Ris-assisted service caching and resource allocation method based on double-time-scale optimization

Abstract

This invention discloses a RIS-assisted service caching and resource allocation method based on dual-timescale optimization, belonging to the field of vehicle-mounted edge computing technology. Addressing the challenges of co-optimizing service caching and resource scheduling under conditions of fluctuating wireless link communication quality, dynamically changing service requests, and limited edge resources, this invention constructs a RIS-assisted end-edge-cloud collaborative computing system and establishes a joint optimization model for service caching, communication transmission, and computing resources. On a large timescale, a service caching strategy based on LSTM-DDPG is adopted to dynamically adjust the cache content of edge nodes according to historical service request information. On a small timescale, an improved genetic algorithm is used to jointly optimize task offloading decisions, RIS phase shift configuration, bandwidth resource allocation, and computing resource allocation. Through a dual-timescale collaborative decision-making mechanism (TT-LDGA), coordinated control of long-term cache optimization and short-term resource scheduling is achieved, improving system resource utilization and service quality while reducing cache switching overhead and latency energy consumption costs.

Description

Technical Field

[0001] This invention relates to the field of vehicle edge computing technology, and in particular to a reconfigurable intelligent reflective surface auxiliary service caching and resource allocation method based on dual time scale optimization. Background Technology

[0002] With the rapid development of intelligent transportation systems, autonomous driving, and vehicle-to-everything (V2X) technologies, the amount of data generated by vehicles and their computing demands are constantly increasing. Vehicular edge computing (VEC) deploys computing and storage resources at the network edge, enabling computing tasks to be processed closer to the vehicle. This effectively reduces task processing latency and improves service response efficiency, and has become an important technical means to support intelligent transportation applications.

[0003] In real-world vehicle environments, the quality of wireless communication links between vehicles and edge servers fluctuates significantly due to factors such as high-speed vehicle movement, complex road structures, and building obstructions. This can easily lead to decreased data transmission rates and reduced task processing performance. In recent years, Reconfigurable Intelligent Surfaces (RIS) have emerged as a new technological approach to improving vehicle communication quality by intelligently controlling the wireless propagation environment and adjusting the phase of the reflective elements to change the wireless signal propagation path.

[0004] Meanwhile, in-vehicle applications typically rely on specific service modules such as object detection, high-precision map processing, and route planning. Since edge servers have limited storage resources, they cannot cache all services, thus requiring dynamic adjustments to cached content based on service demands. However, service caching decisions usually have long update cycles, while task offloading, RIS configuration, and resource allocation decisions need to be rapidly adjusted based on real-time network conditions; the two exhibit significantly different timescale characteristics.

[0005] Most existing studies treat service caching optimization, task offloading, and resource allocation optimization separately, or use a unified time scale for joint optimization, failing to fully consider the differences in time scale between service caching decisions and resource scheduling decisions. Furthermore, the service cache state directly affects task execution location and resource allocation results, while task offloading and resource scheduling affect system service quality, indicating a strong coupling between the two. Therefore, in a RIS-assisted vehicle-mounted edge computing environment, how to achieve coordinated optimization of service caching, task offloading, RIS configuration, and resource allocation to improve system service quality and reduce cache switching overhead remains a pressing technical problem to be solved. Summary of the Invention

[0006] To address the limitations of existing RIS-assisted vehicle edge computing systems, such as limited service cache capacity, high coupling between service cache and resource scheduling, and time-scale differences in various decisions, this invention provides a RIS-assisted service cache and resource allocation method based on dual-time-scale optimization. This method constructs a collaborative optimization framework for service cache and resource scheduling, reducing service cache switching overhead while ensuring system service quality, thereby achieving long-term performance optimization of the vehicle edge computing system.

[0007] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A RIS auxiliary service caching and resource allocation method based on dual time scale optimization, comprising the following steps:

[0008] Step 1: Construct a RIS-assisted in-vehicle edge computing system;

[0009] Step 2: Based on the vehicle edge computing system architecture built in Step 1, build the service cache model, wireless communication model, and computing model respectively, and fully define the variables, constraints, overhead / latency / energy consumption calculation formulas of each model to obtain all the basic quantitative constraints and calculation formulas of the system.

[0010] Step 3: Based on the three types of models and all formulas output in Step 2, construct the multivariate joint optimization objective function and complete constraint set to obtain the NP-hard mixed integer nonlinear programming optimization problem P1;

[0011] Step 4: Based on the optimization problem obtained in Step 3 and the service popularity / cache gain / cache state data output in Step 2, the LSTM-DDPG algorithm is used to generate static cache decisions on a large time scale and update the cache content of the edge server.

[0012] Step 5: Based on the fixed cache state output in Step 4, collect multi-dimensional environmental states in real time at each independent small time scale to obtain small-scale optimized complete input parameters;

[0013] Step 6: Based on the real-time environmental parameters collected in Step 5, the fixed cache state in Step 4, and the joint optimization model constructed in Step 3, an improved genetic algorithm is used to jointly optimize offloading, RIS phase shift, bandwidth, and computing resources, and output the optimal real-time resource scheduling scheme.

[0014] Step 7: Based on the resource scheduling scheme output in Step 6, execute full vehicle task processing, synchronously update the real-time status of the entire domain, and output the updated environmental data at a small time scale.

[0015] Step 8: Loop judgment and dual time scale collaborative iteration, based on the complete time sequence process of steps 4-7 to achieve long-term collaborative optimization.

[0016] In some possible implementations, step 1 specifically involves:

[0017] Step 1.1: Deploy N vehicle users, a base station equipped with M MEC edge servers, 1 cloud server, and a RIS containing Z reflection units;

[0018] Step 1.2, define the link connection relationship: vehicle and base station wireless direct connection, base station and cloud server fiber optic backhaul, RIS deployed on building surface to reflect vehicle wireless signal;

[0019] Step 1.3, set the business logic: the vehicle generates F type computing tasks, each type of task corresponds to a dedicated service model, the cloud server stores all services, the edge server is limited by storage and only caches some services, and outputs the overall system topology.

[0020] Step 1 establishes an edge-cloud collaborative vehicle edge computing architecture that includes vehicles, edge servers, cloud servers, and multi-reflection units (RIS). Relying on the independent adjustable reflection units of the RIS, the wireless propagation link is actively reshaped. At the same time, multi-level task processing modes are divided into local, edge, and cloud. This architecture can alleviate the problem of severe fluctuations in communication links caused by high-speed vehicle movement and building obstruction, and improve the transmission rate and stability of vehicle-side communication. On the other hand, it provides a complete hardware operating platform for subsequent layered cache deployment, multi-mode task offloading, and RIS phase joint optimization, and lays the physical foundation for multi-dimensional collaborative optimization of storage, computing, and communication.

[0021] In some possible implementations, step 2 specifically involves:

[0022] Step 2.1, Construct the service caching model: Define binary cache state variables Characterize whether edge server m caches service f; set cache capacity constraints; construct a formula for calculating cache switching costs:

[0023] ;

[0024] in This is an indicator function that takes the value 1 when the condition x is true, and 0 otherwise. This indicates that the service needs to be obtained from a cloud server. For a fixed backhaul link transmission rate; output cache storage limit and cache update overhead quantification rules;

[0025] Step 2.2, Construct the wireless communication model: Integrate the vehicle-base station direct link and the RIS reflection-assisted link, and derive the small time scale time slots. Formula for signal-to-interference-plus-noise ratio between VU n and BS:

[0026] ;

[0027] in For the transmission power of VU n, It is additive white Gaussian noise. For the channel gain between RIS and BS, For the channel gain from VU n to RIS The channel gain between VU n and BS, Here is the diagonal phase shift matrix of RIS, and the phase shift... Set hardware constraints for discrete quantization phase shift:

[0028] ;

[0029] in To control the quantization accuracy of phase shift; output wireless transmission rate, channel quality, and RIS phase value constraints;

[0030] Step 2.3, Constructing the computational model: Defining unloading decision variables Calculate resource allocation variables Set preconditions for task execution:

[0031] ;

[0032] in Indicates VU n in Unload the task to MEC server m. Indicates local computation. This section describes cloud computing; it categorizes task execution modes into four types: local, single-edge, multi-edge collaboration, and cloud; and derives a unified latency formula for these multiple modes.

[0033] ;

[0034] in This indicates the time required for the task to execute locally. and This indicates the transmission and processing time required for the task to execute on the MEC server m. , and This indicates that MEC server m does not have the services required for task processing, and these services are handled by the cooperating MEC server. The required transmission and processing time for execution This indicates that all MEC servers do not cache the time required for the corresponding service to be processed by the cloud server. The formula for calculating the total energy consumption of the build task is as follows:

[0035] ;

[0036] in This represents the energy consumption of transferring the task to the MEC server m. This represents the energy consumption required for local processing; it defines the QoS quantification formula for a single vehicle.

[0037] ;

[0038] in These are the weighted parameters in QoS that characterize the VU's preference for latency and energy consumption. To ensure the normal operation of VU n, the latency and energy consumption thresholds are set. It provides complete quantitative rules for task processing latency, including task unloading, resource allocation, latency and energy consumption, and user service quality.

[0039] Step 2 constructs a cache model that includes cache capacity constraints and switching overhead calculation formulas, a communication model that integrates direct connection and RIS reflection dual links and limits the discrete phase values ​​of RIS, and a calculation model that binds cache pre-constraints and uniformly quantifies latency, energy consumption, and user QoS indicators for multiple execution modes. This transforms all resource boundaries, overhead benefits, and variable coupling relationships of the real physical system into standardized mathematical expressions. It can accurately quantify the cost loss brought by cache updates and the channel gain brought by communication phase adjustments, unify latency and energy consumption evaluation standards, and clearly establish strong coupling constraints between cache and task offloading, eliminating the problem of fragmented multi-module optimization. This provides quantifiable and solvable complete mathematical support for subsequent multi-objective joint optimization.

[0040] In some possible implementations, step 3 specifically involves:

[0041] Step 3.1: Construct the long-term optimization objective function P1 and introduce weighting factors. Balancing QoS benefits with cache switching overhead:

[0042] ;

[0043] Step 3.2, superimpose all constraints C1-C13:

[0044]

[0045] Constraint C1 states that a task can only be processed when the MEC server has cached the corresponding service model and allocated the necessary computing resources; constraint C2 states that the bandwidth of the MEC server is only allocated to the VU performing task offloading; constraint C3 states that the task generated by VU n can only be processed by a single server; constraints C4, C5, and C6 represent the cache capacity, computing power, and bandwidth resource limits of each MEC server, respectively; constraints C7 and C8 are the upper limits of latency and energy consumption of VU; constraints C9-C13 make this problem a mixed-integer nonlinear programming problem.

[0046] Step 3.3 outputs an NP-hard joint optimization problem containing coupled variables of long and short time scales and a mixture of discrete and continuous variables.

[0047] Step 3 uses long-term average system QoS as the optimization benefit and cache switching overhead as the loss term, and introduces weighting factors to balance the dual objectives. It integrates all discrete and continuous variables such as cache, offloading, bandwidth, computing, and RIS phase shift, and superimposes constraints of hardware, resources, and user experience to construct a mixed integer nonlinear long-term optimization problem. Based on this unified multi-objective optimization framework, service experience and cache update cost can be balanced simultaneously, avoiding the overall performance loss caused by single index optimization. It fully realizes four-dimensional collaborative modeling of storage, communication, computing, and RIS resources. The complete set of constraints ensures that the subsequent algorithm output scheduling scheme has engineering feasibility. At the same time, it is more in line with the dynamic changes of the Internet of Vehicles for infinite long-term time-series optimization, focusing on the optimal steady-state comprehensive performance of the system rather than the short-term benefits of a single time slot.

[0048] In some possible implementations, step 4 specifically involves:

[0049] Step 4.1: Count the frequency of each service request within the historical time window to obtain the service popularity. The QoS improvement brought by quantification caching is used to obtain the service cache gain. Read the current edge server cache state. Combining them yields a single small-scale state. ;

[0050] Step 4.2: Generate the state sequence of all K hourly intervals within a large period. Input an LSTM network to extract temporal correlation features of service requests, and output a unified state over a large time scale. ;

[0051] Step 4.3, convert the time-series characteristic states Input the DDPG deep reinforcement learning network and output a binary decision for the service cache of each edge server. ;

[0052] Step 4.4: Update the storage content of the edge server according to the caching decision. This caching configuration will not change within the current complete large time scale period. The output fixed cache state is used as the input of the pre-constraint for the optimization of the lower small time scale.

[0053] Step 4 leverages the longer cache update cycle and the temporal correlation of service requests to extract multi-round, small-scale historical request temporal features using LSTM, compensating for the lack of temporal awareness in traditional reinforcement learning. DDPG then outputs cache decisions for each edge server, with cache configuration remaining fixed throughout the entire cycle and serving as static input for lower-level resource scheduling. This decoupling of long and short time scales significantly reduces the overall optimization complexity. By mining temporal features, it accurately predicts changes in service popularity and improves cache hit rate. Simultaneously, the long-term stability of the cache state reduces backhaul traffic and download latency caused by frequent updates. Reinforcement learning autonomously iterates and adapts to dynamically changing vehicle requests, enhancing the system's adaptability to time-varying business scenarios.

[0054] In some possible implementations, step 5 specifically involves:

[0055] Step 5.1, collect data from the vehicle side: the current task type, task data volume, and local computing power of each vehicle;

[0056] Step 5.2, collect communication side data: vehicle-base station direct connection channel Vehicle-RIS Channel RIS - Base Station Channel Noise power ;

[0057] Step 5.3, Acquire phase data from the RIS hardware side: Current phase of each reflection unit. Quantization accuracy ;

[0058] Step 5.4: Collect data from the edge server side: remaining storage, remaining computing power, and remaining bandwidth resources;

[0059] Step 5.6: Integrate all real-time state parameters and output a small-time-scale optimized input set.

[0060] Step 5 synchronously collects five types of real-time dynamic information within a high-frequency, short-term period: vehicle task, wireless channel, RIS hardware, upper-layer cache, and remaining edge resources. This fully captures the instantaneous channel fluctuations and sudden task demands caused by vehicle movement. Simultaneously accessing the fixed cache status of large-scale output establishes a dual-layer optimized data link. This ensures that short-term resource scheduling adapts to the real-time changes in the vehicle scenario and strictly adheres to cache pre-constraints to avoid ineffective scheduling schemes. The synchronous collection of the RIS real-time hardware status also provides accurate real-time data support for timely adjustment of reflection phase and compensation for link attenuation.

[0061] In some possible implementations, step 6 specifically involves:

[0062] Step 6.1, Hybrid Unified Chromosome Encoding: Discrete real number encoding is used for discrete variable task offloading and RIS phase shifting, while continuous real number encoding is used for continuous variable computational and communication resource allocation. The chromosome of the j-th individual in the l-th generation is... Each chromosome corresponds to a set of candidate scheduling solutions;

[0063] Step 6.2, Construct a fitness function with constraints and penalties: Using the total QoS of the system as the optimization benchmark, penalties are imposed on individuals that violate constraints C1-C13. The piecewise fitness formula is as follows:

[0064] ;

[0065] Step 6.3, Improve the genetic iteration operation:

[0066] ① Selection: Parent individuals are selected based on fitness ratio, with the following selection probability formula:

[0067]

[0068] in Represents the set of all individuals in the l-th generation;

[0069] ② Adaptive crossover: The crossover probability is dynamically adjusted based on the overall fitness of the population; mean crossover of continuous gene blocks and exchange within discrete gene blocks are performed, while preserving the coupling relationship between variables;

[0070] ③ Adaptive mutation: dynamically adjusts mutation probability; discrete gene integer random mutation, continuous gene Gaussian mutation;

[0071] ④ Elite retention mechanism: The best individual is retained in each generation to avoid losing the best solution;

[0072] Step 6.4: After the genetic iteration converges, output the current optimal scheduling scheme for the small time scale: task unloading matrix. RIS Discrete Phase Shift Matrix Bandwidth allocation Computing resource allocation .

[0073] Step 6 designs a hybrid encoding to uniformly represent discrete offloading, RIS phase variables, and continuous bandwidth and computing power allocation variables. It constructs a fitness function with a penalty term to screen out infeasible solutions that violate various constraints. Combined with adaptive crossover mutation probability and elite retention mechanism, it balances the global exploration and local convergence capabilities of the algorithm, and completes multi-dimensional resource joint optimization in one go. This improved mechanism can solve the defects of premature convergence and performance loss of step-by-step optimization in traditional genetic algorithms. It efficiently solves NP-hard mixed integer optimization problems, synchronously and dynamically adjusts the RIS phase to continuously optimize the wireless transmission link, and synchronously optimizes the offloading position, spectrum and computing power allocation in a single time slot. It effectively reduces transmission latency and terminal power consumption, and outputs a feasible scheduling strategy that meets all hardware and user performance thresholds.

[0074] In some possible implementations, step 7 specifically involves:

[0075] Step 7.1, based on the unloading decision Assign task execution location, matching edge server cache validation constraint C1;

[0076] Step 7.2, based on RIS phase shift Adjusting the phase of all reflective units and reconstructing the wireless propagation link improves performance. ;

[0077] Step 7.3, based on bandwidth Computing resources Allocate computing power and spectrum resources for edge servers;

[0078] Step 7.4: Calculate the total task delay for each vehicle in different modes. Total energy consumption Verify threshold constraints C7 and C8;

[0079] Step 7.5: After the task is completed, update the vehicle task queue, real-time channel gain, remaining edge storage / computing power / bandwidth, and output the updated environment state set.

[0080] Step 7 completes task allocation, RIS phase adjustment, and resource allocation according to the optimized resource scheduling scheme. It verifies latency and energy consumption thresholds according to execution mode to ensure the basic performance of vehicle services. After the task is processed, it synchronously updates the global status of vehicle tasks, channels, and remaining edge resources and sends it back to the status acquisition stage, forming a complete closed loop of "status acquisition - optimization solution - scheme execution - status update". It directly implements the optimal scheduling strategy to improve the service quality of vehicle users. The real-time updated environmental data continuously adapts to the dynamic changes in channels and resources brought about by vehicle movement, ensuring that the latency and energy consumption hard indicators of services such as autonomous driving and high-precision maps do not exceed the limits throughout the process.

[0081] In some possible implementations, step 8 specifically involves:

[0082] Step 8.1: Determine whether the current time series has completed a full large timescale cycle;

[0083] Step 8.2, Period not reached: Jump to step 5, collect a new round of real-time status, and repeat steps 6 and 7 to complete small-scale scheduling;

[0084] Step 8.3, Reaching the cycle: Jump to step 4, retrain LSTM-DDPG based on the accumulated small-scale time-series states to generate a new cache configuration, and then execute steps 5, 6, and 7 in sequence;

[0085] Step 8.4 involves continuous cyclical two-layer timing optimization to maximize the system's average QoS and minimize the total cost of cache switching in the long term, outputting a continuously dynamically optimized operating strategy for the vehicle edge computing system.

[0086] Step 8 sets up a nested loop logic with long and short cycles. When the large-scale update cycle has not been reached, small-scale real-time resource scheduling is continuously executed. After the cycle expires, LSTM-DDPG is run again to update the global cache strategy. The updated cache state continues to serve as a fixed input for subsequent rounds of small-scale optimization, forming a two-layer collaborative optimization closed loop of long-term cache planning and short-term real-time scheduling. This balances the long-term steady-state benefits of the system with short-term dynamic environment adaptation, continuously iterating to reduce the overall system latency, energy consumption, and cache switching costs. It fully coordinates storage, communication, computing, and RIS reflection resources, fully explores the performance limits of RIS-assisted vehicle edge computing, and significantly improves the system's resource utilization and environmental adaptability in complex scenarios such as high-speed movement, time-varying requests, and resource constraints.

[0087] The beneficial effects of this invention are:

[0088] (1) A dual-time-scale optimization framework is proposed, which optimizes service caching decisions and resource scheduling decisions at different time scales, making full use of the time characteristics of the two types of decisions and reducing the complexity of solving the joint optimization problem;

[0089] (2) Establish a collaborative optimization mechanism for service caching, task unloading, RIS configuration and resource allocation to realize multi-dimensional resource joint scheduling and improve the overall resource utilization efficiency of the system;

[0090] (3) By dynamically adjusting the wireless propagation environment through RIS, the quality of the communication link between the vehicle and the edge server can be improved, thereby increasing data transmission efficiency and communication reliability;

[0091] (4) Use the LSTM-DDPG algorithm to learn the service request change pattern, realize the dynamic adjustment of service cache content, improve cache hit rate and reduce cache update overhead;

[0092] (5) An improved genetic algorithm is used to jointly solve the problems of task unloading, RIS configuration and resource allocation, thereby improving the solution efficiency and global search capability of complex optimization problems;

[0093] (6) By optimizing long-term cache and coordinating short-term resource scheduling, the system service quality is improved, the task processing latency, energy consumption and cache switching costs are reduced, and the system's adaptability to dynamic vehicle network environment is enhanced. Attached Figure Description

[0094] Figure 1 This is a system model diagram of the RIS auxiliary service caching and resource allocation method based on dual time scale optimization of the present invention.

[0095] Figure 2 This is a flowchart of the RIS auxiliary service caching and resource allocation method based on dual time scale optimization of the present invention.

[0096] Figure 3 This is a framework diagram of the TT-LDGA invention.

[0097] Figure 4 This is a flowchart of the collaborative unloading and resource allocation process based on an improved genetic algorithm at a small time scale according to the present invention.

[0098] Figure 5 This is a graph showing the cumulative reward convergence performance of different numbers of vehicles during the training process in a specific embodiment of the present invention.

[0099] Figure 6 This is a schematic diagram illustrating the impact of various methods on the short-term system performance of the proposed model system under different total bandwidths in specific embodiments of the present invention.

[0100] Figure 7 This is a schematic diagram illustrating the impact of various methods on the long-term system performance of the proposed model system under different MEC server caching capabilities in specific embodiments of the present invention. Detailed Implementation

[0101] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0102] This invention provides a RIS auxiliary service caching and resource allocation method based on dual time scale optimization, comprising the following steps:

[0103] Step S1: Construct a RIS-assisted vehicle edge computing system, which includes vehicle nodes, edge servers, cloud servers, and a reconfigurable intelligent reflective surface (RIS). The vehicle nodes generate task requests, the edge servers provide service caching and computing resources, the cloud servers provide remote service support, and the RIS is used to regulate the wireless propagation environment.

[0104] Step S2: Establish an in-vehicle edge computing system model, including a service cache model, a wireless communication model, and a computing model. The service cache model describes the edge server's cache capacity constraints and cache switching overhead; the wireless communication model describes the data transmission process between the vehicle and the edge server; and the computing model describes the latency and energy consumption during task execution. The service cache model includes service cache status, cache capacity constraints, and cache switching cost constraints. The cache capacity constraint limits the total cacheable service capacity in the edge server to no more than its maximum storage capacity, and the cache switching cost characterizes the overhead caused by changes in service cache content within adjacent large timescale periods.

[0105] Step S3: Construct a joint optimization model with the goal of maximizing system service quality and minimizing cache switching costs.

[0106] Step S4: Within each large timescale, a service caching decision is generated based on the current service cache state, service cache gain, and service popularity information of each edge server, and the cached content of the edge servers is updated. The service caching decision remains unchanged within the corresponding large timescale period. Service popularity information is obtained by statistically analyzing the frequency of requests for each service within a historical time window. Service cache gain is used to characterize the degree of improvement in system service quality brought about by caching services. The service caching decision is generated jointly based on service popularity, service cache gain, and the current cache state. The service caching decision is generated using a learning method combining a Long Short-Term Memory (LSTM) network and a Deep Deterministic Policy Gradient (DDPG). LSTM is used to extract time-related features from historical service request sequences, and DDPG is used to output service caching actions based on the system state, thereby determining the service cache content of each edge server.

[0107] Step S5: Obtain the vehicle task request status, wireless channel status, RIS status, service cache status, and edge resource status within each small time scale.

[0108] Step S6: Based on the current service cache state, using task offloading decision, RIS phase shift configuration, communication resource allocation, and computing resource allocation as optimization variables, perform joint optimization according to the joint optimization model to obtain a resource scheduling scheme. The RIS includes multiple independently adjustable reflection units. Each reflection unit adjusts the reflection phase to change the wireless signal propagation path, thereby enhancing the communication link quality between the vehicle node and the edge server. The joint optimization model uses task offloading decision, RIS phase shift configuration, bandwidth resource allocation, and computing resource allocation as optimization variables, and simultaneously satisfies cache capacity constraints, communication resource constraints, computing resource constraints, and RIS phase shift constraints. The joint optimization problem is solved using an improved genetic algorithm. This improved genetic algorithm uses a joint encoding method to uniformly encode the task offloading decision, RIS phase shift configuration, communication resource allocation, and computing resource allocation, and obtains the resource scheduling scheme through genetic iteration. The improved genetic algorithm dynamically adjusts the crossover probability and mutation probability according to the population fitness and uses an elite retention mechanism to save the current best individual, thereby improving global search capability and algorithm convergence performance.

[0109] Step S7: Execute task processing according to the resource scheduling plan, and update the vehicle task status, communication status and resource status.

[0110] Step S8: After the preset large time scale decision period is reached, re-execute steps S4 to S7 to achieve dual time scale collaborative optimization of service caching decision and resource scheduling decision;

[0111] In this system, the decision cycle on a larger timescale is longer than that on a smaller timescale, and the service cache decision serves as the system state input for subsequent resource scheduling processes on multiple smaller timescales. The larger timescale is used for service cache decision updates, while the smaller timescales are used for task offloading decisions, RIS phase shift configuration, communication resource allocation, and computing resource allocation. Within the same larger timescale period, the service cache state remains unchanged and serves as the system state input for multiple consecutive resource scheduling processes on smaller timescales. The service cache state generated on the larger timescale serves as the input state for the resource scheduling optimization process on smaller timescales, determining whether a task can be executed on the edge server and the corresponding resource allocation strategy, thereby achieving collaborative optimization of service cache decisions and resource scheduling decisions.

[0112] Example

[0113] The system model diagram of the present invention is as follows: Figure 1As shown, the system consists of a cloud server, a roadside base station (BS), a RIS (Radio Reflector System) installed on the surface of a nearby building, and N vehicle users (VUs). The BS is equipped with M MEC (Multi-access Edge Computing) servers, each with independent computing and storage resources. The BS communicates with the vehicles wirelessly and is connected to the cloud server via fiber optic links. The RIS, deployed on the building surface, contains Z source-reflecting units. Each vehicle user continuously generates computationally intensive tasks during its journey, assuming F different types of computational tasks can be generated, each corresponding to a specific service model. The cloud server pre-configures all services, while the MEC server, due to limited caching resources, can only store a portion of the services.

[0114] This invention provides a RIS auxiliary service caching and resource allocation method based on dual-time-scale optimization, such as... Figure 2 As shown, it includes the following steps:

[0115] Step S1: Construct a RIS-assisted vehicle edge computing system. The system includes N vehicle nodes, a base station with M edge servers, one cloud server, and a RIS with Z source reflection units. Among them, the vehicle nodes generate task requests, the edge servers provide service caching and computing resources, the cloud server provides remote service support, and the RIS is used to regulate the wireless propagation environment.

[0116] Step S2: Establish an in-vehicle edge computing system model, including a service cache model, a wireless communication model, and a computing model.

[0117] The service caching model describes the edge server's cache capacity constraints and cache switching overhead. All caching decisions remain unchanged over a large timescale. The total size of the cached services cannot exceed the MEC server's cache capacity limit. The cache switching cost is the communication latency cost incurred by the download service.

[0118]

[0119] in: This is an indicator function that takes the value 1 when the condition x is true, and 0 otherwise. This indicates that the service needs to be obtained from a cloud server. The backhaul link transmission rate is fixed. The wireless communication model describes the data transmission process between the vehicle and the edge server, considering the coexistence of direct communication links and RIS-assisted communication links. Therefore, small timescale slots... The signal-to-interference-plus-noise ratio (SIR) between VU n and BS can be expressed as:

[0120]

[0121] In the formula For the transmission power of VU n, It is additive white Gaussian noise. For the channel gain between RIS and BS, For the channel gain from VU n to RIS The channel gain between VU n and BS, The diagonal phase shift matrix of the RIS is subject to hardware constraints, and the phase shift... Only from a finite set of discrete values Select from, that is:

[0122]

[0123] in: To control the quantization accuracy of phase shifts, a computational model is used to describe the latency and energy consumption during task execution. The VU's computation modes are divided into local computation, associated MEC server computation, collaborative MEC server computation, and cloud computing. It is particularly important to note that nodes can only process tasks if the corresponding service model has been cached and the necessary computing resources have been allocated; that is:

[0124]

[0125] in: for The collaborative unloading decision variable of VU n at time, Indicates VU n in Unload the task to MEC server m. Indicates local computation. It refers to cloud computing. For computing resource allocation decisions, This indicates whether MEC server m caches service f. QoS is defined as the degree of satisfaction of VU n with latency and energy consumption relative to a set threshold, and can be expressed as:

[0126]

[0127] in: These are the weighted parameters in QoS that characterize the VU's preference for latency and energy consumption. To ensure the normal operation of VU n, the latency and energy consumption thresholds are set. Task processing latency can be uniformly represented as follows, under different circumstances:

[0128] ;

[0129] in: This indicates the time required for the task to execute locally. and This indicates the transmission and processing time required for the task to execute on the MEC server m. , and This indicates that MEC server m does not have the services required for task processing, and these services are handled by the cooperating MEC server. The required transmission and processing time for execution This indicates that none of the MEC servers have cached the time required for the corresponding service to be processed by the cloud server. Energy consumption for task processing can be uniformly expressed as follows under different conditions:

[0130] ;

[0131] in: This represents the energy consumption of transferring the task to the MEC server m. This indicates the energy consumption required for local processing.

[0132] Step S3: Construct a joint optimization model with the objectives of maximizing system service quality and minimizing cache switching cost, which can be expressed as:

[0133]

[0134] in: To balance the weighting factors of cache switching costs and vehicle QoS, constraints C1 state that tasks can only be processed when the MEC server caches the corresponding service model and allocates the necessary computing resources; constraint C2 states that the MEC server's bandwidth is only allocated to VUs performing task offloading; constraint C3 states that tasks generated by VU n can only be processed by a single server; constraints C4, C5, and C6 represent the cache capacity, computing power, and bandwidth resource limitations of each MEC server, respectively; constraints C7 and C8 are the latency and energy consumption limits of VUs. Constraints C9-C13 make this problem a mixed-integer nonlinear programming problem, involving variables at different time scales, and is essentially an NP-hard problem. Its core challenge lies in the difficulty of modeling the temporal relationship between service caching and task offloading. The scale and complexity of this problem make it difficult to solve using traditional optimization methods.

[0135] Step S4: Within each large timescale, record the state of each small timescale. This includes the service cache status of each edge server. Service caching gain and service popularity information Transformed into a large timescale state LSTM-DDPG is used to generate service caching decisions and update the cached content of edge servers; the service caching decisions remain unchanged over the corresponding large time scale period.

[0136] The flowchart of the TT-LDGA algorithm is as follows: Figure 3 As shown, TT-LDGA employs a two-layer time-scale optimization structure. The upper layer performs service caching decisions based on LSTM-DDPG at a large time scale, while the lower layer performs "cooperative offloading and resource allocation" decisions based on an improved genetic algorithm at a small time scale. The two layers collaboratively optimize the low service cache update frequency and the rapid dynamic changes in task offloading and resource allocation, respectively. Within each large time-scale time slot... Initially, the LSTM-DDPG agent observes the environmental state across multiple consecutive small timescale slots and inputs it into the LSTM network to extract temporal features, generating a state representation for the current large timescale slot. This representation is then input into the DDPG network to obtain service caching decisions. The figure shows a TT-LDGA framework diagram where a large timescale slot contains three small timescale slots. After obtaining the caching decision, the lower-level improved genetic algorithm performs "cooperative unloading and resource allocation" decisions based on the current cache configuration conditions within each small timescale slot, updating the environmental state in real time. Simultaneously, the upper layer continuously collects state information from small timescales, providing a basis for service caching decisions in the next large timescale slot.

[0137] Step S5: Obtain the vehicle task request status, wireless channel status, RIS status, service cache status, and edge resource status within each small time scale; based on the current service cache status, use task offloading decision, RIS phase shift configuration, communication resource allocation, and computing resource allocation as optimization variables, and perform joint optimization using an improved genetic algorithm according to the joint optimization model to obtain a resource scheduling scheme.

[0138] Figure 4 This is a flowchart of collaborative unloading and resource allocation based on an improved genetic algorithm at a small time scale.

[0139] (1) Chromosomes and mixed coding

[0140] Since joint optimization problems involve both discrete and continuous variables, traditional single encoding methods are insufficient to uniformly represent different types of decision variables. Therefore, a hybrid encoding mechanism is designed to uniformly represent task offloading decisions, RIS phase shift decisions, computational resource allocation, and communication resource allocation. Specifically, task offloading and RIS phase shifts use discrete real-number encoding, while computational and communication resource allocation use continuous real-number encoding to satisfy constraints C10-C13. Each chromosome corresponds to a candidate solution to the problem; let the chromosome of the j-th individual in the l-th generation be... .

[0141] (2) Fitness function

[0142] The fitness function is used to evaluate the merits of the "cooperative offloading and resource allocation" method; a higher fitness value indicates better system performance. The fitness function is constructed based on the service quality optimization objective, taking into account constraints such as service cache capacity, computing resources, and communication resources. Considering that crossover and mutation operations may generate infeasible solutions, a penalty mechanism is introduced to construct a piecewise fitness function: when an individual meets the constraints, its fitness is calculated according to the optimization objective; when the constraints are violated, a penalty is applied to reduce its fitness, thereby suppressing the propagation of infeasible solutions and guiding the population to converge towards the feasible solution space.

[0143] .

[0144] (3) Selection, crossover and mutation

[0145] To improve global search capabilities in complex non-convex scenarios, an adaptive crossover and mutation strategy is designed. The crossover and mutation probabilities are dynamically adjusted based on population fitness to balance global exploration and local development, and to reduce the risk of premature convergence. Simultaneously, a fitness-based selection mechanism is introduced, selecting high-quality individuals as parents according to the "survival of the fittest" principle.

[0146] ;

[0147] in: This represents the set of all individuals in the l-th generation.

[0148] To address the hybrid encoding characteristics, the traditional genetic algorithm was improved. In the crossover phase, continuous gene blocks employ a parent mean strategy to generate offspring, ensuring the smoothness of the continuous variable search. Simultaneously, a self-crossover mechanism is introduced, randomly exchanging some genes within discrete and continuous gene blocks to enhance the preservation of coupling relationships between variables and improve solution space exploration efficiency. In the mutation phase, discrete gene blocks undergo integer random mutation to enhance the discrete decision space search capability, while continuous gene blocks employ Gaussian mutation to improve the local optimization accuracy of continuous variables, thereby enhancing the algorithm's performance under complex constraints.

[0149] Step S6: After the preset large time scale decision period is reached, re-execute steps S4 to S7 to achieve dual time scale collaborative optimization of service caching decision and resource scheduling decision.

[0150] As mentioned above, this invention considers that task execution not only requires computing resources but also depends on the support of specific service modules, and that service caching and resource allocation have inherent time-scale differences. Therefore, it proposes a RIS-assisted service caching and resource allocation method based on dual-time-scale optimization. A joint optimization model is constructed, incorporating RIS phase shifting, service caching, task offloading, and resource allocation. Addressing the dynamic changes in service demand as vehicles move and the time-scale differences in different decision variables, a dual-time-scale collaborative optimization framework is constructed: at the large time scale, LSTM is introduced to enhance the DRL's ability to model historical service request information, achieving dynamic service caching optimization; at the small time scale, an improved genetic algorithm based on hybrid encoding is designed to achieve joint optimization of task offloading and resource allocation, improving resource scheduling capabilities under complex constraints. Results show that the proposed method outperforms benchmark methods in terms of long-term system performance, task completion latency, and system energy consumption, validating the effectiveness of the dual-time-scale collaborative optimization framework in RIS-assisted VEC service caching scenarios.

[0151] To verify the effectiveness of the proposed RIS-assisted service caching and resource allocation method based on dual time scale optimization, this invention proposes a specific embodiment, comparing the performance of the following five strategies with the proposed TT-LDGA strategy. Specifically, simulation experiments were conducted on this invention in a Python 3.9 and PyTorch 1.13 environment. The TT-LDGA strategy of this invention is compared with the No-RIS auxiliary strategy, the Random-RIS phase shift strategy, the Single-TS-DDPG strategy, the Random-cached strategy, and the Popular-cached strategy.

[0152] (1) No-RIS

[0153] Without deploying RIS-assisted communication, the system only has a direct communication link;

[0154] (2) Random-RIS phase shift

[0155] In the TT-LDGA framework, the phase shift coefficients of RIS are not optimized, but are randomly selected from a discrete set;

[0156] (3) Single-layer network (Single-TS-DDPG)

[0157] A single-layer network is used to process the system state at the most recent small time scale, without introducing LSTM for time series modeling;

[0158] (4) Random-cached service

[0159] Under the premise of meeting cache capacity constraints, services are randomly selected for caching, without relying on any historical information or popularity statistics;

[0160] (5) Service caching based on popularity (Popular-cached)

[0161] Services are cached based on statistical service popularity information, that is, services with the highest request frequency are cached first until the cache capacity is full.

[0162] Figure 5 The cumulative reward convergence performance of the proposed TT-LDGA algorithm is shown under different numbers of vehicles. As can be seen from the figure, the algorithm exhibits good convergence performance under different numbers of vehicles. The convergence speed decreases as the number of vehicles increases, because the environment becomes more complex with more vehicles, and the dimensions of the state and action spaces also increase.

[0163] Figure 6 The impact of different total bandwidths on short-term system performance is demonstrated. With increasing total bandwidth, the performance of all methods improves. This is because increased bandwidth improves data transmission rate, thereby reducing communication latency during task offloading. Overall, the proposed TT-LDGA method consistently performs best under all bandwidth conditions, with a more pronounced advantage in low-bandwidth scenarios, indicating its effective optimization of system performance even with limited communication resources. In contrast, comparative methods lacking channel optimization or caching strategies show relatively limited performance improvements. As bandwidth further increases, the performance gap between methods narrows, but the TT-LDGA method still maintains its lead, demonstrating good stability and adaptability.

[0164] Figure 7This study demonstrates the impact of MEC server caching capabilities on the long-term system performance of various methods. As the cache space increases, the performance of all methods continuously improves. Random-cached, employing a random caching strategy, makes caching decisions without considering service request characteristics and system state, resulting in a low cache hit rate and frequent service replacements, thus maintaining consistently low overall performance. Popular-cached caches services based on historical request frequency, prioritizing popular services, which improves the cache hit rate to some extent, but its performance remains limited due to a lack of comprehensive consideration of task unloading, communication status, and resource allocation. In contrast, the proposed TT-LDGA method combines dynamic system information for joint optimization, effectively controlling cache switching costs while improving the cache hit rate, thus maintaining optimal long-term system performance under different caching capabilities, achieving an improvement of approximately 12.4% over Popular-cached. Furthermore, as the cache capacity further increases, the performance improvement gradually plateaus, indicating that the system has approached the upper limit of caching benefits.

[0165] It should be noted that the examples of the above embodiments can preferably be combined with one or more of each other according to actual needs, and the accompanying drawings of multiple examples adopt a set of combined technical features, which will not be described in detail here.

[0166] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for RIS-aided service caching and resource allocation based on two-time-scale optimization, characterized in that, The steps include the following: Step 1: Construct a RIS-assisted in-vehicle edge computing system; Step 2: Based on the vehicle edge computing system architecture built in Step 1, build the service cache model, wireless communication model, and computing model respectively, and fully define the variables, constraints, overhead / latency / energy consumption calculation formulas of each model to obtain all the basic quantitative constraints and calculation formulas of the system. Step 3: Based on the three types of models and all formulas output in Step 2, construct the multivariate joint optimization objective function and complete constraint set to obtain the NP-hard mixed integer nonlinear programming optimization problem P1; Step 4: Based on the optimization problem obtained in Step 3 and the service popularity / cache gain / cache state data output in Step 2, the LSTM-DDPG algorithm is used to generate static cache decisions on a large time scale and update the cache content of the edge server. Step 5: Based on the fixed cache state output in Step 4, collect multi-dimensional environmental states in real time at each independent small time scale to obtain small-scale optimized complete input parameters; Step 6: Based on the real-time environmental parameters collected in Step 5, the fixed cache state in Step 4, and the joint optimization model constructed in Step 3, an improved genetic algorithm is used to jointly optimize offloading, RIS phase shift, bandwidth, and computing resources, and output the optimal real-time resource scheduling scheme. Step 7: Based on the resource scheduling scheme output in Step 6, execute full vehicle task processing, synchronously update the real-time status of the entire domain, and output the updated environmental data at a small time scale. Step 8: Loop judgment and dual time scale collaborative iteration, based on the complete time sequence process of steps 4-7 to achieve long-term collaborative optimization.

2. The RIS-assisted service caching and resource allocation method based on double-time-scale optimization according to claim 1, characterized in that, Step 1 is as follows: Step 1.1: Deploy N vehicle users, a base station equipped with M MEC edge servers, 1 cloud server, and a RIS containing Z reflection units; Step 1.2, define the link connection relationship: vehicle and base station wireless direct connection, base station and cloud server fiber optic backhaul, RIS deployed on building surface to reflect vehicle wireless signal; Step 1.3, set the business logic: the vehicle generates F type computing tasks, each type of task corresponds to a dedicated service model, the cloud server stores all services, the edge server is limited by storage and only caches some services, and outputs the overall system topology.

3. The RIS-assisted service caching and resource allocation method based on double-time-scale optimization according to claim 2, characterized in that, Step 2 is as follows: Step 2.1, Construct the service caching model: Define binary cache state variables Characterize whether edge server m caches service f; set cache capacity constraints; construct a formula for calculating cache switching costs: ； in This is an indicator function that takes the value 1 when the condition x is true, and 0 otherwise. This indicates that the service needs to be obtained from a cloud server. For a fixed backhaul link transmission rate; output cache storage limit and cache update overhead quantification rules; Step 2.2, Construct the wireless communication model: Integrate the vehicle-base station direct link and the RIS reflection-assisted link, and derive the small time scale time slots. Formula for signal-to-interference-plus-noise ratio between VU n and BS: ； in For the transmission power of VU n, It is additive white Gaussian noise. For the channel gain between RIS and BS, For the channel gain from VU n to RIS The channel gain between VU n and BS, Here is the diagonal phase shift matrix of RIS, and the phase shift... Set hardware constraints for discrete quantization phase shift: ； in To control the quantization accuracy of phase shift; output wireless transmission rate, channel quality, and RIS phase value constraints; Step 2.3, Constructing the computational model: Defining unloading decision variables Calculate resource allocation variables Set preconditions for task execution: ； in Indicates VU n in Unload the task to MEC server m. Indicates local computation. This section describes cloud computing; it categorizes task execution modes into four types: local, single-edge, multi-edge collaboration, and cloud; and derives a unified latency formula for these multiple modes. ； in This indicates the time required for the task to execute locally. and This indicates the transmission and processing time required for the task to execute on the MEC server m. , and This indicates that MEC server m does not have the services required for task processing, and these services are handled by the cooperating MEC server. The required transmission and processing time for execution This indicates that all MEC servers do not cache the time required for the corresponding service to be processed by the cloud server. The formula for calculating the total energy consumption of the build task is as follows: ； in This represents the energy consumption of transferring the task to the MEC server m. This represents the energy consumption required for local processing; it defines the QoS quantification formula for a single vehicle. ； in These are the weighted parameters in QoS that characterize the VU's preference for latency and energy consumption. To ensure the normal operation of VU n, the latency and energy consumption thresholds are set. It provides complete quantitative rules for task processing latency, including task unloading, resource allocation, latency and energy consumption, and user service quality.

4. The RIS auxiliary service caching and resource allocation method based on dual time scale optimization according to claim 3, characterized in that, Step 3 specifically involves: Step 3.1: Construct the long-term optimization objective function P1 and introduce weighting factors. Balancing QoS benefits with cache switching overhead: ； Step 3.2, superimpose all constraints C1-C13: Constraint C1 states that a task can only be processed when the MEC server has cached the corresponding service model and allocated the necessary computing resources; constraint C2 states that the bandwidth of the MEC server is only allocated to the VU performing task offloading; constraint C3 states that tasks generated by the VU can only be processed by a single server; constraints C4, C5, and C6 represent the cache capacity, computing power, and bandwidth resource limits of each MEC server, respectively; constraints C7 and C8 are the upper limits of latency and energy consumption of the VU; constraints C9-C13 make this problem a mixed-integer nonlinear programming problem. Step 3.3 outputs an NP-hard joint optimization problem containing coupled variables of long and short time scales and a mixture of discrete and continuous variables.

5. The RIS auxiliary service caching and resource allocation method based on dual time scale optimization according to claim 4, characterized in that, Step 4 specifically involves: Step 4.1: Count the frequency of each service request within the historical time window to obtain the service popularity. The QoS improvement brought by quantification caching is used to obtain the service cache gain. Read the current edge server cache state. Combining them yields a single small-scale state. ; Step 4.2: Generate the state sequence of all K hourly intervals within a large period. Input an LSTM network to extract temporal correlation features of service requests, and output a unified state over a large time scale. ; Step 4.3, convert the time-series characteristic states Input the DDPG deep reinforcement learning network and output a binary decision for the service cache of each edge server. ; Step 4.4: Update the storage content of the edge server according to the caching decision. This caching configuration will not change within the current complete large time scale period. The output fixed cache state is used as the input of the pre-constraint for the optimization of the lower small time scale.

6. The RIS auxiliary service caching and resource allocation method based on dual time scale optimization according to claim 5, characterized in that, Step 5 specifically involves: Step 5.1, collect data from the vehicle side: the current task type, task data volume, and local computing power of each vehicle; Step 5.2, collect communication side data: vehicle-base station direct connection channel Vehicle-RIS Channel RIS - Base Station Channel Noise power ; Step 5.3, Acquire phase data from the RIS hardware side: Current phase of each reflection unit. Quantization accuracy ; Step 5.4: Collect data from the edge server side: remaining storage, remaining computing power, and remaining bandwidth resources; Step 5.6: Integrate all real-time state parameters and output a small-time-scale optimized input set.

7. The RIS auxiliary service caching and resource allocation method based on dual time scale optimization according to claim 6, characterized in that, Step 6 specifically involves: Step 6.1, Hybrid Unified Chromosome Encoding: Discrete real number encoding is used for discrete variable task offloading and RIS phase shifting, while continuous real number encoding is used for continuous variable computational and communication resource allocation. The chromosome of the j-th individual in the l-th generation is... Each chromosome corresponds to a set of candidate scheduling solutions; Step 6.2, Construct a fitness function with constraints and penalties: Using the total QoS of the system as the optimization benchmark, penalties are imposed on individuals that violate constraints C1-C13. The piecewise fitness formula is as follows: ； Step 6.3, Improve the genetic iteration operation: ① Selection: Parent individuals are selected based on fitness ratio, with the following selection probability formula: in Represents the set of all individuals in the l-th generation; ② Adaptive crossover: The crossover probability is dynamically adjusted based on the overall fitness of the population; mean crossover of continuous gene blocks and exchange within discrete gene blocks are performed, while preserving the coupling relationship between variables; ③ Adaptive mutation: dynamically adjusts mutation probability; discrete gene integer random mutation, continuous gene Gaussian mutation; ④ Elite retention mechanism: The best individual is retained in each generation to avoid losing the best solution; Step 6.4: After the genetic iteration converges, output the current optimal scheduling scheme for the small time scale: task unloading matrix. RIS Discrete Phase Shift Matrix Bandwidth allocation Computing resource allocation .

8. The RIS auxiliary service caching and resource allocation method based on dual time scale optimization according to claim 7, characterized in that, Step 7 specifically includes: Step 7.1, based on the unloading decision Assign task execution location, matching edge server cache validation constraint C1; Step 7.2, based on RIS phase shift Adjusting the phase of all reflective units and reconstructing the wireless propagation link improves performance. ; Step 7.3, based on bandwidth Computing resources Allocate computing power and spectrum resources for edge servers; Step 7.4: Calculate the total task delay for each vehicle in different modes. Total energy consumption Verify threshold constraints C7 and C8; Step 7.5: After the task is completed, update the vehicle task queue, real-time channel gain, and remaining edge storage / computing power / bandwidth, and output the updated environmental state set.

9. The RIS auxiliary service caching and resource allocation method based on dual time scale optimization according to claim 8, characterized in that, Step 8 specifically includes: Step 8.1: Determine whether the current time series has completed a full large timescale cycle; Step 8.2, Period not reached: Jump to step 5, collect a new round of real-time status, and repeat steps 6 and 7 to complete small-scale scheduling; Step 8.3, Reaching the cycle: Jump to step 4, retrain LSTM-DDPG based on the accumulated small-scale time-series states to generate a new cache configuration, and then execute steps 5, 6, and 7 in sequence; Step 8.4 involves continuous cyclical two-layer timing optimization to maximize the system's average QoS and minimize the total cost of cache switching in the long term, outputting a continuously dynamically optimized operating strategy for the vehicle edge computing system.

Images