Hybrid task unloading method for social Internet of Things

Abstract

The invention discloses a hybrid task unloading method oriented to the social Internet of Things, and relates to the field of task unloading of the social Internet of Things. According to the method, a multi-dimensional system model integrating tasks, transmission and calculation is constructed to accurately describe environmental dynamics and social attributes; on the basis, a task unloading process is formalized into a Markov decision process, and a mixed action space supporting three modes of local execution, full unloading and partial unloading is designed; an optimal unloading decision is adaptively generated through a deep reinforcement learning agent according to a real-time environment state, task characteristics and a candidate node set, the decision comprehensively considers transmission delay and node calculation and storage capacity, multi-objective optimization is achieved through a comprehensive reward function, and the optimal unloading decision is optimized through a multi-objective optimization algorithm. Task unloading challenges caused by isomerism, dynamics and social attributes in a social internet of things environment are effectively solved, and task scheduling with low delay, high reliability and multi-objective optimization is realized.

Description

Technical Field

[0001] This invention relates to the field of task offloading in the Internet of Things (IoT), specifically a hybrid task offloading method for the IoT. Background Technology

[0002] With the rapid evolution of cutting-edge technologies such as 5G, IoT, and AI, new application scenarios that are sensitive to latency and computationally intensive, such as smart healthcare, autonomous driving, and intelligent manufacturing, are constantly emerging. These applications not only place higher demands on the scheduling flexibility of underlying computing resources, but also require low-latency, high-reliability task processing in dynamic network environments. The centralized data processing architecture of traditional cloud computing can hardly meet these requirements. The high latency and high bandwidth consumption caused by its long-distance data transmission characteristics have become the core bottleneck restricting application performance. Against this backdrop, edge computing and fog computing technologies have emerged. By moving computing and storage resources from cloud data centers to the network edge (such as base stations and edge servers) or fog nodes (distributed nodes between the edge and the cloud), they significantly shorten data transmission distances, alleviate network bandwidth pressure, and provide technical support for low-latency applications.

[0003] However, in the new paradigm of Social Internet of Things (SIoT), which integrates social attributes and IoT connectivity, the aforementioned technologies still face multiple challenges. Social IoT treats smart devices as entities with autonomous social attributes. Devices can establish social relationships based on user habits, spatial adjacency, and historical interaction records, forming a dynamic network of resource sharing and task collaboration. However, the high heterogeneity and dynamism of this environment render traditional task offloading methods problematic. On the one hand, there are significant differences in computing power, storage capacity, and battery life among terminals such as smart sensors and wearable devices, and the resource allocation of edge nodes, fog nodes, and cloud nodes also exhibits multi-layered differentiation. On the other hand, device mobility, network link quality fluctuations, and the suddenness of task loads further exacerbate the complexity of resource scheduling. Traditional task offloading strategies are mainly divided into static offloading and dynamic offloading, both of which have obvious limitations. Static offloading mechanisms rely on preset rules or fixed topology models to allocate tasks (such as uniformly offloading all computationally intensive tasks to edge nodes), and cannot perceive real-time changes in network status and device resources. When edge nodes become overloaded due to sudden surges in tasks, static strategies continue to allocate tasks, leading to a sharp increase in task queuing latency and even triggering task failures. While dynamic offloading mechanisms can adjust offloading decisions based on parameters such as real-time network bandwidth, latency, task computation volume, and latency thresholds, the decision-making process requires additional computing resources to collect and analyze real-time data, potentially introducing decision-making delays. Furthermore, dynamically selecting offloading targets can easily overlook the social trust relationships between devices; the reliability risks of unfamiliar nodes reduce data transmission security, and frequent switching of offloading nodes can cause network congestion, ultimately worsening task processing performance.

[0004] Currently, researchers both domestically and internationally have conducted relevant research on task offloading methods. Shu Xu, Qingjie Liu, Chengye Gong, Xupeng Wen, et al., in their paper "Energy-efficient multi-agent deepreinforcement learning task offloading and resource allocation for UAV edge computing" [Sensors, 2025, 25(11):3403], proposed a multi-agent reinforcement learning algorithm to optimize task offloading and resource allocation, thereby enhancing mobile edge computing systems. This system reduces task processing latency and UAV energy consumption. Sulimani H., Sulimani R., Ramezani F., et al., in their paper "HybOff: a HybridOffloading approach to improve load balancing in fog environments" [Journal of Cloud Computing, 2024, 13(1):113], proposed a hybrid offloading algorithm combining static and dynamic offloading strategies. This algorithm effectively achieves load balancing while reducing latency and cost, improving system resource utilization. In their paper "An expert system for hybrid edge to cloud computational offloading in heterogeneous MEC–MCC environments" [Journal of Network and Computer Applications, 2024, 225:103867], Sheharyar Khan, Zheng Jiangbin, Muhammad Irfan, and others designed an artificial neural network model for optimizing the offloading decision of computing tasks in heterogeneous mobile edge computing scenarios. This model achieves high offloading accuracy and effectively reduces latency and energy consumption.In the field of social networks, Jie Tang's article "Computational Models for Social Network Analysis: A Brief Survey" [Proceedings of the 26th International Conference on World Wide WebCompanion, Geneva, Swiss, 2017] reviews the importance of social networks, model construction, and the latest research results, providing profound methodological guidance for understanding and modeling large-scale dynamic social networks.

[0005] However, while some solutions focusing on multi-agent systems optimize offloading strategies through reinforcement learning, they are mostly designed for specific scenarios and fail to adapt to the social attributes of devices in the social IoT. This makes it difficult to support collaborative offloading of large-scale heterogeneous devices (such as terminals, edge computing, and fog nodes), resulting in insufficient generalization ability of the strategies in real-world social IoT environments. Furthermore, while existing hybrid offloading strategies (such as methods combining static clustering and dynamic decision-making) can improve load balancing efficiency in fog computing environments, they exhibit poor stability under high loads and concurrent requests from multiple tasks. They also fail to fully consider the divisibility of tasks, making it difficult to flexibly allocate resource consumption between local computing and remote nodes through partial offloading modes, thus hindering the balance of load pressure across different nodes. In addition, most solutions only optimize a single metric, failing to comprehensively consider multi-dimensional requirements such as task success rate, queue latency, and node storage constraints, thus failing to meet the diverse service quality requirements of complex applications in the social IoT, such as smart healthcare and autonomous driving.

[0006] Therefore, a new solution is needed to address the above problems. Summary of the Invention

[0007] The purpose of this invention is to provide a hybrid task offloading method for the social Internet of Things (IoT). In the heterogeneous and dynamic environment of the social IoT, it breaks through the limitations of traditional offloading methods, integrates the social attributes of devices, supports multi-mode offloading (local execution, full offloading, partial offloading), and achieves a multi-objective optimized task offloading scheme through intelligent decision-making. The problem it solves is that in the highly heterogeneous and dynamic social IoT environment, traditional static offloading strategies cannot adapt to real-time changes, while dynamic offloading strategies have defects such as decision delay, network congestion, and ignoring the social attributes of devices, making it difficult to achieve efficient, reliable and low-latency task allocation.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a hybrid task offloading method for the social Internet of Things, comprising at least the following steps:

[0009] S1: Construct a multi-dimensional system model, which includes a task model, a transmission model, and a computation model. The task model is used to describe the metadata of the task, the transmission model is used to calculate the latency of task transmission between nodes, and the computation model is used to calculate the processing latency of task on the node.

[0010] S2: Build a decision model based on deep reinforcement learning, formalize the task unloading process into a Markov decision process, and define the state space, mixed action space and reward function.

[0011] S3: Execute adaptive unloading decisions.

[0012] Furthermore, the metadata of the task model includes the amount of input data, the total number of computation cycles required, the amount of output data, and the maximum acceptable latency.

[0013] The transmission model calculates transmission delay and propagation delay based on the Euclidean distance between nodes;

[0014] The computational model defines the total processing latency of a task on a node as the sum of queuing latency and execution latency, where the execution latency is calculated based on the total number of CPU cycles required by the task and the node's CPU frequency.

[0015] Furthermore, the decision model of the deep reinforcement learning adopts a hierarchical neural network structure;

[0016] The hierarchical neural network structure includes an upper layer network, a middle layer network, and a lower layer network;

[0017] The upper-layer network is used to determine the unloading type based on the input state;

[0018] The middle-layer network is used to decide on unloading the target node based on the concatenated information of the decision results of the upper-layer network and the input state;

[0019] The lower-layer network is used to determine the unloading ratio based on the spliced ​​information of the decision results of the upper-layer and middle-layer networks and the input state.

[0020] Furthermore, the state space is constructed based on a multi-dimensional system model, and the state space includes task characteristics, node characteristics, and network link characteristics.

[0021] Furthermore, the hybrid action space is a triple action vector.

[0022]

[0023] in, This is a discrete decision, indicating whether to unload (0 for local execution, 1 for remote unloading). This is a continuous parameter representing the percentage of tasks that are unloaded. For discrete decision-making, let represent the selected target node;

[0024] By using different combinations of triples, the hybrid action space includes three task execution modes: local execution, full uninstallation, and partial uninstallation.

[0025] Furthermore, the reward function is designed according to different execution modes to comprehensively optimize transmission latency, node computing power adaptability, and node storage adaptability.

[0026] The reward function is designed to correspond to the three task execution modes in the hybrid action space;

[0027] The reward calculation formula is as follows:

[0028]

[0029] For partial uninstallation, the reward calculation formula is as follows:

[0030]

[0031] For local execution, the reward calculation formula is as follows:

[0032]

[0033] in, These are normalized weighting coefficients used to balance multiple optimization objectives; For transmission delay; These are the available computing capabilities of the target node and the local node, respectively. These are the remaining storage for the target node and the local node, respectively. These are the computational and storage requirements of the task, respectively. This is an indicator function; its value is 1 when the condition is true, and 0 otherwise. This is a penalty for task timeout or uninstallation failure.

[0034] Furthermore, S3 includes at least the following steps:

[0035] Real-time acquisition of system status information and characteristics of tasks to be processed;

[0036] Based on the collected information, a set of qualified target candidate nodes is selected;

[0037] The current state and candidate node set are input into the trained deep reinforcement learning model to generate a hybrid action that includes unloading type, unloading ratio and unloading target;

[0038] The task is unloaded according to the aforementioned hybrid action;

[0039] Collect performance metrics after task unloading, calculate rewards, and use this experience to update the model for continuous optimization.

[0040] Furthermore, the system status information includes social attribute information between devices, which includes the collaboration history, trust relationship, and spatial proximity between devices.

[0041] Compared with the prior art, the beneficial effects of the present invention are:

[0042] This invention proposes to construct a multi-dimensional system model integrating tasks, transmission, and computation to accurately describe the dynamics and social attributes of the environment. Based on this, the task unloading process is formalized as a Markov decision process, and a hybrid action space supporting three modes—local execution, full unloading, and partial unloading—is designed. Through a deep reinforcement learning agent, the optimal unloading decision is adaptively generated based on the real-time environmental state, task characteristics, and candidate node set. This decision comprehensively considers transmission latency, node computation and storage capabilities, and achieves multi-objective optimization through a comprehensive reward function. This effectively solves the task unloading challenges brought about by heterogeneity, dynamism, and social attributes in the social IoT environment, and achieves low-latency, high-reliability, and multi-objective optimized task scheduling. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a diagram illustrating the construction of the social Internet of Things model of the present invention;

[0045] Figure 2 This is a representation of the attribute features of nodes and edges in this invention;

[0046] Figure 3 This is a framework diagram of the hybrid unloading algorithm of the present invention;

[0047] Figure 4 This is a flowchart of the hybrid unloading process of the present invention. Detailed Implementation

[0048] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0049] Currently, in the heterogeneous and dynamic environment of the social Internet of Things (IoT), traditional task offloading methods (static offloading mechanisms cannot adapt to real-time network and resource changes, while dynamic offloading mechanisms suffer from implicit costs such as decision delays and network congestion) struggle to achieve efficient task allocation. Existing optimization solutions either fail to adapt to the social attributes of devices, lack the ability to collaborate on large-scale heterogeneous devices, or suffer from insufficient stability under high-load scenarios and do not fully support task divisibility. Moreover, most of them only optimize a single objective and cannot comprehensively balance multiple dimensions such as task success rate, queue latency, and node storage constraints, making it difficult to meet the service quality requirements of latency-sensitive and computationally intensive applications. To address this problem, this invention provides a hybrid task offloading method for the social IoT, which achieves intelligent and adaptive task offloading decisions through deep reinforcement learning, thereby overcoming the limitations of traditional methods and improving the overall system performance.

[0050] The technical solution is as follows:

[0051] (1) Multi-dimensional system model construction. The task, transmission and computing processes in the social Internet of Things are comprehensively modeled, taking into full account the heterogeneity, dynamism and social attributes of the environment. Task modeling includes metadata such as input data volume, required computing cycle, output data volume and maximum acceptable delay. Transmission modeling calculates transmission delay and propagation delay based on Euclidean distance between nodes. Computation modeling includes the queuing delay and execution delay of tasks at nodes. The queuing model adopts a first-in-first-out (FIFO) mechanism to manage the task queue. The execution delay is calculated based on the CPU cycles required by the task and the CPU frequency of the node. The total computation delay is defined as the sum of queuing delay and execution delay, providing a theoretical basis for subsequent offloading decisions.

[0052] (2) Design a deep reinforcement learning model based on Markov decision processes. The task unloading process is formalized as an MDP problem. The defined task characteristics include: input data volume, computational load, and maximum tolerance time; the node characteristics include: computational resources, remaining storage, and queue length. A hybrid action space integrating discrete decision-making and continuous parameter adjustment is designed to support three modes: local execution (the task is completed on the initiating device), full unloading (the entire task is migrated to a remote node), and partial unloading (the task is divided and completed collaboratively by local and remote nodes). The action vector specifically includes three dimensions: whether to unload, unloading ratio, and unloading target node. For the three different modes, reward functions are designed respectively, comprehensively balancing transmission latency, node computing power adaptability, and node storage adaptability, and introducing timeout or unloading failure penalties. At the same time, a discount factor is used to balance short-term immediate rewards and long-term performance optimization.

[0053] (3) Implement an adaptive offloading decision-making process. The decision-making process includes six key stages: 1) Real-time collection of the operating status and social attribute information of terminals, edge devices, fog devices, and cloud devices. 2) Analysis of the key features of the tasks to be processed to determine their offloading adaptability and divisibility. 3) Filtering nodes with scarce resources, unstable networks, or low trust based on the information from the first two stages to determine the target node set that meets the service quality constraints. 4) Inputting the status, task features, and candidate node set into the trained reinforcement learning model to generate the optimal offloading strategy. 5) Encrypting and transmitting task data according to the decision results, coordinating the concurrent execution of subtasks and the aggregation of results. 6) Collecting performance indicators such as task completion latency, resource consumption, and success rate, calculating reward signals, updating the policy network and value network of deep reinforcement learning, and achieving continuous iterative optimization.

[0054] Specifically:

[0055] I. Information Collection

[0056] The system architecture described in this invention is deployed in a social Internet of Things (IoT) environment, and consists of three heterogeneous devices: cloud, edge (including fog), and terminals. The cloud refers to one or more centralized big data centers with powerful computing and storage capabilities; the edge consists of servers deployed at network access points (such as base stations), responsible for providing near-field computing services; and terminals encompass various intelligent devices such as personal computers, smart cars, robots, smartphones, and drones. These devices initiate computing tasks but are often limited by their own computing, storage, and energy capabilities. Physically, cloud and edge devices are interconnected via a high-speed wired backbone network, while terminal devices access the network via wired or wireless means. All these devices are interconnected, forming a large-scale, dynamic IoT network that continuously transmits data and interacts with tasks.

[0057] It is worth noting that some terminal devices in the network (such as smart cars and drones) are mobile, and their location changes or dynamic switching of communication channels cause the network topology to evolve continuously over time. To accurately characterize this dynamic environment, this invention perceives the system state through continuous information collection. Specifically, the system periodically or based on event triggers collects three types of key parameters from the network: First, node parameters, collecting attributes such as computing power (represented by CPU frequency), remaining storage capacity, current task queue length, geographical location, and remaining battery power for each device that can act as a task processor (including terminals, edge servers, and cloud centers); second, edge parameters, collecting transmission characteristics such as available bandwidth, transmission latency, and physical distance calculated based on geographical location for communication links between nodes; and third, task parameters, parsing metadata such as the input data size and the total number of computation cycles required for each computation task to be processed generated by the terminal. In addition, historical / trust information generated by the interaction and collaboration between devices will also be recorded.

[0058] To transform the complex physical network described above into a form that can be processed by algorithms, this invention abstracts the entire social Internet of Things into a weighted graph model. In this model, the set of nodes Each vertex in the array corresponds to a physical device and is accompanied by collected node parameters as its features; the edge set Each edge in the graph model represents a communication link between nodes and is accompanied by collected edge parameters as its features. Tasks generated by a node are considered dynamic entities attached to that node, and their task parameters constitute the task's own feature vector. This model integrates the network's topology, resource status, link conditions, and task load; all its encapsulated features collectively constitute the input state space for subsequent deep reinforcement learning algorithms. As time progresses and the system operates, this model will be continuously updated to ensure it accurately reflects the network's real-time dynamics. For clarity, this specification provides... Figure 2 The (feature table) lists in detail the collected feature parameters and their definitions.

[0059] II. Model Construction

[0060] Before calculating transmission delay, the physical distance between nodes must first be determined. Let nodes... With nodes The coordinates are respectively and The physical distance between two points is measured using Euclidean distance:

[0061]

[0062] distance This serves as the basis for subsequent transmission delay modeling. The task originates from the node. Transmit to node Total delay This includes data transmission time and inherent link latency, and is defined by the following formula:

[0063]

[0064] in, The input data size for the task (MB). For the node To the node The bandwidth (Mbps), multiplied by 8, is converted from bytes to bits. The inherent latency of the link (ms) is related to the physical distance. Related, specifically:

[0065]

[0066] in, The speed at which a signal travels through a medium (m / ms) is denoted as . Other processing delays caused by routing, queuing, etc.

[0067] Task on compute node Total processing latency , which is the sum of queuing delay and execution delay:

[0068]

[0069] The execution latency is determined by both the task's computational requirements and the node's computational capabilities, and the calculation formula is as follows:

[0070]

[0071] in, This represents the total number of CPU cycles required for the task. For nodes CPU frequency (GHz), multiplied by It converts frequency units into cycles per second.

[0072] This algorithm follows a Markov decision process, and the system state perceived by the deep reinforcement learning agent is... Defined by the following vector:

[0073]

[0074] Specific features such as Figure 2 The parameters are shown in the table. The actions output by this deep reinforcement learning agent are... It is a mixed vector, defined as a triple:

[0075]

[0076] in, For discrete decisions, it indicates whether to unload (0 for local execution, 1 for remote unloading). This is a continuous parameter representing the percentage of tasks that are unloaded. For discrete decision-making, let represent the selected target node. To guide the agent in learning the optimal policy, differentiated reward functions are designed for different unloading modes. For the complete unloading mode, the reward is... The calculation is as follows:

[0077]

[0078] For certain uninstallation modes, rewards are provided. The calculation is as follows:

[0079]

[0080] For local execution mode, rewards The calculation is as follows:

[0081]

[0082] in, These are normalized weighting coefficients used to balance multiple optimization objectives; For transmission delay; These are the available computing capabilities of the target node and the local node, respectively. These are the remaining storage for the target node and the local node, respectively. These are the computational and storage requirements of the task, respectively. This is an indicator function; its value is 1 when the condition is true, and 0 otherwise. This is a penalty for task timeout or uninstallation failure.

[0083] III. Hybrid Offloading Algorithm Based on Deep Reinforcement Learning

[0084] The algorithm formalizes the dynamic task unloading problem as a Markov decision process, which explicitly defines the state space, action space, state transition probabilities, and reward function (Equations 6-10). The algorithm operates as a closed-loop optimization process of perception, decision-making, and learning. When a new computational task is generated in the system, the algorithm first enters the perception phase, collecting and constructing state vectors through monitoring modules deployed in various locations. All the necessary real-time data. Then, the system enters the decision-making phase, displaying the current state. The input is fed into a pre-trained deep reinforcement learning model. This deep reinforcement learning algorithm consists of three layers of neural networks: upper, middle, and lower (e.g., ...). Figure 3 (As shown). First, The input is fed into the topmost neural network layer to obtain the unloading type. Then, the results from the upper layers are concatenated with the state. The data is then input into the middle layer to obtain the target node for unloading. Finally, the results from the upper layer, the middle layer, and the status are compared. These are concatenated together as input for the next layer to obtain the unloading ratio. The unloading type, unloading target, and unloading ratio are combined to form the strategy. .

[0085] The model is based on the policy it has learned. Generate a blend action This action explicitly specifies the execution mode, unloading ratio, and target node for this task. After the decision is generated, the system enters the execution phase. The task will be scheduled for execution based on the decision result: if it is to be executed locally, the task enters the local queue; if it is to be completely unloaded, the task data is completely transmitted to the selected node; if it is to be partially unloaded, the task is precisely divided, with part executed locally and part transmitted to a remote node for collaborative processing. After the task is completed, the system enters the crucial learning and feedback phase. The environment will calculate a reward signal. This signal quantifies the merits of the decision; low latency and high success rate bring positive rewards, while task failure or timeout brings negative rewards. Simultaneously, the system observes new environmental states. This complete interactive experience The experience is stored in an experience replay buffer. During training, the agent periodically draws a batch of historical experiences randomly from the buffer and updates its policy network and value network using the gradient descent algorithm, thereby changing its decision-making policy. This allows for continuous improvement.

[0086] Example 1:

[0087] A hybrid task offloading method for the social Internet of Things includes at least the following steps:

[0088] S1: Construct a multi-dimensional system model, which includes a task model, a transmission model, and a computation model. The task model is used to describe the metadata of the task, the transmission model is used to calculate the latency of task transmission between nodes, and the computation model is used to calculate the processing latency of task on the node.

[0089] S2: Build a decision model based on deep reinforcement learning, formalize the task unloading process into a Markov decision process, and define the state space, mixed action space and reward function.

[0090] S3: Execute adaptive unloading decisions.

[0091] The metadata of the task model includes the amount of input data, the total number of computation cycles required, the amount of output data, and the maximum acceptable latency.

[0092] The transmission model calculates transmission delay and propagation delay based on the Euclidean distance between nodes;

[0093] The computational model defines the total processing latency of a task on a node as the sum of queuing latency and execution latency, where the execution latency is calculated based on the total number of CPU cycles required by the task and the node's CPU frequency.

[0094] The decision-making model of deep reinforcement learning adopts a hierarchical neural network structure;

[0095] A hierarchical neural network structure includes an upper layer network, a middle layer network, and a lower layer network;

[0096] The upper-layer network is used to determine the unloading type based on the input state;

[0097] The middle-layer network is used to decide on the unloading of the target node based on the concatenated information of the decision results of the upper-layer network and the input state;

[0098] The lower-level network is used to determine the unloading ratio based on the concatenated information of the decision results of the upper-level and middle-level networks and the input state.

[0099] The state space is constructed based on a multi-dimensional system model, and includes task characteristics, node characteristics, and network link characteristics.

[0100] The hybrid action space is a triple action vector

[0101]

[0102] in, This is a discrete decision, indicating whether to unload (0 for local execution, 1 for remote unloading). This is a continuous parameter representing the percentage of tasks that are unloaded. For discrete decision-making, let represent the selected target node;

[0103] By using different combinations of triples, the hybrid action space supports three task execution modes: local execution, full uninstallation, and partial uninstallation.

[0104] The reward function is designed according to different execution modes to comprehensively optimize transmission latency, node computing power adaptability and node storage adaptability;

[0105] The reward function is designed to correspond to the three task execution modes in the hybrid action space;

[0106] For a complete uninstallation, the reward calculation formula is as follows:

[0107]

[0108] For partial uninstallation, the reward calculation formula is as follows:

[0109]

[0110] For local execution, the reward calculation formula is as follows:

[0111]

[0112] in, These are normalized weighting coefficients used to balance multiple optimization objectives; For transmission delay; These are the available computing capabilities of the target node and the local node, respectively. These are the remaining storage for the target node and the local node, respectively. These are the computational and storage requirements of the task, respectively. This is an indicator function; its value is 1 when the condition is true, and 0 otherwise. This is a penalty for task timeout or uninstallation failure.

[0113] S3 includes at least the following steps:

[0114] Real-time acquisition of system status information and characteristics of tasks to be processed;

[0115] Based on the collected information, a set of qualified target candidate nodes is selected;

[0116] The current state and candidate node set are input into the trained deep reinforcement learning model to generate a hybrid action that includes unloading type, unloading ratio and unloading target;

[0117] Task unloading is performed based on mixed actions;

[0118] Collect performance metrics after task unloading, calculate rewards, and use this experience to update the model for continuous optimization.

[0119] System status information includes social attribute information between devices, which includes the collaboration history, trust relationship and spatial proximity between devices.

[0120] Example 2:

[0121] Based on Embodiment 1, an execution system equipped with a hybrid task offloading method is proposed. The specific process of this system is a continuously running, self-optimizing closed-loop process (e.g., Figure 4 (As shown), the detailed steps are as follows:

[0122] 1. The process begins with real-time awareness of the system environment. A perception module deployed at the network edge is activated to collect global status information. This information includes two main categories: first, physical resource status, covering the current computing resource utilization, remaining storage capacity, battery power, and network link quality (such as bandwidth and latency) of all terminal devices, edge servers, and fog nodes; second, social attribute information, including the collaboration history, trust relationships, and spatial proximity between devices, which is collected as contextual auxiliary data.

[0123] 2. When a terminal device generates a computational task to be processed, the system immediately parses the task. This stage aims to extract key metadata about the task, including but not limited to the size of the input data, the total number of CPU cycles required to complete the computation, the size of the result data, the maximum tolerable processing latency, and to determine whether the task can be decomposed into subtasks for parallel processing. These characteristic parameters are the fundamental basis for determining which offloading mode is suitable for the task.

[0124] 3. Based on the information obtained in stages 1 and 2, the system performs candidate node screening. Its key feature is that, according to preset service quality constraints, the system filters out nodes with limited computing resources, unstable network connections, or excessively low trust ratings, thereby forming a reliable set of target candidate nodes that meet the requirements of the current task. This set defines the feasible range of actions for subsequent intelligent decision-making.

[0125] 4. The system inputs the current environment state into a pre-trained deep reinforcement learning model. Based on its learned optimal policy, the DRL model outputs a hybrid action vector. This action vector explicitly specifies the unloading type for this task.

[0126] 5. Based on the decision result of step 4, determine the following: If the process is handled by a local device, the target node and the unloading ratio will not be relevant. If the unloading is done remotely, then the target node and the unloading ratio will be available.

[0127] 6. Based on predetermined decisions, the system coordinates the transmission of tasks, their execution on local or remote nodes, and manages the aggregation of subtask results to ensure data processing consistency and time synchronization.

[0128] 7. After task completion, the system enters the learning and optimization phase. The system collects the actual performance metrics of the task, including total latency, resource consumption, and success / failure status. These metrics are used to calculate the reward signal. Subsequently, the system determines if a new task has arrived: if so, the experience tuples from this interaction are stored in the experience replay buffer, triggering the feedback and learning process. The system samples a batch of historical data from the buffer and updates the policy and value network parameters of the DRL model using the gradient descent algorithm, thereby optimizing its decision-making ability. The updated DRL model will be used to process the next task, and the process jumps to phase 1. If no new task arrives, the process ends.

[0129] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A hybrid task offloading method for the social Internet of Things, characterized in that: At least the following steps are included: S1: Construct a multi-dimensional system model, which includes a task model, a transmission model, and a computation model. The task model is used to describe the metadata of the task, the transmission model is used to calculate the latency of task transmission between nodes, and the computation model is used to calculate the processing latency of task on the node. S2: Build a decision model based on deep reinforcement learning, formalize the task unloading process into a Markov decision process, and define the state space, mixed action space and reward function. S3: Execute adaptive unloading decisions.

2. The hybrid task offloading method for the social Internet of Things according to claim 1, characterized in that: The metadata of the task model includes the amount of input data, the total number of computation cycles required, the amount of output data, and the maximum acceptable latency. The transmission model calculates transmission delay and propagation delay based on the Euclidean distance between nodes; The computational model defines the total processing latency of a task on a node as the sum of queuing latency and execution latency, where the execution latency is calculated based on the total number of CPU cycles required by the task and the node's CPU frequency.

3. The hybrid task offloading method for the social Internet of Things according to claim 1, characterized in that: The deep reinforcement learning decision model adopts a hierarchical neural network structure; The hierarchical neural network structure includes an upper layer network, a middle layer network, and a lower layer network; The upper-layer network is used to determine the unloading type based on the input state; The middle-layer network is used to decide on unloading the target node based on the concatenated information of the decision results of the upper-layer network and the input state; The lower-layer network is used to determine the unloading ratio based on the spliced ​​information of the decision results of the upper-layer and middle-layer networks and the input state.

4. The hybrid task offloading method for the social Internet of Things according to claim 1, characterized in that: The state space is constructed based on a multi-dimensional system model, and the state space includes task characteristics, node characteristics, and network link characteristics.

5. A hybrid task offloading method for the social Internet of Things according to claim 3, characterized in that: The hybrid action space is a triple action vector. ； in, For discrete decisions, indicating whether to unload; This is a continuous parameter representing the percentage of tasks that are unloaded. For discrete decision-making, let represent the selected target node; By using different combinations of triples, the hybrid action space includes three task execution modes: local execution, full uninstallation, and partial uninstallation.

6. A hybrid task offloading method for the social Internet of Things according to claim 4, characterized in that: The reward function is designed according to different execution modes to comprehensively optimize transmission latency, node computing power adaptability and node storage adaptability. The reward function is designed to correspond to the three task execution modes in the hybrid action space; For a complete uninstallation, the reward calculation formula is as follows: ； For partial uninstallation, the reward calculation formula is as follows: ； For local execution, the reward calculation formula is as follows: ； in, These are normalized weighting coefficients used to balance multiple optimization objectives; For transmission delay; These are the available computing capabilities of the target node and the local node, respectively. These are the remaining storage for the target node and the local node, respectively. These are the computational and storage requirements of the task, respectively. This is an indicator function; its value is 1 when the condition is true, and 0 otherwise. This is a penalty for task timeout or uninstallation failure.

7. A hybrid task offloading method for the social Internet of Things according to claim 1, characterized in that: The S3 includes at least the following steps: Real-time acquisition of system status information and characteristics of tasks to be processed; Based on the collected information, a set of qualified target candidate nodes is selected; The current state and candidate node set are input into the trained deep reinforcement learning model to generate a hybrid action that includes unloading type, unloading ratio and unloading target; The task is unloaded according to the aforementioned hybrid action; Collect performance metrics after task unloading, calculate rewards, and use this experience to update the model for continuous optimization.

8. A hybrid task offloading method for the social Internet of Things according to claim 7, characterized in that: The system status information includes social attribute information between devices, which includes the collaboration history, trust relationship and spatial proximity between devices.

Images