RSMA-MEC system resource allocation method based on data compression

Abstract

The invention discloses an RSMA-MEC system resource allocation method based on data compression. The method specifically comprises the following steps: step 1, constructing a system model comprising a plurality of user equipment and a plurality of mobile edge computing MEC servers; step 2, establishing an end-to-end time delay model, and comprehensively considering local data compression time delay of a user, task transmission time delay through an RSMA link, and data decompression time delay and task calculation time delay of an MEC server side; 3, constructing a joint optimization problem for minimizing the maximum time delay; 4, modeling a joint optimization problem into a Markov decision process; 5, solving a Markov decision process by adopting a near-end strategy optimization algorithm, and outputting an optimal resource allocation strategy and a joint optimization decision; and step 6, carrying out resource allocation on the user equipment and the MEC server in the system. According to the invention, the time delay performance and the resource utilization efficiency of the multi-user multi-server MEC system are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of wireless communication technology, and more specifically to a resource allocation method for RSMA-MEC systems based on data compression. Background Technology

[0002] With the widespread adoption of 5G mobile communication technology and the Internet of Things (IoT), computationally intensive and latency-sensitive applications such as augmented reality, virtual reality, and industrial automation are experiencing explosive growth. Mobile edge computing (MEC) effectively reduces task processing latency and the transmission burden on the core network by offloading computing power to the network edge. However, in scenarios with massive numbers of connected terminal devices, MEC systems face two major challenges: first, the scarcity of wireless spectrum resources makes it difficult to support high-concurrency task offloading; second, the sheer volume of raw data means that direct transmission can exacerbate channel congestion, leading to increased transmission latency.

[0003] To improve spectral efficiency, non-orthogonal multiple access (NOMA) technology has been introduced into MEC systems. However, NOMA technology places high demands on receiver signal processing capabilities, and its performance advantages are limited when the differences in channel conditions between users are small.

[0004] Existing technologies often focus on improving single technical paths, lacking an efficient framework for jointly optimizing multi-dimensional resources such as communication, compression, and computation. Especially in complex MEC scenarios with multiple users and servers, how to coordinate RSMA and data compression technologies to dynamically allocate resources and minimize system latency remains a pressing technical challenge. Summary of the Invention

[0005] I. Technical problems to be solved To address the shortcomings of existing technologies, this invention proposes a resource allocation method, system, and storage medium for mobile edge computing systems based on data compression and RSMA. The aim is to effectively reduce task processing latency in multi-user, multi-server MEC systems by using deep reinforcement learning technology to jointly optimize communication, computing, and data compression resources.

[0006] II. Specific Technical Solutions A resource allocation method for an RSMA-MEC system based on data compression, comprising the following steps: Step 1: Construct a system model that includes multiple user devices and multiple mobile edge computing (MEC) servers; Step 2: Establish an end-to-end latency model, comprehensively considering the data compression latency of the user locally, the task transmission latency through the RSMA link, and the data decompression latency and task computation latency of the MEC server. Step 3: Construct a joint optimization problem that minimizes the maximum latency; Step 4: Model the joint optimization problem as a Markov decision process; Step 5: Use the near-end policy optimization algorithm to solve the Markov decision process and output the optimal resource allocation strategy and joint optimization decision; Step 6: Allocate resources between user devices and MEC servers in the system.

[0007] Preferably, in step two, the latency model includes data compression latency and task transmission latency on the user side, as well as data decompression latency and task computation latency on the MEC server side.

[0008] As a preferred option: In step three, the variables for joint optimization include user task offloading association, transmission power of public and private messages in RSMA communication, public message rate allocation, compression ratio of task data, compression coefficient of task data, and computing resource allocation of MEC server.

[0009] As a preferred embodiment, in step four, the joint optimization problem is modeled as a Markov decision process, specifically including defining a state space, which includes the amount of task data for all users, the channel gain between all user devices and all MEC servers, the computational task intensity of users, and the local computation frequency of users; defining an action space, whose elements are the set of all variables of the joint optimization; and defining a reward function, which is related to the negative value of the maximum latency of the computational tasks of all users and includes a penalty term for violating constraints.

[0010] Preferably, step five involves using an Actor-Critic network architecture, where the Actor network selects actions based on the current state, and the Critic network evaluates the value of state-action pairs; collecting experience data through interaction with the environment and storing it in an experience replay buffer; sampling small batches of data from the experience replay buffer and calculating the advantage function; updating the parameters of the Actor network by maximizing an alternative objective function that includes a shearing mechanism; and updating the parameters of the Critic network by minimizing the mean squared error of the value function.

[0011] A system for resource allocation in an RSMA-MEC system based on data compression is provided, which includes a first processing unit for constructing a multi-user, multi-server RSMA-MEC system model. The second processing unit is used to determine the task processing latency model of the RSMA-MEC system based on data compression; The third processing unit is used to determine the optimization problem model based on the task processing latency model. The goal of the optimization problem model is to minimize the maximum latency for users to process tasks on different MEC servers. The fourth processing unit is used to obtain a deep reinforcement learning framework based on the optimization problem model; The fifth processing unit is used to obtain the optimal policy based on the PPO-based deep reinforcement learning optimization framework; The sixth processing unit is used to allocate resources to the RSMA-MEC system based on the optimal strategy.

[0012] The beneficial effects of this invention are as follows: Strong collaborative optimization capability: This invention deeply integrates RSMA multiple access, data compression technology and deep reinforcement learning, and realizes unified dynamic optimization of communication resources (power, rate), computing resources (server CPU frequency) and data compression parameters (proportion, coefficient), breaking the limitations of the technology separation in traditional solutions.

[0013] Superior latency performance: By minimizing the maximum latency for users as the optimization objective and leveraging the powerful decision-making capabilities of the PPO algorithm, this invention can effectively adapt to highly dynamic network environments and maintain low end-to-end latency even when the number of users and the amount of task data change, thus significantly improving the user experience.

[0014] High system efficiency: The PPO algorithm adopted has the advantages of high sampling efficiency and good training stability. It can efficiently process high-dimensional and continuous action spaces, making it suitable for online learning and decision-making in practical MEC systems and improving the overall resource utilization efficiency of the system.

[0015] High flexibility: This invention considers the dynamic adjustment of the compression coefficient, which can adaptively adjust the compression degree according to channel conditions and server load, achieving the best balance between transmission overhead and compression calculation overhead, and avoiding the performance loss of fixed compression strategies. Attached Figure Description

[0016] Figure 1 This is a flowchart of the resource allocation method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the RSMA-MEC system architecture based on data compression provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the time-slot process of a user computing task provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the PPO algorithm framework provided in an embodiment of the present invention; Figure 5 This is a schematic diagram comparing the convergence performance of different algorithms provided in the embodiments of the present invention; Figure 6 This is a schematic diagram illustrating the maximum latency variation of user computing tasks under different numbers of users, as provided in an embodiment of the present invention.

[0017] Figure 7 This is a schematic diagram illustrating the maximum latency variation of user computing tasks under different MEC server computing frequencies provided in this embodiment of the invention.

[0018] Figure 8 This is a schematic diagram illustrating the relationship between the number of users and the maximum latency of user computing tasks under different MEC servers provided in this embodiment of the invention.

[0019] Figure 9 This is a schematic diagram illustrating the relationship between the maximum transmit power and the maximum latency of user computing tasks under different communication schemes provided in the embodiments of the present invention.

[0020] Figure 10 This is a schematic diagram illustrating the relationship between the amount of task data and the maximum latency of user computing tasks under different compression coefficients provided in this embodiment of the invention. Detailed Implementation

[0021] The preferred embodiments of the present invention will now be described in detail with reference to the accompanying drawings, so that the advantages and features of the present invention can be more easily understood by those skilled in the art, thereby providing a clearer and more explicit definition of the scope of protection of the present invention.

[0022] like Figure 1 As shown, this invention constructs a multi-user, multi-server system architecture based on data compression using RSMA-MEC. This architecture consists of a cluster of MEC servers with heterogeneous computing capabilities, deployed on base station facilities of mobile communication networks, aiming to provide distributed computing services to mobile users within the region. Given the limited energy of user equipment, all computing tasks are offloaded to MEC servers for processing. Simultaneously, it supports modular decomposition of computing tasks and dynamic allocation to servers of different performance levels for collaborative processing. It is particularly noteworthy that for complex machine learning tasks such as image processing, the multi-server collaborative offloading mechanism can significantly improve computational efficiency. Multiple video streams generated in scenarios such as vehicle-mounted systems, medical monitoring, and security surveillance can implement differentiated offloading strategies based on the real-time load status and computational performance characteristics of each MEC server. This architecture achieves two optimization goals: firstly, decomposing the overall computational complexity across multiple computing nodes, effectively alleviating the bottleneck of single-server processing; secondly, dynamically matching server computing resources according to the accuracy requirements of each sub-task, optimizing system energy efficiency while ensuring the processing accuracy of core tasks.

[0023] like Figure 2 As shown, consider an uplink using a single-antenna configuration architecture, where both the base station deploying the MEC server and the user equipment are configured with a single antenna element, where the offloaded user set is defined. MEC server set , The total bandwidth of the MEC system is BHz, total time period T Divided into several durations Equal-length time slots, each time slot consisting of t Identify and label the data. To address interference suppression issues in multi-user, multi-server communication scenarios, set... N Each user offloads tasks through an independent sub-band, and bandwidth is allocated to each user. This is to eliminate interference between users. During the uninstallation process, users... n Send to MEC server m The size of the task to be processed is ,in Indicates user n Task input data volume, using Indicates user n The computational tasks are performed on the MEC server. m The unloading allocation ratio coefficient. This parameter must satisfy the normalization constraint. ,when Time indicates user n Do not send to the server m Unload computing tasks, and This corresponds to the extreme scenario of completely uninstalling to a single server.

[0024] Data transmission from a single user to multiple MEC servers utilizes RSMA technology to reuse spectrum resources, enabling parallel transmission within the same frequency band. Through a hierarchical signal processing mechanism, the user's original data stream is decomposed into common and private stream components. The common stream is jointly decoded by all target MEC servers, while the private stream is dedicated to a specific server, thereby maximizing spectrum utilization while ensuring transmission reliability. Specifically, the user... n Transmitted to MEC server m The message data is Based on the technical principles of RSMA, message data is divided into common stream message parts. and private stream message section Common flow messages are collectively encoded into a single common flow. and with transmission power Uplink transmission is performed to each MEC server, with private messages encoded into their respective private streams. respectively by power The signal is transmitted to different servers via targeted transmission, so the user's transmitting signal can be represented as: (1) MEC Server m The receiving end receives data from the user. n The message is represented as: (2) in, Indicates user n With MEC server m Channel gain between The corresponding additive white Gaussian noise follows a pattern with a mean of zero and a variance of . The complex Gaussian distribution.

[0025] Accordingly, users n To each MEC server m Transmit common stream The achievable decoding rate can be expressed as: (3) Without loss of generality, assume that the user n Channel gain between MEC server and MEC server To ensure that each MEC server can successfully decode the common stream message, the common stream transmission rate must not exceed the decoding capacity of the weakest channel server within the time slot. t Internal allocation rate The following conditions should be met: (4) To ensure the successful implementation of SIC on the MEC server receiver, each user n The following conditions must be met: (5) in The decoding sensitivity threshold of the receiver's SIC is represented, and its physical meaning is the minimum power difference required for the receiver to effectively eliminate interference signals. Based on the channel gain ranking characteristics, the constraint condition shown in equation (5) can be simplified to: (6) After each server completes decoding of the common stream messages, each MEC server... m Decoding of dedicated private stream messages is performed independently, and the corresponding private stream decoding message rate is: (7) Last user n Transmitted to MEC server m The offloading rate is equal to the sum of the decoding rates of the public and private streams, and can be achieved in time slots. t The inner representation is as follows: (8) In the computing model, the computing tasks generated by the user terminal are lightweighted by the data compression module during the preprocessing stage, and then offloaded to the MEC node for distributed processing via the wireless channel. This processing mechanism builds a collaborative architecture of compression-offloading-computation.

[0026] The compression and decompression module consists of three stages: first, compression is performed at the user end; then, the compressed data is offloaded via the wireless channel; and finally, the base station's MEC server decompresses and restores the data. This system sets the ratio of the data size before and after compression as the data compression factor. ,in The minimum allowable compression ratio to avoid network data distortion. Specifically, the compression ratio is related to the computational overhead of compression; using an exponential computational complexity model, the number of CPU cycles required to compress one unit of data can be expressed as: (9) In the formula, This refers to the compression efficiency parameter in a data compression scheme, which the user selects when performing compression processing. This indicates the compression ratio selected in the task data. Then the compression latency on the user end can be expressed as: (10) in, Indicates user n The local CPU computing frequency. Assuming that all MEC servers use the same compression technology for decompression operations, the corresponding decompression time can be calculated as follows: (11) In the formula, Indicates MEC server m For the first n CPU computing resource frequency allocated to each user.

[0027] Under the MEC collaborative processing framework, each user n The computational tasks are distributed and offloaded to different MEC servers for processing. The offloading process includes three key stages: task transmission, task execution, and result return. The latency of the result return stage is ignored for two reasons: First, the amount of data returned by the computation results is usually much smaller than the amount of data in the original task, so its transmission latency has little impact on the overall latency; second, the downlink, thanks to the high-power transmission advantage of the base station, is significantly better than the uplink of the user equipment in terms of channel quality and anti-interference capability, enabling more efficient data transmission.

[0028] (1) First stage: The tasks that need to be offloaded to the MEC server in this stage can be divided into two parts: compressed data processing and uncompressed raw data. (In time slots) t Within this context, the transmission latency of raw data sent from the user terminal without compression to the MEC terminal is expressed as follows: (12) The transmission delay of the compressed data selected by the user terminal and sent to the MEC terminal is expressed as: (13) (2) Second stage: Each MEC server at the receiving end allocates CPU computing resources to process the offloading tasks of multiple users. It is assumed that there is no interaction dependency between the compressed and uncompressed data during the task offloading process. Therefore, the two parts of the task data can be executed independently. For users... n The original unloaded data, MEC server m In the time slot t The computation latency of an intra-task execution can be expressed as: (14) in, Indicates user n The computational task intensity reflects various service types, including voice, text, images, and video. For users... n Compressed unloaded data, MEC server m In the time slot t The computation latency of an intra-task execution can be expressed as: (15) This system proposes a computational task offloading scheme for RSMA-MEC systems based on data compression technology. For example... Figure 3 As shown, in the user n During the time slot process, users first compress some data locally, while uncompressed tasks are directly offloaded to the MEC server. After compression is complete, the compressed tasks are uploaded to the server for decompression and computation by the MEC. When the user... n After the unloading task is processed asynchronously on the MEC server, the MEC server will send the calculation results back to the user, forming a complete closed loop of computing services.

[0029] Therefore, users n via server m The total execution time for unloading and processing tasks is expressed as: (16) Based on the above model, for users n The maximum latency required for computing tasks on different MEC servers is: (17) The invention focuses on optimizing user uninstallation associations through joint optimization. and common stream message rate allocation Transmission power allocation Compression ratio Compression coefficient MEC server computing resource allocation The goal is to minimize the maximum latency experienced by all user computing tasks. Based on the above variable definitions, the problem of minimizing the maximum latency of user computing tasks can be formally expressed as: (18) in, Indicates user n Maximum transmission power, This represents the minimum allowable value for the uplink task offloading rate. Specifically, constraint C1 requires that the common stream transmission rate not exceed the decoding capacity of the weakest channel server, thus ensuring the reliability of multi-server collaborative decoding; constraint C2 ensures that the MEC server can successfully separate and decode the mixed signal in a predetermined order; constraint C3 limits the total transmit power of users to not exceeding the upper limit of their hardware-allowed power, reflecting the energy consumption limitations of the user end; constraint C4 ensures that the task is completed within the tolerable latency; constraint C5 ensures that the total computing resources allocated to each MEC server do not exceed its maximum available computing capacity; constraint C6, the normalization condition, ensures the fully distributed offloading of user computing tasks; constraint C7 ensures the service quality of the user's uplink; and constraints C8 and C9 together ensure that the optimization variables meet the value range.

[0030] The optimization problem (18) exhibits complex nonconvexity due to the multidimensional nonlinear coupling between the optimization variables, making it difficult to handle directly using traditional methods. Therefore, the DRL method based on the PPO algorithm is used to make decisions on the optimization variables. Joint optimization involves multiple variables in the RSMA-MEC system, and PPO is suitable for this highly dynamic, multi-objective scenario where user channel changes and task data changes occur.

[0031] In the optimization problem (18), the decisions regarding the computational task offloading association and the allocation of common stream message rates, transmit power allocation, compression ratio, compression coefficient and MEC server computational resources are only related to the current state of the system and are not related to the previous state. This is essentially in line with the definition of a typical Markov decision process, whose core elements mainly include three parts.

[0032] (1) State space: in each time slot t The system state space consists of multi-dimensional dynamic parameters, specifically including the set of task data. Channel gain set time slot t Remains unchanged, but every cycle T Updated set of user computational task strengths and user local computing frequency set To avoid gradient explosion or vanishing due to inconsistent data units, the state space is normalized during training: (19) Among them, the state space It can be represented as the union of the above four sets, that is , and These represent the upper and lower boundaries of the state space set, respectively.

[0033] (2) Action space: in each time slot t The action space is associated with the user's uninstallation. and common stream message rate allocation Transmission power allocation Compression ratio Compression coefficient MEC server computing resource allocation Composition. Therefore, the action space is defined as... These together constitute the multidimensional action decision parameters.

[0034] (3) Reward function: In the design of reward mechanisms for reinforcement learning, the reward function... Defined as the sum of the maximum negative latency of the user's computation task and the penalty term, the penalty terms for constraints C2 and C4 are defined as follows: (20) In each time slot t The reward function is defined as: (twenty one) Similarly, to facilitate the training process, the rewards are normalized: (twenty two) like Figure 4 As shown, the PPO algorithm is used to solve the Markov decision process and obtain the decision actions of the optimization variables.

[0035] To evaluate the performance of the adopted PPO resource allocation algorithm, three other algorithms were compared: (1) Actor-Critic (AC) Algorithm: The Actor-Critic algorithm combines value-based and policy-based methods. The target Q-value is calculated by the same network, and the maximum value among all current action value estimates is obtained through the Bellman equation each time. Neural networks tend to accumulate positive errors, leading to overestimation problems.

[0036] (2) DDPG: This algorithm combines deep learning and deterministic policy gradient methods to solve the continuous action space problem. The DDPG algorithm based on the actor-critic framework can achieve effective offline policy learning in complex high-dimensional environments, but it cannot avoid the value estimation error problem.

[0037] (3) SD3 algorithm: A reinforcement learning algorithm based on the actor-critic framework, which addresses the overestimation problem of the value function in DDPG by combining a double Q network with the Softmax operator. This algorithm utilizes the conservative estimation strategy of the double Q network to reduce the overestimation bias of action value, and balances exploration and exploitation by Softmax weighted fusion of the outputs of the two Q networks, thus exhibiting higher stability and convergence efficiency in complex continuous control tasks.

[0038] The simulation results are analyzed from the following multiple dimensions: from Figure 5 The results show the convergence performance of each algorithm within 5000 training cycles. Simulation results indicate that all algorithms converge with increasing training cycles. Overall, the PPO algorithm exhibits significant advantages during training, with its average reward value being significantly higher than other algorithms. This is primarily due to the PPO algorithm's extremely high stability and reliability, its wide applicability to discrete and continuous action spaces, its parameter insensitivity and ease of parameter tuning, its good convergence during training, its resistance to overfitting, and its ability to update at multiple time steps, further improving learning efficiency. Compared to the PPO algorithm, SD3 and DDPG are more sensitive to environmental changes, so PPO is widely used in large-scale reinforcement learning tasks, especially demonstrating superior performance in training complex tasks. After 5000 training cycles, the AC algorithm lags behind SD3 and DDPG algorithms in terms of average reward value, reflecting its limitations in handling complex tasks and the insufficient utilization of batch data samples by the online policy framework.

[0039] Figure 6 This demonstrates the relationship between the number of users and the maximum latency of user computing tasks. As the number of users gradually increases from 4 to 12, the maximum latency of user computing tasks within the system exhibits a non-linear increasing trend. This growth characteristic is due to the increased competition for computing resources on edge servers caused by the expansion of the number of users, while the computing frequency and parallel processing capabilities of MEC servers are limited. In the horizontal comparison of algorithms, PPO consistently maintains superior latency performance, and its more reasonable resource allocation as the number of users increases widens the latency gap compared to other algorithms. This is because the state space dimension expands exponentially with the number of users, and the shearing mechanism ensures stability.

[0040] from Figure 7The simulation results show a dynamic relationship between the maximum latency of user computing tasks and the computing frequency of the MEC server. Simulation results indicate that the maximum latency of user computing tasks gradually decreases as the maximum computing frequency of the MEC server increases. This trend is attributed to the fact that as the MEC server computing frequency increases, the task processing rate also increases, thereby shortening the task execution latency at the edge. Compared to other algorithms, the PPO algorithm consistently maintains the lowest average latency for user computing tasks when the MEC server computing frequency increases from 10 GHz to 50 GHz. This demonstrates that as the computing power of the MEC server increases, this algorithm exhibits the best performance advantage in latency reduction, primarily due to the synergistic optimization effect of its dynamic resource scheduling mechanism and the server computing frequency. According to the algorithm complexity analysis, the MEC server computing frequency is independent of the Markov decision process, thus maintaining stable efficiency even during server expansion.

[0041] like Figure 8 Simulation results show that this simulation experiment further conducts a quantitative analysis of the multi-server MEC offloading mechanism in large-scale system scenarios, focusing on the impact of the number of users and the number of MEC service nodes on computational latency performance. The total system bandwidth in the experimental settings... B Fixed at 20 MHz, user computing task data volume fixed at 1 Mbit, number of users N The latency varies within the range [4, 10]. As the number of users increases from 4 to 10, the maximum latency of user computing tasks gradually increases due to resource contention. In a horizontal comparison, the overall latency level is lower when multiple MEC servers are combined compared to a single MEC server configuration, indicating that increasing the number of servers can improve the parallel processing capability of tasks. Especially when the user scale exceeds 6, the latency reduction rate accelerates, demonstrating the advantages of multi-server architecture in high-concurrency scenarios. However, it is worth noting that when the number of users reaches a certain scale, the maximum latency of user tasks in different multi-MEC server solutions tends to be similar. This is because when the computing resources of MEC servers tend to be relatively sufficient, the system latency is mainly limited by the continuously increasing communication transmission latency or data compression overhead.

[0042] To thoroughly evaluate the advantages of RSMA-based multi-server MEC systems in terms of user latency performance, Figure 9 Showing the maximum power of different users Comparison of latency performance for the maximum computational task under certain conditions. Experimental data shows that, with... With parameter improvements, all multiple access techniques showed a trend of reduced latency, with the RSMA scheme consistently maintaining its performance lead. At high transmit power, the performance difference between RSMA and NOMA schemes in this algorithm became more significant, confirming the applicability of the proposed algorithm in high-power applications. Although RSMA technology exhibits a more significant performance improvement over NOMA under high-power conditions, its optimization in the low-power operating region still has engineering application value. RSMA expands the application boundaries of NOMA through a more flexible resource allocation mechanism.

[0043] like Figure 10 Simulation results show that, compared with the baseline fixed compression coefficient scheme, the proposed PPO adaptive compression coefficient scheme in this system results in the baseline scheme adjusting the compression coefficient... The compression coefficients were fixed at 0.5, 0.7, and 0.9, respectively. Numerical results show the relationship between task data volume and maximum latency for user computation under different compression coefficients. As the task data volume increases, the maximum latency gradually rises. However, the adaptive variable compression coefficient setting scheme in this system consistently maintains the highest computational performance, and the latency difference compared to the baseline fixed scheme widens as the task data volume increases. The reason can be analyzed from the trend in the graph. First, when the task data volume is relatively small (1 Mbit to 1.5 Mbit), a fixed compression coefficient of 0.7 is better than 0.5. Channel transmission resources are sufficient, and the task data does not need excessive compression, reducing compression processing overhead and avoiding data distortion. Then, when the task data volume exceeds 2 Mbit, a fixed compression coefficient of 0.5 is better than 0.7. This is because the processing data scale increases, and users need to increase the compression level to reduce data volume and lower transmission latency. Simulation results show that, compared to a fixed compression coefficient, the PPO adaptive compression coefficient scheme can better cope with changes in user communication and computing resources.

[0044] The system for resource allocation in RSMA-MEC systems based on data compression includes: a first processing unit for constructing a multi-user, multi-server RSMA-MEC system model; a second processing unit for determining a task processing latency model for the RSMA-MEC system based on data compression; a third processing unit for determining an optimization problem model based on the task processing latency model, the objective of which is to minimize the maximum latency for users to process tasks on different MEC servers; a fourth processing unit for obtaining a deep reinforcement learning framework based on the optimization problem model; a fifth processing unit for obtaining the optimal policy based on the PPO-based deep reinforcement learning optimization framework; and a sixth processing unit for allocating resources to the RSMA-MEC system based on the optimal policy.

[0045] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims.

Claims

1. A resource allocation method for an RSMA-MEC system based on data compression, characterized in that, The specific steps include: Step 1: Construct a system model that includes multiple user devices and multiple mobile edge computing (MEC) servers; Step 2: Establish an end-to-end latency model, comprehensively considering the data compression latency of the user locally, the task transmission latency through the RSMA link, and the data decompression latency and task computation latency of the MEC server. Step 3: Construct a joint optimization problem that minimizes the maximum latency; Step 4: Model the joint optimization problem as a Markov decision process; Step 5: Use the near-end policy optimization algorithm to solve the Markov decision process and output the optimal resource allocation strategy and joint optimization decision; Step 6: Allocate resources between user devices and MEC servers in the system.

2. The RSMA-MEC system resource allocation method based on data compression according to claim 1, characterized in that: In step two, the latency model includes data compression latency and task transmission latency on the user side, as well as data decompression latency and task computation latency on the MEC server side.

3. The RSMA-MEC system resource allocation method based on data compression according to claim 1, characterized in that: In step three, the variables for joint optimization include user task offloading association, transmission power of public and private messages in RSMA communication, public message rate allocation, compression ratio of task data, compression coefficient of task data, and computing resource allocation of MEC server.

4. The RSMA-MEC system resource allocation method based on data compression according to claim 1, characterized in that: In step four, the joint optimization problem is modeled as a Markov decision process, which specifically includes defining a state space, including the amount of task data for all users, the channel gain between all user devices and all MEC servers, the intensity of user computational tasks, and the user's local computation frequency; defining an action space, whose elements are the set of all variables in the joint optimization; and defining a reward function, which is related to the negative value of the maximum latency of the computational tasks of all users and includes a penalty term for violating constraints.

5. The RSMA-MEC system resource allocation method based on data compression according to claim 1, characterized in that: Step five specifically involves using an Actor-Critic network architecture, where the Actor network selects actions based on the current state, and the Critic network evaluates the value of state-action pairs; collecting experience data through interaction with the environment and storing it in an experience replay buffer; sampling small batches of data from the experience replay buffer and calculating the advantage function; updating the parameters of the Actor network by maximizing the alternative objective function that includes a shearing mechanism; and updating the parameters of the Critic network by minimizing the mean squared error of the value function.

6. The system of the RSMA-MEC system resource allocation method based on data compression according to any one of claims 1-5, characterized in that: A first processing unit is provided for building a multi-user, multi-server RSMA-MEC system model; The second processing unit is used to determine the task processing latency model of the RSMA-MEC system based on data compression; The third processing unit is used to determine the optimization problem model based on the task processing latency model. The goal of the optimization problem model is to minimize the maximum latency for users to process tasks on different MEC servers. The fourth processing unit is used to obtain a deep reinforcement learning framework based on the optimization problem model; The fifth processing unit is used to obtain the optimal policy based on the PPO-based deep reinforcement learning optimization framework; The sixth processing unit is used to allocate resources to the RSMA-MEC system based on the optimal strategy.

Images