence speed than other methods. In contrast, DB-LinUCB only uses dynamic parameter updating but lacks coalition learning, CB-LinUCB uses coalition learning but lacks dynamic updating, and LinUCB lacks both, so they perform poorly under dynamic topology. In addition, the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided in the present application also achieves the highest cumulative throughput, thanks to its ability to quickly identify and select the optimal relay; while UCB and LinUCB rely only on the basic exploration-exploitation mechanism, and respond more slowly to changes in relay quality, resulting in significantly poorer performance under dynamic communication conditions.

[0123] Then, the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the application is compared with DB-UCB, CBUCB, UCB and random selection strategy in processing time delay sensitive data, and the performance indexes of optimal selection rate, cumulative regret value and cumulative time delay are mainly investigated. Figure 8 It is shown that, thanks to the combination of coalition game learning and dynamic parameter updating rule, the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the application can quickly and accurately adapt to the change of relay load, so as to realize the highest optimal selection rate. In contrast, DB-UCB lacks a coalition learning mechanism, and UCB lacks both dynamic updating and coalition mechanism, so the convergence speed is slower when selecting the optimal relay. Although CBUCB adopts coalition learning, it is superior to UCB in performance, but due to the lack of dynamic updating, its performance is still inferior to the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the application. Figure 9 It is shown that, the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the application also performs best in minimizing the cumulative regret value, can effectively adapt to the change of relay load, and reduce the exploration cost due to the selection of suboptimal relay, which is a common challenge faced by UCB and CBUCB. Although DB-UCB improves the performance to some extent through dynamic parameter updating, due to the lack of coalition learning, its overall effect is still inferior to the application. Finally, as shown in Figure 10 It is shown that, the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the application realizes the lowest cumulative time delay by quickly identifying and selecting the most suitable relay; while UCB and CBUCB have higher delay due to slower response to real-time network demand. Although DB-UCB is superior to UCB in terms of time delay, due to the slower adaptation speed to dynamic relay load, it still lags behind the port communication network dynamic relay selection method based on multi-armed bandit and coalition game provided by the application.

[0124] In summary, the application provides a port communication network dynamic relay selection method based on multi-arm slot machine and coalition game, which is used to cope with the dynamic relay selection challenge in the heterogeneous port communication environment. In the port communication network, the dynamic change and difficulty to predict of channel state information caused by the poor environmental conditions and the mobility of terminal nodes, and the frequent fluctuation of available relay set caused by the constantly changing network topology. The existing relay selection method often lacks adaptability when coping with these complex and rapidly changing environments, which may have a negative impact on communication performance. The application estimates the relay transmission performance by using context information and historical communication feedback, and dynamically updates the process parameters to enhance adaptability, thereby achieving accurate evaluation of relay performance. After introducing the coalition game theory, terminals in similar context states can form alliances, share information and conduct collaborative training, significantly improving learning efficiency and enabling the algorithm to quickly adapt to changes in the communication network. In addition, the application can dynamically adjust the relay selection strategy according to the specific communication needs of different data types, ensuring optimal performance in various port business scenarios. Simulation results show that in highly dynamic and complex communication demand environments, the application has better performance compared to existing relay selection strategies.

[0125] The above-described embodiments are merely preferred embodiments of the present application and are not intended to limit the concept and scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those skilled in the art without departing from the design concept of the present application shall fall within the protection scope of the present application, and the technical content claimed by the present application has been fully recorded in the claims.

Claims

1. A port communication network dynamic relay selection method based on multi-armed bandit and coalition game, characterized in that, Comprising the following steps, S1, the mobile terminal periodically collects the context information and historical communication feedback of available relay nodes for subsequent relay performance estimation and strategy optimization; S2, construct a context-adversarial multi-armed bandit model, use ridge regression method to model the context and historical feedback of the relay, estimate the instantaneous transmission rate of the relay, and use weighted moving average method to estimate the queuing delay of the relay; S3, introduce a joint learning mechanism based on coalition game, divide mobile terminals with similar communication conditions into communication alliances according to the context similarity, share relay context and feedback information within the alliance, and jointly train parameters to speed up model convergence; S4, according to the current communication performance estimation bias, dynamically update the decay factor and time-sensitive weight of the model parameters, improve the adaptability to the change of communication state, and enhance the online learning ability of the model; S5, construct a relay selection utility function integrating multiple communication indicators, comprehensively consider at least the key performance indicators including transmission rate and communication delay, and dynamically select the optimal relay to adapt to the performance requirements in multiple business scenarios.

2. The method of claim 1, wherein, In S1, the context information collected by the mobile terminal includes but is not limited to signal strength, terminal speed, motion direction, system capacity history, constituting a six-dimensional feature vector x t,n and synchronously collect transmission feedbacks of the relay nodes in past communication rounds to build a relay performance estimation base dataset.

3. The method of claim 1, wherein, In S2, the estimation of the relay transmission rate employs a ridge regression method, which models the historical context-feedback (G t,n ,c t,n ) to obtain regression parameters The calculation formula is: where λ is a regularization parameter; G t,n denotes a historical context matrix; c t,n denotes historical transmission rate feedback; According to the estimated parameters, the transmission rate of relay node n at the current time is: To estimate the queuing delay, a weighted moving average model is used: Among them, ω(t',b) is the time window weight function, τ t',n is the historical queuing delay.

4. The method of claim 1, wherein, The communication alliance in S3 is formed by context similarity, specifically, at each round of time t, terminal node i compares the context feature quantization difference with other nodes j, if it does not exceed the preset threshold ε, it forms a communication alliance:

5. The method of claim 4, wherein, The terminal nodes inside the alliance share context information and feedback information, and jointly train the ridge regression process parameters in a transfer learning manner, and the members in the alliance C k use process parameters when performing transmission rate estimation and is represented as: Alliance C k Process parameters used by members in Alliance C when performing queue delay estimation With is represented as: Through sharing information and collaborative modeling to improve the convergence speed.

6. The method of claim 1, wherein, Model parameter decay factor γ in S4 t,n The adjustment formula is as follows according to the error dynamics between the relay estimated performance and the actual feedback: where γ max , γ min are set range boundaries, and k is a scaling factor.

7. The method of claim 6, wherein, In S4, the time-sensitive coefficient is calculated in a time window sliding manner, the current time step is T, and the window length is T w , and the weighting function of the bth segment window is ω(t',b) = μ (B-b) + φ[t' - (b - 1)T w - T] Where μ∈(0,1) is the decay base, φ is the position correction function in the window, and B represents the total number of window segments.

8. The method of claim 1, wherein, In S5, the relay selection is based on the principle of maximizing the utility function, and the utility function is: where α i + β i = 1, k 1,i , k 2,i are non-negative weight coefficients balancing the relative impact of throughput and delay, and the optimal relay selection is:

Images