Medical edge computing task offloading and dynamic pricing method and system

Abstract

The application discloses a medical edge computing task offloading and dynamic pricing method and system, and the method comprises the following steps: a user equipment generates a computing task containing a medical emergency level; an edge server uses a multi-agent reinforcement learning intelligent agent integrated with a shared LSTM to predict a load trend according to a system historical state, and generates a differentiated pricing vector corresponding to the emergency level; the user equipment calculates a tendency value probability of selecting each edge server and local computing to maximize its own utility to execute an offloading decision; and a centralized training distributed execution framework is used to update intelligent agent parameters and optimize long-term benefits. The application differentiates task priorities through step-by-step pricing, realizes forward-looking decision-making by using the LSTM to perceive time sequence characteristics, and ensures game stability by selecting a tendency value model, effectively overcoming the defects of traditional methods, such as the inability to guarantee the service quality of key tasks, decision lag and system shock, and realizing low-delay, high-reliability and load-balanced resource allocation in a medical scene.

Description

Technical Field

[0001] This invention relates to the fields of medical Internet of Things and edge computing technology, and in particular to a method and system for medical edge computing task offloading and dynamic pricing. Background Technology

[0002] With the rapid development of IoT in healthcare (IoMT) in smart communities, the proliferation of personal health terminals has led to an explosive growth in real-time physiological monitoring data, making task offloading a key technology supporting low-latency healthcare services. In typical smart healthcare scenarios, computationally intensive health analysis tasks, such as electrocardiogram prediction or multidimensional physiological feature extraction, are increasingly being offloaded from resource-constrained personal devices to community edge servers. This architecture aims to alleviate the energy consumption pressure on terminals by utilizing edge computing resources and ensure timely analysis of critical medical data. To achieve this goal, existing research has introduced various task offloading and resource allocation frameworks based on edge computing. These frameworks typically utilize game theory or heuristic algorithms to build decision models, aiming to optimize the interaction between user devices (UEs) and edge servers and promote the efficient allocation of computational tasks. The application of these technologies has become an important trend in the IoMT field, demonstrating great potential in improving the efficiency of healthcare service response.

[0003] In existing technical solutions, the task offloading process typically relies on a centralized scheduling and management platform or a simple discrete game model. Specifically, the system usually assumes that the Edge Server provides a uniform service price for all connected user tasks, or charges linearly based solely on the size of the data packets. After generating a computation task, the UE selects the server with the highest utility as the offloading target based on the prices reported by each Edge Server and the current instantaneous communication latency. In this process, all health monitoring tasks are often treated as homogeneous data streams, and the system allocation logic mainly focuses on minimizing the completion latency of a single task or the terminal's energy consumption. Meanwhile, to cope with dynamically changing environments, some solutions attempt to introduce traditional game models for iterative solutions, trying to achieve some kind of equilibrium between users and servers.

[0004] However, several challenges remain in realizing the full potential of efficient task offloading systems in IoMT. First, traditional resource allocation methods often focus on unilateral objectives or homogeneous pricing, ignoring the inherent differences in urgency among medical computing tasks. Since existing solutions cannot differentiate urgency through price signals, in resource-competitive community environments, high-urgency emergency health tasks often compete with routine monitoring tasks, leading to increased processing delays for critical early warning information and compromising patient safety. Second, the dynamic nature of the task environment, including fluctuating task urgency, varying server load, and network conditions, can result in inefficient allocation or even service failure. Existing methods, particularly those based on traditional game theory, often fail to adapt quickly to these dynamic changes, leading to suboptimal resource utilization and reduced system performance. Furthermore, existing static game theory models struggle to adapt to the characteristics of user mobility and drastic fluctuations in wireless channels within community environments. Lacking the ability to capture temporal characteristics, their resource allocation strategies often lag when facing highly dynamic medical computing task flows, thus limiting the overall system efficiency. Summary of the Invention

[0005] The purpose of this invention is to address the problems of task urgency mismatch, weak adaptability to dynamic environments, lack of pricing adjustment mechanisms, and long-term decision lag in health monitoring task offloading in smart community IoMT scenarios. It provides a method and system for medical edge computing task offloading and dynamic pricing. By integrating Stackelberg Game theory and MARL technology, it achieves coordinated optimization of processing latency, terminal energy consumption, and quality of service (QoS) in complex community environments. Simultaneously, it utilizes an urgency quantification mechanism based on the medical data priority classification method in the IEEE 802.15.6 standard and a differentiated dynamic pricing strategy to improve the targeting of resource allocation and system response efficiency. Furthermore, through LSTM-driven time-series state awareness, it ensures decision accuracy under highly dynamic medical data flows, ultimately achieving the goal of low-latency, high-reliability, and dynamically adaptive medical computing task processing.

[0006] The technical solution adopted in this invention is:

[0007] A method for offloading medical edge computing tasks and dynamic pricing includes the following steps:

[0008] Initialization: Configure the computing resources and communication parameters of the edge servers, and initialize the multi-agent reinforcement learning (MARL) model for each edge server;

[0009] Task generation and broadcasting: The user equipment assesses the medical urgency of the collected physiological signal characteristics to obtain the task urgency, generates a medical computing task with the task urgency, and sends it to the scheduling and management platform; the scheduling and management platform broadcasts the medical computing task to all edge servers;

[0010] Pricing decision: Each edge server generates a differentiated pricing list based on task urgency and current system status using a MARL model containing a shared LSTM network, and then aggregates it to the scheduling management platform; the differentiated pricing list includes multiple price components corresponding to different task urgency.

[0011] Offloading decision: Each user device calculates the utility of selecting different edge servers and local computing based on a differentiated pricing list of all edge servers distributed by the scheduling management platform. And according to utility Calculate the corresponding preference probability Finally, based on the probability of choice Make the final uninstall decision ;

[0012] Task execution: Under the coordination of the scheduling management platform, the user equipment offloads the medical computing task to the selected edge server or local computing according to the offloading decision; after the medical computing task is completed, the actual task latency and load are recorded, and the reward for each edge server is calculated; the processing results of the medical computing task are stored as experience data in the experience replay pool.

[0013] Model update: The edge server randomly samples a batch of data from the experience replay pool to update the network parameters of the MARL model agent, in order to optimize the long-term benefits of the edge server and the overall system performance.

[0014] Furthermore, each user device generates medical computing tasks based on the collected physiological signal characteristics (such as heart rate and electrocardiogram). , is represented as: ,in, Task data volume (unit: MB); Computational intensity (CPU cycles required to process one MB of data); The task urgency level is determined by a built-in algorithm based on standards such as IEEE 802.15.6 (e.g., level 1-8).

[0015] Furthermore, the urgency of the task is determined based on the medical crisis level to which the medical computing task belongs; the medical crisis level is a set of discrete levels divided according to the degree of medical urgency. For example, arrhythmia data detected by electrocardiogram (ECG) can be classified as high-risk, while body temperature monitoring data can be classified as low-risk; assuming a total of Each medical emergency level is denoted as a set. ; Task urgency The expression is obtained by weighted summation of the task's level, as follows:

[0016] ;

[0017] Where S represents the total number of medical emergency levels (e.g., 8). For normalized weights, satisfying and , used to quantify the relative severity of level s-th level medical emergency; As an indicator variable, if the current computing task of the i-th user device belongs to level s, then Otherwise, it is 0.

[0018] Furthermore, the implementation of the pricing decision-making process includes:

[0019] The agent of each edge server j collects the observation state of the current time slot. To characterize the dynamic changes of the system, the observation state Including the pricing of the previous time slot itself Information on the urgency distribution of user tasks in the current time slot and the actual load of the previous time slot The observation information from all servers can be combined to form a global state. Value assessment for the centralized training phase

[0020] To capture the temporal variation patterns of the system, this invention uses the observation sequences from the most recent time slots. The input is fed into the LSTM module to extract temporal features from historical states;

[0021] observation sequence The input is a shared LSTM network module; the LSTM network performs time-series modeling of historical load changes, pricing strategies, and task urgency distribution through a gating structure, and outputs the global hidden state vector of the system in time slot 𝑡. ; This represents the hidden state vector of the system in time slot 𝑡. This vector is obtained by time-series modeling of historical observation sequences using a shared LSTM network, and is used to comprehensively characterize the changing trends of system load demand and the competitive relationships between servers. By introducing this hidden state representation, each edge server can infer the current potential state of the system based on historical information, thereby improving the accuracy and stability of pricing decisions in partially observable environments.

[0022] Hidden state of LSTM network output The input consists of an Actor network (a deep learning model) for each agent; the Actor network is a multi-layer feedforward neural network whose input is the global temporal features extracted by LSTM, and whose output is the server's pricing strategy. This represents the strategy function for the j-th server. For Actor network parameters;

[0023] The Actor network outputs a continuous action vector, which is then mapped to a pricing vector corresponding to the urgency level of the medical task. S represents the total number of urgency levels for the task;

[0024] in This is the unit price for the task with the lowest urgency (level 1). It is the unit price for the highest urgency (Level S) mission, and meets the following conditions: This reflects the principle of "high priority and high premium".

[0025] Each edge server sends its own pricing vector to the scheduling management platform, which then aggregates and forwards them to all user devices.

[0026] Furthermore, the implementation of the unloading decision-making step includes:

[0027] Utility calculation: User device i calculates the utility of selecting different options k. The general utility function is defined as follows: ,in For indicator functions, For the collection of all edge servers, Pricing is set for each unit of data on edge server j. The amount of task data for user device i;

[0028] When k=0, it indicates local computation. =0, the indicator function takes the value of 0. The utility function simplifies to: ,in, Satisfaction with the quality of service for local computing. Calculate energy consumption locally;

[0029] When k=j, it means the data is unloaded to edge server j. The indicator function takes the value of The utility function simplifies to ,in The service quality satisfaction of offloading user device i to edge server j is related to task latency. A decreasing function; The urgency of the task corresponding to edge server j Pricing; The amount of task data for user device i;

[0030] Calculation of propensity to choose: Calculate the propensity to choose each option k based on the Logit model. , Where β is the sensitivity parameter, , This represents the maximum amount of user device task data in the system, measured in MB. This value is a system preset constant and can be pre-set based on the maximum amount of medical task data that may be generated in actual application scenarios. For example, raw electrocardiogram data may reach tens of MB, and this value can be set accordingly. .

[0031] The maximum allowed unit price is a system-preset constant determined by the service provider based on market pricing strategies, user willingness to pay, and operating costs, serving as a price ceiling constraint. This represents the minimum unit cost of the server. This value is a system-preset constant that reflects the minimum operating cost (such as energy consumption cost, hardware wear and tear, etc.) required for the server to process a unit of data. It can be obtained through server energy consumption models and historical operating data statistics. A set of all available options, including local computation ( ) and all edge servers ( Decision execution: based on the probability of choice preference. The final unloading decision will be made by randomly selecting or choosing the option with the highest probability. .

[0032] Furthermore, the user device i's satisfaction with the processing service quality of edge server j. The calculation formula is:

[0033] ,

[0034] in For QoS satisfaction coefficient, The maximum tolerable delay for the task. Total task delay.

[0035] Furthermore, the task execution steps will include tuples ( Store it in the experience replay pool.

[0036] Furthermore, the implementation of the model update step includes:

[0037] Centralized training: The scheduling and management platform uses the experience replay pool A small batch of historical experience data is randomly sampled to obtain global information x, and the global information x is used to update the Critic network parameters of the j-th edge server (agent). The training objective of the Critic network is to minimize the temporal difference error, and its loss function is... Defined as:

[0038] ;

[0039] in, The experience replay pool is used to store historical experience tuples generated by the interactions between all edge servers (agents) and the environment. Where x represents the global information of the current time slot, For the action of the j-th edge server, The instant reward obtained by the j-th edge server This provides global information for the next time slot. The capacity of the experience replay pool is a preset constant, and small batches of data are extracted for training using a random sampling method during sampling. This represents the Critic network parameters of the j-th edge server, i.e., the action-value function. The neural network weights. This parameter is continuously updated and optimized during training using gradient descent. This represents the target Q-value generated by the target network calculated by the j-th edge server, used for stabilizing the training process. Its calculation formula is: ,in, For the j-th edge server performing the action The instant reward obtained afterward; This is a discount factor, with a value range of [0,1], used to balance the importance of current rewards and future rewards; This represents the target Critic network, which has the same network structure as the Critic network but updates its parameters more slowly. This provides global information for the next time slot; For target Actor network Based on local observations in the next time slot The output action; The target Actor network has the same network structure as the Actor network, but its parameters are updated more slowly. This represents the action chosen by the j-th edge server (agent) at the current moment, including its pricing strategy. And resource allocation decisions. N is the total number of edge servers. This represents the combined action of global information x and all edge servers. The Q-value estimate of the Critic network output of the j-th edge server is used to evaluate the merits of the current joint action. This represents global observation information, including the status information of all edge servers (such as load and historical pricing) and the task requirement information of all user devices (such as task urgency and data volume). This information is aggregated by the scheduling management platform and then used by the centralized training module deployed on it.

[0040] Policy Update: Each edge server j updates its Actor network parameters using gradient ascent. To maximize the expected cumulative reward, the policy gradient is calculated as follows:

[0041] ;

[0042] in, This represents the Actor network (i.e., the policy function) of the j-th edge server, with local observations as input. (For example, the server's load status, the pricing of neighboring servers, etc.), output actions ; These are the parameters of the Actor network; The gradient of the policy function with respect to the parameters; The Q-value of the Critic network output for actions The gradient; the overall expectation is derived from the experience replay pool. Global information x and joint action in mid-sample Calculated above.

[0043] Target network update: After each update of the Critic and Actor networks, the target network parameters are slowly updated using a soft update method to maintain the stability of the training process. The update formula is as follows:

[0044] ;

[0045] ;

[0046] in, and These represent the parameters of the target Critic network and the target Actor network of the j-th edge server, respectively. The soft update coefficient is typically a small constant with a value much less than 1 (e.g., ...). This is used to control the update speed of the target network parameters, so that the target network parameters slowly approach the current network parameters.

[0047] Furthermore, as a supplement to the embodiments of the present invention, the edge server of the present invention employs a load balancing mechanism: when predicting potential load... When the computational capacity threshold is exceeded, the MARL model agent's policy network (Actor) increases the price for high-urgency levels. Alternatively, the price difference between different levels could be increased to use price leverage to suppress the tendency of low-priority tasks to be accessed.

[0048] Furthermore, the present invention also discloses a medical edge computing task offloading and dynamic pricing system, which executes the aforementioned medical edge computing task offloading and dynamic pricing method. The system includes:

[0049] At least one user device is used to assess the medical urgency of the acquired physiological signal characteristics to obtain a task urgency level, generate a medical computing task with the task urgency level, and calculate the utility of selecting different edge servers and local computing based on a differentiated pricing list of all edge servers. And according to utility Calculate the corresponding preference probability Finally, based on the probability of choice Make the final decision to uninstall ;

[0050] At least one edge server, each edge server is configured with a MARL model agent, the MARL model agent includes a shared LSTM network and an Actor-Critic network for generating differentiated pricing vectors; completes medical computing tasks offloaded from user devices, records actual task latency and load, and calculates rewards for edge servers; stores the processing results of medical computing tasks as experience data in an experience replay pool.

[0051] The scheduling and management platform is used to receive medical computing tasks from user devices and broadcast them to various edge servers, aggregate and forward pricing vectors from edge servers to user devices, and coordinate the allocation process of medical computing tasks.

[0052] In this process, the edge server randomly samples a batch of data from the experience replay pool to update the network parameters of the MARL model agent, thereby optimizing the long-term benefits of the edge server and the overall performance of the system.

[0053] This invention, employing the above technical solutions, offers the following advantages compared to existing methods: 1) By constructing a bilateral game and learning framework between the server and users, this invention overcomes the dependence of traditional methods on a static environment, ensuring efficient and fair resource allocation in a highly dynamic and competitive market environment. 2) By introducing a differentiated pricing mechanism, this invention aims to address the technical deficiencies of existing technologies, such as the failure of pricing levers and the inability to provide deterministic resource guarantees for critical medical computing tasks. 3) By integrating a Long Short-Term Memory (LSTM) network into the state space of a MARL agent, this invention enables the agent to capture and predict the temporal characteristics of medical data. In this way, the agent can obtain more forward-looking state information, thereby adjusting pricing and resource allocation strategies in advance, improving the accuracy and timeliness of decision-making. Compared to existing methods that rely solely on instantaneous information for decision-making, this invention significantly enhances the system's adaptability and reliability to dynamic changes in medical scenarios, ensuring the continuity and timeliness of critical medical services.

[0054] This invention is particularly applicable to smart community healthcare scenarios, enabling efficient resource matching, load balancing, and low-latency assurance of medical information processing between community health edge servers and user devices for health monitoring tasks with varying degrees of medical urgency. Attached Figure Description

[0055] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments;

[0056] Figure 1 This is a schematic diagram of the medical edge computing task offloading and dynamic pricing method of the present invention; Figure 2 This is a schematic diagram of the LSTM-MADDPG algorithm for medical edge computing task offloading and dynamic pricing proposed in this invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0058] With the widespread adoption of IoT in smart community healthcare, massive amounts of real-time physiological monitoring data require low-latency processing. Traditional task offloading solutions have significant limitations: 1) Uniform pricing fails to differentiate the urgency of medical computing tasks, potentially causing delays in critical tasks due to resource competition; 2) Decision-making models based on instantaneous states lack foresight and struggle to adapt to dynamic environments; 3) Centralized or simple game-theoretic models are prone to system oscillations, failing to guarantee long-term stability and efficiency. This invention aims to address these issues by providing a solution that enables precise resource matching, ensures service quality for critical tasks, and possesses dynamic adaptive capabilities.

[0059] This invention discloses a method and system for task offloading and dynamic pricing in medical edge computing. The core is a task coordination framework for home-based healthcare, using a scheduler as an information intermediary to coordinate task broadcasting and resource matching. The system accurately characterizes task priorities through a medical urgency assessment module, and incentivizes Edge Servers to provide differentiated services for tasks of varying urgency using a differentiated pricing strategy. It employs a propensity score model based on the Logit function, transforming traditional discrete offloading decisions into a continuous probability mapping space, guiding the user to autonomously select the server node that maximizes utility. The system collects physiological task characteristics through a data perception module, sets tiered service prices through a pricing game theory module, and achieves optimal response through a reinforcement learning solution module. During task execution, it utilizes LSTM to capture temporal changes in physiological data, thereby ensuring the real-time performance of high-urgency tasks while achieving global load balancing and system revenue optimization in large-scale community scenarios.

[0060] like Figure 1 As shown, this invention discloses a method for offloading medical edge computing tasks and dynamic pricing, including the following steps:

[0061] Initialization: Configure the computing resources and communication parameters of the edge servers, and initialize the multi-agent reinforcement learning (MARL) model for each edge server;

[0062] Task generation and broadcasting: The user equipment assesses the medical urgency of the collected physiological signal characteristics to obtain the task urgency, generates a medical computing task with the task urgency, and sends it to the scheduling and management platform; the scheduling and management platform broadcasts the medical computing task to all edge servers;

[0063] Pricing decision: Each edge server generates a differentiated pricing list based on task urgency and current system state using a MARL model containing a shared long short-term memory network (LSTM) and aggregates it to the scheduling management platform; the differentiated pricing list includes multiple price components corresponding to different task urgency.

[0064] Offloading decision: Each user device calculates the utility of selecting different edge servers and local computing based on a differentiated pricing list of all edge servers distributed by the scheduling management platform. And according to utility Calculate the corresponding preference probability Finally, based on the probability of choice Make the final decision to uninstall ;

[0065] Task execution: Under the coordination of the scheduling management platform, the user equipment offloads the medical computing task to the selected edge server or local computing according to the offloading decision; after the medical computing task is completed, the actual task latency and load are recorded, and the reward for each edge server is calculated; the processing results of the medical computing task are stored as experience data in the experience replay pool.

[0066] Model update: The edge server randomly samples a batch of data from the experience replay pool to update the network parameters of the MARL model agent, in order to optimize the long-term benefits of the edge server and the overall system performance.

[0067] Furthermore, each user device generates medical computing tasks based on the collected physiological signal characteristics (such as heart rate and electrocardiogram). , is represented as: ,in, Task data volume (unit: MB); Computational intensity (CPU cycles required to process one MB of data); The urgency level is assigned (e.g., 1-8). This invention references the medical data priority classification framework defined in the IEEE 802.15.6 standard (which divides health monitoring data packets into 8 user priorities, with the highest priority used for emergency medical events). The built-in algorithm on the user device automatically assesses and outputs the corresponding urgency level based on the collected physiological signal characteristics (such as heart rate, ECG waveform, etc.). It should be noted that, as a feasible implementation, the medical urgency assessment of this invention can be an automatic classification from a mature algorithm built into the device, or it can be manually input by the user or based on existing tags obtained from an upstream medical system.

[0068] Specifically, it should be noted that the IEEE 802.15.6 standard, as a communication standard for wireless body area networks, primarily defines the physical layer and MAC layer protocols. However, an important component of this standard is User Priorities (UPs). The IEEE 802.15.6 standard does define eight user priorities, typically labeled UP0 to UP7. Priority 7 is specifically used for emergency and medical implantation event reporting. This priority classification framework provides a basis for differentiated services in the medical Internet of Things and has been adopted by numerous academic studies.

[0069] Furthermore, the urgency of the task is determined based on the medical crisis level to which the medical computing task belongs; the medical crisis level is a set of discrete levels divided according to the degree of medical urgency. For example, arrhythmia data detected by electrocardiogram (ECG) can be classified as high-risk, while body temperature monitoring data can be classified as low-risk; assuming a total of Each medical emergency level is denoted as a set. ; Task urgency The expression is obtained by weighted summation of the task's level, as follows:

[0070] ;

[0071] in, The total number of medical emergency levels. For normalized weights, satisfying and , used to quantify the relative severity of level s-th level medical emergency; As an indicator variable, if the current computing task of the i-th user device belongs to level s, then Otherwise, it is 0. Furthermore, the implementation of the pricing decision step includes:

[0072] The agent of each edge server j collects the observation state of the current time slot. To characterize the dynamic changes of the system, the observation state Including the pricing of the previous time slot itself Information on the urgency distribution of user tasks in the current time slot and the actual load of the previous time slot The observation information from all servers can be combined to form a global state. Used for value assessment during the centralized training phase.

[0073] To capture the temporal variation patterns of the system, this invention uses the observation sequences from the most recent time slots. The input is fed into the LSTM module to extract temporal features from historical states;

[0074] observation sequence The input is a shared LSTM network module; the LSTM network performs time-series modeling of historical load changes, pricing strategies, and task urgency distribution through a gating structure, and outputs the global hidden state vector of the system in time slot 𝑡. ; This represents the hidden state vector of the system in time slot 𝑡. This vector is obtained by time-series modeling of historical observation sequences using a shared LSTM network, and is used to comprehensively characterize the changing trends of system load demand and the competitive relationships between servers. By introducing this hidden state representation, each edge server can infer the current potential state of the system based on historical information, thereby improving the accuracy and stability of pricing decisions in partially observable environments.

[0075] Hidden state of LSTM network output The input consists of an Actor network (a deep learning model) for each agent; the Actor network is a multi-layer feedforward neural network whose input is the global temporal features extracted by LSTM, and whose output is the server's pricing strategy. , This represents the strategy function for the j-th server. For Actor network parameters;

[0076] The Actor network outputs a continuous action vector, which is then mapped to a pricing vector corresponding to the urgency level of the medical task. S represents the total number of urgency levels for the task;

[0077] in This is the unit price for the task with the lowest urgency (level 1). It is the unit price for the highest urgency (Level S) mission, and meets the following conditions: This reflects the principle of "high priority and high premium".

[0078] Each edge server sends its own pricing vector to the scheduling management platform, which then aggregates and forwards them to all user devices.

[0079] Furthermore, the implementation of the unloading decision-making step includes:

[0080] Utility calculation: User device i calculates the utility of selecting different options k. The general utility function is defined as follows: ,in For indicator functions, For the collection of all edge servers, Pricing is set for each unit of data on edge server j. The amount of task data for user device i; when k=0, it indicates local computation. =0, the indicator function takes the value of 0. The utility function simplifies to: ,in, Satisfaction with the quality of service for local computing. For local energy consumption calculation, when k=j, it indicates that the energy is offloaded to edge server j. The indicator function takes the value of The utility function simplifies to ,in The service quality satisfaction of offloading user device i to edge server j is related to task latency. A decreasing function; The urgency of the task corresponding to edge server j Pricing; The amount of task data for user device i;

[0081] Calculation of propensity to choose: Calculate the propensity to choose each option k based on the Logit model. , , where β is a sensitivity parameter used to adjust the user's sensitivity to differences in utility; to ensure model convergence, , This represents the maximum amount of user device task data in the system, measured in MB. This value is a system preset constant and can be pre-set based on the maximum amount of medical task data that may be generated in actual application scenarios. For example, raw electrocardiogram data may reach tens of MB, and can be set accordingly. . The maximum allowed unit price is a system-preset constant determined by the service provider based on market pricing strategies, user willingness to pay, and operating costs, serving as a price ceiling constraint. This represents the minimum unit cost of the server. This value is a system-preset constant that reflects the minimum operating cost (such as energy consumption cost, hardware wear and tear, etc.) required for the server to process a unit of data. It can be obtained through server energy consumption models and historical operating data statistics. A set of all available options, including local computation ( ) and all edge servers ( ).

[0082] Decision execution: based on the probability of choice preference The final unloading decision will be made by randomly selecting or choosing the option with the highest probability. .

[0083] Furthermore, the user device i's satisfaction with the processing service quality of edge server j. The calculation formula is:

[0084] ,

[0085] in For QoS satisfaction coefficient, The maximum tolerable delay for the task. Total task delay.

[0086] Furthermore, the task execution steps will include tuples ( Store it in the experience replay pool.

[0087] Furthermore, the implementation of the model update step includes:

[0088] Centralized training: The scheduling and management platform uses the experience replay pool A small batch of historical experience data is randomly sampled to obtain global information x, and the global information x is used to update the Critic network parameters of the j-th edge server (agent). The training objective of the Critic network is to minimize the temporal difference error, and its loss function is... Defined as:

[0089] ;

[0090] in, This represents the experience replay pool, used to store historical experience tuples generated from all interactions between edge servers (agents) and the environment. Where x represents the global information of the current time slot, For the action of the j-th edge server, The instant reward obtained by the j-th edge server This provides global information for the next time slot. The capacity of the experience replay pool is a preset constant, and small batches of data are extracted for training using a random sampling method during sampling. This represents the Critic network parameters of the j-th edge server, i.e., the action-value function. The neural network weights. This parameter is continuously updated and optimized during training using gradient descent. This represents the target Q-value generated by the target network calculated by the j-th edge server, used for stabilizing the training process. Its calculation formula is: ,in, For the j-th edge server performing the action The instant reward obtained afterward; This is a discount factor, with a value range of [0,1], used to balance the importance of current rewards and future rewards; This represents the target Critic network, which has the same network structure as the Critic network but updates its parameters more slowly. This provides global information for the next time slot; For target Actor network Based on local observations in the next time slot The output action; The target Actor network has the same network structure as the Actor network, but its parameters are updated more slowly. This represents the action chosen by the j-th edge server (agent) at the current moment, including its pricing strategy. And resource allocation decisions. N is the total number of edge servers. This represents the combined action of global information x and all edge servers. The Q-value estimate of the Critic network output of the j-th edge server is used to evaluate the merits of the current joint action. This represents global observation information, including the status information of all edge servers (such as load and historical pricing) and the task requirement information of all user devices (such as task urgency and data volume). This information is aggregated by the scheduling management platform and distributed to the centralized training module deployed on the platform for centralized training. Specifically, during the centralized training phase, in addition to its responsibilities for task broadcasting, price aggregation, and offloading coordination, the scheduling management platform is also responsible for collecting global status data, sampling experience data, and updating the Critic network parameters of each edge server.

[0091] Policy Update: Each edge server j updates its Actor network parameters using gradient ascent. To maximize the expected cumulative reward, the policy gradient is calculated as follows:

[0092] ;

[0093] in, This represents the Actor network (i.e., the policy function) of the j-th edge server, with local observations as input. (For example, the server's load status, the pricing of neighboring servers, etc.), output actions ; These are the parameters of the Actor network; The gradient of the policy function with respect to the parameters; The Q-value of the Critic network output for actions The gradient; the overall expectation is derived from the experience replay pool. Global information x and joint action in mid-sample Calculated above.

[0094] Target network update: After each update of the Critic and Actor networks, the target network parameters are slowly updated using a soft update method to maintain the stability of the training process. The update formula is as follows:

[0095] ;

[0096] ;

[0097] in, and These represent the parameters of the target Critic network and the target Actor network of the j-th edge server, respectively. The soft update coefficient is typically a small constant with a value much less than 1 (e.g., ...). This is used to control the update speed of the target network parameters, so that the target network parameters slowly approach the current network parameters.

[0098] Furthermore, as a supplement to the embodiments of the present invention, the edge server of the present invention employs a load balancing mechanism: when predicting potential load... When the computational capacity threshold is exceeded, the MARL model agent's policy network (Actor) increases the price for high-urgency levels. Alternatively, the price difference between different levels could be increased to use price leverage to suppress the tendency of low-priority tasks to be accessed.

[0099] Furthermore, the present invention also discloses a medical edge computing task offloading and dynamic pricing system, which executes the aforementioned medical edge computing task offloading and dynamic pricing method. The system includes:

[0100] At least one user device is used to assess the medical urgency of the acquired physiological signal characteristics to obtain a task urgency level, generate a medical computing task with the task urgency level, and calculate the utility of selecting different edge servers and local computing based on a differentiated pricing list of all edge servers. And according to utility Calculate the corresponding preference probability Finally, based on the probability of choice Make the final decision to uninstall ;

[0101] At least one edge server, each edge server is configured with a MARL model agent, the MARL model agent includes a shared LSTM network and an Actor-Critic network for generating differentiated pricing vectors; completes medical computing tasks offloaded from user devices, records actual task latency and load, and calculates rewards for edge servers; stores the processing results of medical computing tasks as experience data in an experience replay pool.

[0102] The scheduling and management platform is used to receive medical computing tasks from user devices and broadcast them to various edge servers, aggregate and forward pricing vectors from edge servers to user devices, and coordinate the allocation process of medical computing tasks.

[0103] In this process, the edge server randomly samples a batch of data from the experience replay pool to update the network parameters of the MARL model agent, thereby optimizing the long-term benefits of the edge server and the overall performance of the system.

[0104] The specific principles of this invention will be explained in detail below:

[0105] This invention proposes a task offloading and dynamic pricing framework (MARL-IoMT-TOC) for health monitoring in smart communities, integrating MARL and Stackelberg Game theory. The MARL-IoMT-TOC system aims to optimize the allocation of computing resources in complex, multi-competitive environments by dynamically balancing the operational revenue of Edge Servers and the service utility of Users (UEs). The system employs a Scheduler as a matchmaking intermediary for information exchange, responsible for broadcasting task requests, aggregating pricing information from various servers, and feeding it back to the UE. The Scheduler does not possess absolute decision-making power but rather supports decentralized game theory between Edge Servers and UEs by coordinating information flow.

[0106] The core of the MARL-IoMT-TOC system is a game-driven task decision-making architecture, which consists of three key process modules: a task feature perception and urgency classification module, a propensity-based task matching module, and an agent-based differentiated pricing module. The urgency classification module prioritizes physiological data; the matching module uses a propensity function to transform the user's discrete unloading choices into a continuous probability space to ensure the stability of the game equilibrium; and the pricing module uses an agent with an integrated shared LSTM to perceive the environment and output the optimal tiered pricing strategy.

[0107] This invention is a product of Each UE, The smart community IoMT system consists of an Edge Server and a Scheduler. User Equipment (UE) 101 is a personal medical terminal with data acquisition and communication capabilities, such as a smart bracelet, ECG patch, or portable pulse oximeter. Each UE 101 periodically or event-drivenly generates health monitoring tasks. Each task It can be characterized by triples: .in, The amount of data for the task (in MB). This indicates computational intensity (i.e., the number of CPU cycles required to process each unit of data, measured in Cycles / MB). This represents the medical urgency level according to predefined criteria, where S is the total number of levels. As a feasible implementation, for example, S=8, where level 1 represents routine physical examination data, and level 8 represents suspected myocardial infarction warning data. UE As a rational decision-maker, based on the pricing and QoS assessment of all servers, choose the one that maximizes your own utility. Uninstallation target including local computing .

[0108] Edge Server: The Edge Server (ES) 102 is deployed within smart communities, such as community medical centers, base stations, or dedicated gateways, and possesses strong computing capabilities. (Unit: GHz) and storage resources. Each ES 102 is modeled as an independent, economically rational agent whose core function is to run a local reinforcement learning policy network (Actor network) to dynamically generate an S-dimensional differentiated pricing vector based on the system state. ,in , This indicates the unit data price charged for a task with an urgency level of 's' (unit: currency unit / MB). Server As an RL agent, it executes the pricing strategy in each time slot. The goal is to maximize one's long-term cumulative rewards.

[0109] Scheduling Management Platform: The Scheduling Management Platform (Scheduler) 103, as a logically centralized information intermediary platform, is typically deployed in a community cloud center or on a core edge node. As an information intermediary, the Scheduling Management Platform (Scheduler) 103 is responsible for collecting user requests, forwarding server quotes, confirming the final offloading decision, and coordinating task allocation and resource matching processes. Its main responsibilities are:

[0110] (1) Information collection and broadcasting: Receive computing tasks sent by all UE 101 And broadcast it to all available ES102.

[0111] (2) Price aggregation and forwarding: Receive pricing vectors from all ES 102 responses. And forward it to the corresponding UE101.

[0112] (3) Decision confirmation and coordination: Receive the final offloading decision from UE 101 ( =0 indicates local calculation. =j indicates unloading to ES j), and updates the global resource mapping table to coordinate the transmission and execution of task data.

[0113] (4) Maintain global information: Maintain system parameters, such as total bandwidth B, channel model, computing capacity of each ES, etc.

[0114] Furthermore, UE 101 and ES 102 are connected via a wireless network (such as Wi-Fi or 5G), and their channel gain... and background noise This will change over time. The ES 102 and Scheduler 103 are typically connected via a high-bandwidth, low-latency wired or stable wireless network.

[0115] The system operates in discrete time slots. In each time slot, the UE generates computational tasks with specific data volume and medical urgency based on the collected physiological signal characteristics. The Edge Server, as an economic entity with individual rationality, dynamically adjusts its service price based on its own load and market dynamics by deploying the Multi-Agent Deep Deterministic Policy Gradient (MADDPG) algorithm. The Scheduler maintains a global communication model and resource mapping table to ensure precise alignment between task flows and resource flows.

[0116] Communication Model: Medical computing tasks require maximum permissible latency The transmission rate is achieved within the specified timeframe. When UE i decides to offload the task to ES j, its transmission rate... Based on Shannon's theorem and considering multi-user interference, the total system bandwidth B is divided into W orthogonal sub-channels. Assuming an equal-division strategy, the sub-channel bandwidth occupied by UE i and ES j for communication is calculated as follows. ,in For the total operating frequency band, Let be the number of sub-channels. Then, what is the transmission rate of UE i offloaded to ES j? for:

[0117] ;

[0118] in For transmission power, For channel gain, This represents the background noise power spectral density.

[0119] Total task delay Due to transmission delay and server computation latency Composition. Transmission delay ,in This refers to the amount of data involved in the task. Server computation latency. ,in This is the total CPU cycles required for the task. Additionally, selecting the same ES... The frequency of calculating resource allocation per user is ,in It refers to the server's computing power. Is it currently uninstalled to The number of users. If UE i chooses local calculation ( If =0), then the calculation delay is Local computing power consumption ;in and ξ represents the local CPU frequency and effective switched capacitor coefficient of UE i, respectively.

[0120] Mission urgency It is a key indicator guiding server pricing and user decisions. Task urgency. Multiple health and safety levels can be defined based on clinical guidelines, and the result can be obtained by weighted summation of multiple task-related health and safety levels:

[0121] ;

[0122] in The total number of health and safety levels (e.g., 8). The normalized weights reflect the relative urgency of level s (Σ). =1). For example, it can be set to =0.05 (normal), =0.3 (critical); As an indicator variable, if the task Belongs to the level ,but Otherwise Task urgency As an input to the ES pricing game, and reflected in the server's service health and efficiency. middle.

[0123] This invention samples a bilateral utility function to accurately model the rational choices of the User Experience Provider (UE) and the long-term benefit objectives of the ES agent. By introducing a propensity score model, this invention evolves the traditional discrete offloading decision into a continuously differentiable probabilistic mapping, thereby ensuring system stability during large-scale medical task integration.

[0124] (a) UE utility function UE The core objective is to minimize payment costs and maximize QoS satisfaction while meeting task latency constraints. This utility function comprehensively considers service satisfaction, energy costs, and payment costs.

[0125] 1) QoS satisfaction calculation: User satisfaction with the server Satisfaction with the processing services provided Defined as:

[0126] ;

[0127] in, It is the QoS satisfaction coefficient. The maximum tolerable delay for the task. This represents the total task latency. This logarithmic model accurately reflects the non-linear perception of latency reduction by users in medical scenarios.

[0128] 2) Deterministic utility function: User Choose local computing or select a server. Net utility Defined as

[0129] ;

[0130] in Energy consumption cost calculated locally; Uninstall to server The payment costs.

[0131] (ii) Edge server utility functions Edge Server As the leader in a game, one must formulate pricing strategies in a competitive environment to maximize one's profits. To enable the server to more accurately quantify user needs and ensure excellent mathematical properties, this invention introduces a selection bias value into the server's decision-making logic.

[0132] 1) Introduction of Selection Propensity: The server obtains the utility distribution of each user through the Scheduler and calculates the user's... For the server Choice bias :

[0133] ;

[0134] For the server, this probability value reflects both the user's high preference for the current optimal option and... hour, It maximizes the potential for users to switch to other servers in the future due to price adjustments (non-zero probability). This continuous probability mapping makes the expected load function of the server smooth and strictly differentiable, avoiding the discrete oscillations caused by traditional hard allocation (such as argmax), which is beneficial for policy gradient optimization in multi-agent reinforcement learning. 2) Server revenue function and dual pricing scheme: Edge Server As the game leader, the aim is to maximize its long-term gains by dynamically adjusting the pricing list. For smart community healthcare scenarios, this invention designs the following two pricing logics:

[0135] ① Uniform Pricing: The server sets the same unit price for all access tasks. At this point, the server's pricing list degenerates into:

[0136] ,in ;

[0137] ②Criticism-based Pricing: The server price is determined based on urgency level. A tiered pricing list is created based on urgency levels. :in, Indicates server The urgency level is The task is set to calculate the unit price. This scheme follows the logic of "paying a reasonable premium for high priority," that is, satisfying... Through this price signal, the server can implement differentiated resource allocation for tasks of different natures in IoMT.

[0138] Based on the above tendency values, the server The expected computational load is defined as Therefore, the utility function of Edge Server is expressed as:

[0139] ;

[0140] in It is a server The set price, Its operating costs. Through this design, the server agent can perceive the "micro-preferences" of all users, and thus adjust its strategies in multi-agent competition. Finding the optimal pricing at Nash equilibrium (NE) ensures the stable operation of the system as a whole in the highly dynamic environment of smart communities.

[0141] Dynamic Pricing and Offloading Decision Mechanism Based on MARL: This invention employs the MARL algorithm with integrated shared LSTM to achieve the coordinated evolution of Edge Server pricing strategy and UE offloading decision. This mechanism addresses the non-stationarity and incomplete information asymmetry issues in multi-agent environments, ensuring the system converges to a non-stationary (NE) state in complex game theory.

[0142] To achieve a system-level dynamic balance between server revenue and user utility, this invention employs the MARL framework.

[0143] Agent: Each Edge Server It is modeled as an independent intelligent agent.

[0144] Action space Agent's actions That is, its dynamic pricing output in the current time slot. This pricing strategy will be based on maximizing server utility. The objectives are adjusted in real time, reflecting the urgency of the task. The differentiation.

[0145] reward function Instant rewards for the agent For its server utility This reward function innovatively introduces a preference value. This incentivizes agents to improve QoS and understand users' potential choices while pursuing revenue, avoiding the limitations of traditional MARLs that only pursue short-term gains.

[0146] A forward-looking time-series modeling mechanism based on shared LSTM: Since user-side decisions are generated independently based on the current prices of each server, the server agent... Since real-time choices made by other users cannot be directly obtained, system load information is not fully observable during pricing decisions. To address this issue, this invention introduces a shared LSTM network into a multi-agent reinforcement learning model to perform time-series modeling of historical system observation sequences, thereby extracting potential load change features under incomplete information conditions.

[0147] 1) System observation state definition: in time slot The observed state of the system is represented as p(t−1) represents the pricing vector of each edge server in the previous time slot. This indicates the urgency distribution information of medical tasks in the current time slot. This indicates the load status of each server in the previous time slot.

[0148] 2) Implicit load characterization: Construct a historical observation sequence from the observation states of the most recent m time slots. The system is then fed into a shared LSTM network for temporal feature extraction to obtain the hidden state vector of the system in time slot t. When generating pricing strategies, the agent can base them on implicit states. The decision-making process implicitly predicts the user's preference for this server at the current moment. This resolves the issue of WD selection depending on instantaneous load in the Nash equilibrium proof. However, the contradiction that this load is not visible during the decision-making stage makes the pricing strategy forward-looking.

[0149] Continuous Action Space and Tiered Pricing Mapping Logic: Server Agent The policy network (Actor) outputs a continuous action vector. This action directly maps to the differentiated pricing list defined in Content 3. :

[0150] 1) Action execution: using the system's hidden state vector output by the shared LSTM network. As input, according to the Actor network Output .

[0151] 2) Balancing and adjustment mechanism: When predicting potential load When the computational capacity threshold is exceeded, the Actor network increases the price for high-urgency levels. Alternatively, the price difference between different levels can be increased to use price leverage to suppress the tendency of low-priority tasks to access the system, thus ensuring the maintenance of the user-side NE in an uncertain environment.

[0152] like Figure 2 The diagram illustrates the flow of the LSTM-MADDPG algorithm for task offloading and dynamic pricing in medical edge computing proposed in this invention. Starting from the input layer, the system simultaneously acquires user task information (such as task urgency and data volume) and server status information (such as current load, channel quality, and previous pricing) to form the current observation state. This state is input into a shared LSTM network to extract global temporal features and generate hidden vectors. Subsequently, the Actor network generates server pricing actions based on the hidden vectors, while the Critic network evaluates the value of the current strategy using the hidden vectors and action information. The generated actions are executed in the environment, and the system updates the state, calculates rewards, and records completion flags based on the action results. All observation, action, and reward information is stored in an experience replay buffer for training the Critic and optimizing the Actor, while sharing LSTM network weights to maintain consistency of global temporal features. After iterative training, the system finally outputs user task offloading decisions and server pricing strategies, realizing dynamic resource optimization and intelligent pricing in a medical edge computing environment.

[0153] This invention employs the MADDPG framework to address the non-stationarity problem caused by multi-agent competition through centralized training.

[0154] 1) Reward function design: Agent Instant rewards Server utility in Content 3 Hook it up and add a penalty:

[0155] ;

[0156] in The penalty coefficient ensures that the server can pursue profits while also considering the latency constraints of medical tasks.

[0157] 2) Centralized Critic Updates: Critic Network Obtain global information during training Its updates follow a loss function that minimizes the target Q value:

[0158] ;

[0159] in The Bellman value is calculated for the target network. Through this global perspective, each server agent can learn the pricing tendencies of other agents, thus achieving dynamic stability in a competitive environment.

[0160] Offloading Decision and Model Training Loop: The system completes pricing and offloading for each time slot through the following steps:

[0161] 1) Forecasting and Pricing: Servers Utilizing a local Actor network and LSTM to predict potential load, publish a price list for the current time slot. .

[0162] 2) User Response: WD devices respond based on all received price information and their own task urgency. Calculate the selection tendency value described in Content 3. and perform the optimal uninstallation action. .

[0163] 3) Load Accumulation and Learning: Actual Load Generation User satisfaction and server revenue are stored as environmental feedback in the experience replay pool. This drives the Agent to continuously optimize the prediction accuracy of users' Nash equilibrium selection behavior.

[0164] This invention categorizes tasks using a medical urgency model, constructs a two-sided utility framework based on Stackelberg game theory, and utilizes a multi-agent reinforcement learning algorithm with integrated shared LSTM to enable edge servers to learn differentiated dynamic pricing strategies that balance long-term gains with system load balancing. This invention ensures the stability and convergence of the game equilibrium through a propensity score selection mechanism; guarantees low latency for high-urgency tasks through tiered differentiated pricing; and enhances the system's decision-making foresight and robustness in dynamic environments through the temporal prediction capabilities of LSTM.

[0165] This invention, employing the above technical solutions, offers the following advantages compared to existing methods: 1) By constructing a bilateral game and learning framework between the server and users, this invention overcomes the dependence of traditional methods on a static environment, ensuring efficient and fair resource allocation in a highly dynamic and competitive market environment. 2) By introducing a differentiated pricing mechanism, this invention aims to address the technical deficiencies of existing technologies, such as the failure of pricing levers and the inability to provide deterministic resource guarantees for critical medical computing tasks. 3) By integrating a Long Short-Term Memory (LSTM) network into the state space of a MARL agent, this invention enables the agent to capture and predict the temporal characteristics of medical data. In this way, the agent can obtain more forward-looking state information, thereby adjusting pricing and resource allocation strategies in advance, improving the accuracy and timeliness of decision-making. Compared to existing methods that rely solely on instantaneous information for decision-making, this invention significantly enhances the system's adaptability and reliability to dynamic changes in medical scenarios, ensuring the continuity and timeliness of critical medical services.

[0166] This invention is particularly applicable to smart community healthcare scenarios, enabling efficient resource matching, load balancing, and low-latency assurance of medical information processing between community health edge servers and user devices for health monitoring tasks with varying degrees of medical urgency.

[0167] Obviously, the described embodiments are only a part of the embodiments of this application, not all of them. Without conflict, the embodiments and features in the embodiments of this application can be combined with each other. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the detailed description of the embodiments of this application is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

Claims

1. A method for offloading medical edge computing tasks and dynamic pricing, characterized by: Includes the following steps: Initialization: Configure the computing resources and communication parameters of the edge servers, and initialize the multi-agent reinforcement learning (MARL) model for each edge server; Task generation and broadcasting: The user equipment assesses the medical urgency of the collected physiological signal characteristics to obtain the task urgency, generates a medical computing task with the task urgency, and sends it to the scheduling and management platform. The scheduling and management platform broadcasts medical computing tasks to all edge servers; Pricing decision: Each edge server generates a differentiated pricing list based on the task urgency and the current system state using a MARL model containing a shared long short-term memory network (LSTM) and aggregates it to the scheduling management platform; the differentiated pricing list includes multiple price components corresponding to different task urgency. Offloading decision: Each user device calculates the utility of selecting different edge servers and local computing based on a differentiated pricing list of all edge servers distributed by the scheduling management platform. And according to utility Calculate the corresponding preference probability Finally, based on the probability of choice Make the final decision to uninstall ; Task execution: Under the coordination of the scheduling management platform, the user equipment offloads the medical computing task to the selected edge server or local computing according to the offloading decision; after the medical computing task is completed, the actual task latency and load are recorded, and the reward for each edge server is calculated; the processing results of the medical computing task are stored as experience data in the experience replay pool. Model update: The edge server randomly samples a batch of data from the experience replay pool to update the network parameters of the MARL model agent, in order to optimize the long-term benefits of the edge server and the overall system performance.

2. The medical edge computing task offloading and dynamic pricing method according to claim 1, characterized in that: Each user device generates a medical computing task based on the collected physiological signal characteristics. , is represented as: ,in, The amount of data for the task; To calculate density; The urgency level of the task.

3. The medical edge computing task offloading and dynamic pricing method according to claim 2, characterized in that: The urgency of the task is determined based on the medical crisis level to which the medical computing task belongs; the medical crisis level is a set of multiple discrete levels divided according to the degree of medical urgency, denoted as S. ; Task urgency The expression is obtained by weighted summation of the task levels, as follows: ； in, For normalized weights, satisfying and , used to quantify the relative severity of level s-th level medical emergency; As an indicator variable, when the current computing task of the i-th user device belongs to level s, then Otherwise, it is 0.

4. The medical edge computing task offloading and dynamic pricing method according to claim 1, characterized in that: The implementation of pricing decision-making steps includes: The agent of each edge server j collects the observation state of the current time slot. To characterize the dynamic changes of the system and observe its state Including the pricing of the previous time slot itself Information on the urgency distribution of user tasks in the current time slot and the actual load of the previous time slot The observation information from all servers is combined to form the global state. Used for value assessment during the centralized training phase; The observation sequence of the most recent time slots The input is fed into a shared LSTM network module to extract temporal features from historical states; the LSTM network uses a gating structure to perform temporal modeling of historical load changes, pricing strategies, and task urgency distributions, and outputs the system's global hidden state vector in time slot 𝑡. ; Hidden state of LSTM network output The input consists of an Actor network for each agent; the Actor network is a multi-layer feedforward neural network whose input is the global temporal features extracted by LSTM, and whose output is the server's pricing strategy. , This represents the strategy function for the j-th server. These are the parameters of the Actor network; The Actor network outputs a continuous action vector, which is then mapped to a pricing vector corresponding to the urgency level of the medical task. S represents the total number of urgency levels for the task; Each edge server sends its own pricing vector to the scheduling management platform, which then aggregates and forwards them to all user devices.

5. The medical edge computing task offloading and dynamic pricing method according to claim 1, characterized in that: The implementation of the unloading decision-making steps includes: Utility calculation: User device i calculates the utility of selecting different options k. The general utility function is defined as follows: ,in For indicator functions, For the collection of all edge servers, Pricing is set for each unit of data on edge server j. The amount of task data for user device i; For user device i, based on selection preference probability The final decision to uninstall was made; When k=0, it indicates local computation. =0, the indicator function takes the value of 0. The utility function simplifies to: ,in, Satisfaction with the quality of service for local computing. Calculate energy consumption locally; When k=j, it means the data is unloaded to edge server j. The indicator function takes the value of The utility function simplifies to ,in The service quality satisfaction of offloading user device i to edge server j is related to task latency. A decreasing function; The urgency of the task corresponding to edge server j Pricing; The amount of task data for user device i; Calculation of propensity to choose: Calculate the propensity to choose each option k based on the Logit model. , Where β is a sensitivity parameter used to adjust the user's sensitivity to differences in utility. To ensure model convergence, , This represents the maximum amount of user device task data in the system. The highest unit price allowed by the system. The minimum unit cost for edge servers; A set of options; The utility value of selecting the k-th option for the i-th user equipment; Decision execution: based on the probability of choice preference The final unloading decision will be made by randomly selecting or choosing the option with the highest probability. .

6. The medical edge computing task offloading and dynamic pricing method according to claim 5, characterized in that: User device i's satisfaction with the processing service quality of edge server j The calculation formula is: ， in For QoS satisfaction coefficient, The maximum tolerable delay for the task. Total task delay.

7. The medical edge computing task offloading and dynamic pricing method according to claim 1, characterized in that: Tuples in the task execution steps ( Store it in the experience replay pool.

8. The medical edge computing task offloading and dynamic pricing method according to claim 1, characterized in that: The implementation of the model update step includes: Centralized training: The scheduling and management platform uses the experience replay pool Global information x is obtained by randomly sampling small batches of historical experience data, and the global information x is used to update the Critic network parameters of the j-th edge server. The training objective of the Critic network is to minimize the temporal difference error, corresponding to the loss function. for: ； in, represents the experience replay pool, used to store historical experience tuples generated by all agents' interactions with the environment; x represents global observation information, including the status information of all edge servers and the task requirement information of all user devices; This represents the action chosen by the j-th edge server at the current moment, including the pricing strategy. And resource allocation decisions; N is the total number of edge servers; This represents the target Q-value generated by the target network calculated by the j-th edge server, used for stabilizing the training process. The calculation formula is: ,in, The instant reward obtained by the j-th edge server As a discount factor, Let x' represent the target Critic network, and x' represent the global information at the next time step. For target Actor network Based on local observations at the next moment The output action; This represents the combined action of global information x and all edge servers. The Q-value estimate of the Critic network output of the j-th edge server; Policy Update: Each edge server j updates the Actor network parameters using the gradient ascent method. To maximize the expected cumulative reward, the policy gradient is calculated using the following formula: ； in, This represents the Actor network of the j-th edge server. For input local observations, For output actions; Target network update: After each update of the Critic and Actor networks, the target network parameters are slowly updated using a soft update method, with the following update formula: ； ； in, and These represent the parameters of the target Critic network and the target Actor network, respectively. The soft update coefficient is much less than 1.

9. The medical edge computing task offloading and dynamic pricing method according to claim 1, characterized in that: Edge servers employ a load balancing mechanism: when predicting potential load... When the computational capacity threshold is exceeded, the MARL model agent's policy network (Actor) increases the price for high-urgency levels. Alternatively, the price difference between different levels could be increased to use price leverage to suppress the tendency of low-priority tasks to be accessed.

10. A medical edge computing task offloading and dynamic pricing system, executing the medical edge computing task offloading and dynamic pricing method according to any one of claims 1 to 9, characterized in that: The system includes: At least one user device is used to assess the medical urgency of the collected physiological signal features to obtain the task urgency, and generate a medical computing task with the task urgency. It also calculates the utility of selecting different edge servers and local computing based on the differentiated pricing list of all edge servers. And according to utility Calculate the corresponding preference probability Finally, based on the probability of choice Make the final decision to uninstall ; At least one edge server, each edge server is configured with a MARL model agent, the MARL model agent includes a shared long short-term memory network LSTM and an Actor-Critic network for generating differentiated pricing vectors; completes medical computing tasks offloaded from user devices, records actual task latency and load, and calculates rewards for the edge server; The results of medical computing tasks are stored as experience data in the experience playback pool. The scheduling and management platform is used to receive medical computing tasks from user devices and broadcast them to various edge servers, aggregate and forward pricing vectors from edge servers to user devices, and coordinate the allocation process of medical computing tasks. In this process, the edge server randomly samples a batch of data from the experience replay pool to update the network parameters of the MARL model agent, thereby optimizing the long-term benefits of the edge server and the overall performance of the system.

Images