Task-driven hierarchical heterogeneous network resource management data offloading method

Abstract

The application provides a task-driven hierarchical heterogeneous network resource management data unloading method; comprising: modeling the multi-dimensional resource management and task scheduling problem of a heterogeneous multi-access MEC network; determining the priority of task processing, and adopting a data unloading strategy for task scheduling and data unloading processing; after the intelligent reactive jammer detects the data unloading of the heterogeneous user, the jammer adopts the tracking jamming or dynamic jamming mode to hinder the data transmission unloading behavior; the internal users of the heterogeneous multi-access MEC network adjust the respective association matching relationship with the wireless access node / base station; the heterogeneous user adjusts the respective association strategy and unloading parameter according to the perceived jammer behavior, the transmission demand of the own task and the data unloading behavior of other users. The model is complete, reasonable and effective.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wireless communication, in particular to a task-driven hierarchical heterogeneous network resource management data offloading method. BACKGROUND

[0002] With the rapid development of hardware devices and communication technologies, the communication and perception capabilities of mobile users are continuously improved, and the number of Internet of Things devices is continuously growing. Based on the delay-sensitive and computationally-intensive mobile applications, such as autonomous driving, augmented reality, etc., the overall demand situation of wireless communication network presents exponential growth. These applications usually need to occupy a large amount of computing resources. However, the limited computing capacity of mobile devices becomes a key factor restricting their widespread deployment and efficient operation. MEC as a new network architecture aims to solve the above challenges. By offloading computing tasks from mobile devices to adjacent base stations (BSs) or wireless access points (APs), MEC technology effectively improves the utilization of computing resources, which not only can alleviate the computing load of the core network, but also can meet the strict delay response requirements in the task.

[0003] Although MEC architecture has significant advantages in solving mobile devices, it also has certain limitations, specifically: the computing power of a single node is limited, making it difficult to cope with the concurrent service requests of a large number of mobile devices; moreover, the openness of the wireless channel, the diversity of task requirements, and the mobility of devices result in constant changes in mobile user access requests, frequent adjustments of network topology structure and associated relationships (references: Zhang Y, Kishk MA, Alouini M S. Computation offloading and service caching in heterogeneous MEC wireless networks[J]. IEEE Transactions on Mobile Computing, 2021, 22(6): 3241-3256. references: Wang P, Di B, Song L, et al. Multi-layer computation offloading in distributed heterogeneous mobile edge computing networks[J]. IEEE Transactions on Cognitive Communications and Networking, 2022, 8(2): 1301-1315.), which easily causes local congestion of network links, and further causes load imbalance of edge servers, significantly reducing the overall performance of the system and the service experience of users. In addition, in complex Internet of Things scenarios, the user heterogeneity caused by significant differences in mobile user types and hardware configurations further exacerbates the unreasonable allocation of multi-dimensional resources (such as computing resources, storage resources, communication resources, etc.) in MEC networks. In addition, in a complex electromagnetic spectrum environment, the intelligent level of malicious interference technology is constantly improving, and its interference mode tends to be diversified. Intelligent reactive jammers (IRJ) with sensing capabilities pose a more serious threat to users in heterogeneous MEC systems.However, most existing resource allocation schemes based on reinforcement learning (references: Lv Z, Xiao L, Du Y, et al. Efficient Communications in Multi-Agent Reinforcement Learning for Mobile Applications [J]. IEEE Transactions on Wireless Communications, 2024. references: Wang P, Zheng Z, Di B, et al. HetMEC: Latency-optimal task assignment and resource allocation for heterogeneous multi-layer mobile edge computing [J]. IEEE Transactions on Wireless Communications, 2019, 18(10): 4942-4956.) mainly focus on fixed network topology scenarios, ignoring the rapid dynamic changes of network topology and the QoS requirements of heterogeneous tasks, and lack of in-depth research on the adaptability and effectiveness of multi-dimensional resource allocation management methods in heterogeneous MEC networks. This leads to problems such as load imbalance and link congestion when edge computing servers face sudden requests and data volume surges in practical applications, severely limiting the performance of intelligent learning algorithms. Therefore, there is an urgent need to research new multi-agent collaborative anti-interference data offloading algorithms to solve the above problems. SUMMARY

[0004] The present application provides a task-driven hierarchical heterogeneous network resource management data offloading method, which can be used to solve the technical problem that the specific needs of different users in handling heterogeneous tasks in terms of priority, security, delay and energy consumption are not met in heterogeneous MEC networks.

[0005] The task-driven hierarchical heterogeneous network resource management data offloading method comprises the following steps:

[0006] Step 1, model the multi-dimensional resource management and task scheduling problem based on task driving in the heterogeneous MEC network; describe it as a partially observable Markov game process, the participants of the game are heterogeneous users in the heterogeneous network, and joint task scheduling, dynamic spectrum access and dynamic computing resource management are optimized for distributed data offloading decisions;

[0007] Step 2, the heterogeneous users in the heterogeneous network design the reward weighting value according to the queuing information of their own cache area tasks and the task related information including the task demand of time delay demand and information age, determine the priority of task processing, and adopt appropriate edge computing decision to process data;

[0008] After the intelligent reactive jammer detects the data offloading behavior of the heterogeneous user, it adopts the tracking jamming or dynamic jamming mode to hinder the data transmission offloading behavior;

[0009] Step 3, the internal users of the heterogeneous network adjust their association matching relationship with the wireless access node / base station according to the current association situation of the base station / wireless access node and the computing power demand of the task they need to process, so as to meet the dynamic computing power adaptation of the heterogeneous users in the whole network in processing tasks;

[0010] Step 4, on the basis of the association of the heterogeneous devices in the network, the internal users of the heterogeneous network adjust the edge computing strategy including channel access, data offloading ratio, transmission power and task scheduling according to the perceived interference machine behavior and other users' data offloading behavior, so as to realize the optimization target of maximizing the demand satisfaction degree of the internal users of the network;

[0011] Step 5, when the multi-users in the heterogeneous network complete the adjustment, the swarm intelligence algorithm completes an iteration, and when all the multi-dimensional resources in the heterogeneous network are reasonably allocated or the set iteration number is reached, the process is ended.

[0012] Further, the multi-dimensional resource management and task scheduling problem based on task driving in the heterogeneous MEC network is modeled, including:

[0013] Step 1.1, the multi-dimensional resource management and task scheduling problem based on task driving in the heterogeneous MEC network is modeled as a multi-agent nonlinear mixed integer programming decision problem, and the local calculation and mobile edge offloading problem of the internal users of the heterogeneous network is analyzed.

[0014] The multi-dimensional resource management and task scheduling task offloading problem of the intelligent reactive confrontation is modeled as a multi-objective optimization function with common optimization target;

[0015] A multi-access MEC model for wide-area heterogeneous network is considered, which includes multiple mobile devices MUs of different types, a central base station BS, multiple wireless access points APs and intelligent reactive jammers IRJ, aiming to build a complex and efficient edge computing environment. The system model provides computing power support for MUs through the edge computing servers deployed in BS and AP nodes; the edge computing servers are installed in the base station and wireless access node respectively, and provide computing resource access service for MUs; F i lo MU ithe computing power size of the AP / BS, the AP / BS assigns to the MU i , wherein the computing power distribution manner of the AP / BS is related to the number of associated users, and the computing power distribution rule is to evenly distribute the computing power of the AP / BS based on the number of associated users; in terms of mobile offloading, the MUs are built-in with intelligent algorithm modules, which can formulate offloading strategies according to environmental spectrum information and self data offloading needs, and offload part or all of the computing tasks to the BS / AP for auxiliary task processing; the set of available channels outside is and satisfies the number of mobile devices is N; for a mobile user i, each task includes three tuples (d i (t), c i (t), τ i (t)), wherein d i (t) is the data size of the computing task; c i (t) represents the number of CPU cycles required by the computing task, and τ i (t) gives the maximum tolerable delay time of the computing task; when the tth computing task is generated or arrives at a certain mobile user end, the mobile user system must determine whether the task needs to be partially or completely offloaded to the base station or only computed locally;

[0016] For the tth computing task, the MUs decide whether to offload to a specific AP or execute locally based on task characteristics (such as real-time performance, data volume), spectrum resources, and computing power resources; i,m (t) represents the association of the AP m and the MU i when processing task t; wherein o i,m (t) = 1 indicates that the MU i is connected to the AP m ; otherwise, o i,m (t) = 0; in the task offloading process of the mobile user MU, the selected access channel, i.e., the access frequency point, as well as the device, the offloading ratio, the transmission power, and the offloaded tasks of the MU i are the key factors affecting the user to perform data offloading; and are respectively defined as α i (α i ∈[0,1])、P i,m (P i,m ∈[P i,min ,P i,max ])、 If the access channel quality is good, the MU further optimizes the offloading strategy by adjusting the offloading ratio α i and the transmission power P i,m , aiming to achieve efficient and safe offloading of task data.

[0017] All heterogeneous users, wireless access nodes and base stations in the heterogeneous multi-access MEC network are divided into transmit-receive pairs, and all heterogeneous users can select different base stations (BSs) / wireless access nodes (APs) for task offloading. There is a smart reactive jammer in the heterogeneous network, which can perform targeted indiscriminate attacks on offloaded data according to the data offloading behavior of the heterogeneous user, but when the reactive jammer cannot detect the user data offloading behavior, the reactive jammer adopts dynamic jamming to pollute the external spectrum environment, thereby hindering the data offloading behavior of the user in the heterogeneous network. The interference effect of the smart reactive jammer is related to its interference power, channel gain and sensing ability; there is mutual interference between users in the heterogeneous multi-access MEC network, and the size of the mutual interference is related to the mutual interference coefficient, the selected transmission channel and the transmission power.

[0018] Step 1.2, model and analyze the multi-dimensional resource management and task scheduling problem of data offloading in the heterogeneous network, discuss the mobile edge computing problem of heterogeneous tasks (delay-sensitive applications and computing-intensive applications) in the heterogeneous network, and model and describe delay-sensitive applications and computing-intensive applications.

[0019] Heterogeneous users need to offload data to APs based on task type, timeliness and priority to meet diversified QoS requirements; all heterogeneous users are set to periodically generate and process tasks with different QoS requirements; the communication services involved in the task are divided into two categories: delay-sensitive application services and computing-intensive application services; delay-sensitive applications give priority to the security of data transmission and pursue high reliability and low latency; application programs ensure data security by exchanging critical messages between MUs / APs; while computing-intensive applications mainly meet the user's general entertainment services, involving a large amount of data transmission and complex calculation; although computing-intensive applications have some requirements for high data rates, they have higher tolerance in reliability and delay. The specific modeling process of the two service applications is as follows:

[0020] Delay-sensitive application model: a six-tuple model ((d i (t),c i (t),θ i (t),τ i (t),ξ i (t),L i (t))) is used to represent the delay-sensitive application service, and the tuple model includes the following elements: data size d i (t), the number of CPU cycles required for calculation c i (t), the maximum tolerable delay θ i (t), the maximum AoI limit τ i (t), the task urgency index ξ i(t), and a task type identifier L i (t); where L i (t) = 1 represents latency-sensitive, and L i (t) = 0 represents compute-intensive; a local computing delay T i l (t) is as follows:

[0021]

[0022] MU i a tolerable delay θ i (t) is θ i (t) ≥ T i l (t); thus, based on local computing, the energy consumption in the tth computing task is expressed as:

[0023]

[0024] where is the effective capacitance switching parameter of MU i uses a partial offloading mode for latency-sensitive applications; MU i offloads tasks to the AP m The SINR utility function at time t is defined as:

[0025]

[0026] where P i represents the transmit power of MU i , W represents the interference power of the smart reactive J k , W i,m is the channel gain of the transmission link between MU i and the AP m , P k is the channel gain of the interference link between IRJJ m and the AP n , P n,m is the transmit power of other MUs, W m is the link gain of other users and the AP m in the same frequency band, where N0 is the noise spectral density of the AP i ;

[0027] According to the channel capacity theory, the offloading transmission rate R m (t) of user MU i,m to the base station AP m is:

[0028]

[0029] where λ m represents the SINR threshold of AP m . Therefore, the total offloading delay of MU i for the tth computing task is expressed as:

[0030]

[0031] The transmission energy consumption of the tth computing task is:

[0032]

[0033] To ensure high transmission reliability while meeting the needs of users in reducing latency and energy saving, the energy consumption normalization index and the latency normalization index of heterogeneous users in processing latency-sensitive application tasks are introduced. The normalization index calculation method is shown as follows:

[0034]

[0035] where

[0036] Computing-intensive application model: considering the high requirement of computing-intensive applications on data rate, the utility function of such applications can be modeled by offloading throughput. For heterogeneous users MU i , the computing-intensive task is modeled by a four-dimensional tuple , which is expressed as: where d i (t) is the data size, c i (t) is the number of CPU cycles required for computation, represents the minimum transmission rate allowed for offloading computing-intensive tasks; therefore, the transmission rate R i (t) allowed for transmitting computing-intensive tasks in MU i,m is recorded as:

[0037]

[0038] For computing-intensive applications, mobile users adopt a binary offloading mode; for the tth computing-intensive task, the normalized utility function is closely related to the transmission throughput of MU i , which is defined as:

[0039]

[0040] ​Further, in step 2, the mobile user in the isomer MEC network designs a reward weighting value according to the task type, task information, information age and other related heterogeneous task requirements of the user's own task offloading, determines the priority of related task processing, and adopts a suitable edge computing decision for data processing; the reward weighting value is as follows:

[0041] For delay-sensitive applications, the freshness of data is extremely important. The information freshness of data is quantified by AoI, which represents the time delay from the generation of the data packet to the processing of the task. The longer the delay time, the lower the freshness of the task. Delay-sensitive applications use a critical task priority transmission processing mode, that is, before processing each delay-sensitive application task, the user will evaluate the state of the task queue and the freshness of the information to decide whether to prioritize transmission or discard the data in the queue. The data of the compute-intensive application is processed according to the order of arrival, using a first-come-first-served processing mode.

[0042] The new data packet enters the user's buffer according to the average value λ of the Poisson distribution; the data packet in the user's internal cache is represented by Y i <ξ,τ> , where ξ represents the urgency index and τ represents the maximum AoI limit. The urgency level of each data packet is ξ∈{1,2,…,v}, where ξ=0 represents an empty task queue; when the MU i processes a task that exceeds the maximum AoI threshold, the corresponding information will be removed from the cache; when the user processes a task that exceeds the maximum AoI threshold, the task is discarded and the related information is removed from the urgency vector; based on this, the weighting value of the reward function is determined for the characteristics of delay-sensitive applications to encourage priority transmission behavior, where the reward weighting value ι i is defined as:

[0043]

[0044] where t age represents the waiting time of the task before transmission.

[0045] Further, in step 3, the internal users of the heterogeneous network adjust their association matching relationship with the wireless access node / base station according to the current association of the base station / wireless access node and the computing power required for processing the task, to meet the dynamic computing power adaptation of the heterogeneous users in the whole network in processing tasks.

[0046] The present application provides a fast matching algorithm combining matching game and potential game, aiming to realize the fast matching of multiple pairs of Aps-Mus, and considering the companion effect and satisfaction factor to optimize the allocation of computing power resources and improve the network performance. A related association matching satisfaction function is formulated:

[0047]

[0048] where f(MU i ,AP m ) denotes the association satisfaction between MU i and AP m , ranging from 0-1; ζ is a constant ζ≥7; when the computing power δ m assigned to MU i by AP i,m exceeds the task computing power requirement δ' i,m , the matching satisfaction function satisfies: The matching satisfaction function adopts a sigmoid curve to represent the relationship between different task computing power requirements and matching satisfaction; ι i,m represents the urgency of offloading tasks, reflecting the trend of the curve; a lower ι i,m corresponds to a gentler utility curve slope, indicating that the task has a lower demand for computing power; on the contrary, the higher the value of ι i,m , the steeper the curve, indicating that the task has a greater demand for computing resources; the global optimization problem of association satisfaction is defined as:

[0049]

[0050] where Λ is the global network association satisfaction; formula (13) aims to maximize the global network satisfaction by optimizing the association access strategy of all MUs; both MU and AP have autonomous decision-making capabilities: MU needs to select the appropriate AP to calculate whether the computing power resources allocated by its device association meet the demand; AP judges whether the access request of the current MU is accepted according to the global network satisfaction;

[0051] MU constructs its preferred AP association access priority ranking based on its task processing capacity and computing power requirements by collecting external channel state information CSI; the proposed fast matching game algorithm is divided into two stages:

[0052] In stage I, users need to reconstruct the priority list of AP based on the perceived CSI information and location information: mobile users who do not meet the task offloading requirements will propose matching requirements to their preferred wireless access nodes; the number of links should not exceed the number of allowed MU associations; mobile users who have met the task offloading requirements will reduce the number of choices for attempting to access other APs;

[0053] In stage II, after receiving the matching request of the mobile user, the AP will interact with the related APs and determine whether to accept the user's matching request according to certain criteria (such as formula (24)); this decision-making process takes into account multiple factors, including the overall performance of the network, the satisfaction of users, and the effective use of computing resources, etc. If the change of the matching result can improve the overall satisfaction of the network, the user's access request will be accepted.

[0054] Step 4: Based on the current network matching relationship, perceived jammer behavior and other users' task offloading behavior, the internal users in the heterogeneous network adjust their own task scheduling strategies and edge computing strategies to maximize the network internal user demand satisfaction as the optimization goal.

[0055] Further, step 4: Based on the perceived jammer behavior and other users' task offloading behavior, the internal users in the heterogeneous network adjust their own association relationship, task scheduling strategy and edge computing strategy to maximize the network internal user demand satisfaction as the optimization goal.

[0056] The user demand satisfaction function based on heterogeneous tasks is:

[0057]

[0058] where μ and x i,m represent the weight value 0≤μ≤1 and the security transmission index x i,m ={-1,0,1}; MU sets the weight value μ according to the device performance and the current task preference for delay and energy consumption; in the data offloading process, ensuring data security and task computation success is a basic requirement for auxiliary computing tasks. Among these requirements, data security is particularly important. The security transmission index x i,m is introduced to measure the security of task offloading; x i,m =1 indicates that the security of the offloading process can be guaranteed, and the task computation is successfully executed (f i,m ≠f J , max(T i tot (t), T i l (t))≤θ i (t) or ); x i,m =0 indicates that the security of the offloading process can be guaranteed, but the user task processing fails (f i,m ≠f J , max(T i tot (t), T i l (t))>θ i (t) or );x i,m =-1 indicates that the unloading task was interfered with or attacked.

[0059] This application addresses the requirements of users in the HetMEC network regarding task offloading types, priorities, and computational costs, and proposes a distributed multidimensional network resource management scheme. The optimization objective of this scheme is to maximize the satisfaction of heterogeneous users across the entire network while ensuring that the QoS requirements of heterogeneous tasks are met. Therefore, the optimization problem is formulated as follows:

[0060] In the optimization problem, the goal is to find an optimal data offloading decision {P} for each user's computational task. i ,f i ,α i ,Y i}, to make the demand satisfaction level RS i Maximizing (t) means ensuring that the task aims to minimize the computational cost per user in terms of latency and energy consumption, while also ensuring the security of task offloading in the presence of eavesdropping jammers; the network optimization problem is expressed as:

[0061]

[0062] Where f, α represents heterogeneous users connected to the base station / wireless access node. Selected frequency, power, and offload ratio resource allocation matrix; Y i (t) represents MU i The transmission task selected at time t; the constraints represent the scope of the multidimensional heterogeneous decision and task selected by MU.

[0063] The fusion of hierarchical optimization strategy and multi-agent game learning mechanism facilitates cross-layer accurate management and dynamic adjustment of network resources. The upper-layer algorithm module is deployed in base stations and wireless access nodes, uses Q-mix architecture and matching game, and is responsible for channel access of users in the HetMEC network; the lower-layer algorithm module is deployed on heterogeneous users and uses a distributed heterogeneous multi-agent reinforcement learning mechanism to decide data offloading strategies including data offloading ratio, transmit power, and task scheduling. The upper-layer algorithm module is deployed in base stations and wireless access nodes, supervises global network resource allocation and matching, while the lower-layer algorithm focuses on local optimization to meet the QoS requirements of heterogeneous tasks. Specifically, the upper-layer algorithm module uses Q-mix architecture and matching game, and is mainly responsible for the division of spectrum resources and the reasonable allocation of computing power resources in the HetMEC network; the upper-layer reinforcement learning algorithm uses a reinforcement learning mechanism to adaptively adjust the allocation of spectrum resources, effectively avoiding dynamic interference. At the same time, the fast non-replacement matching game algorithm of many-to-many quickly matches the base station / AP with the heterogeneous user, ensuring efficient allocation of computing power. The lower-layer algorithm module is deployed on heterogeneous users and uses a distributed heterogeneous multi-agent reinforcement learning mechanism to solve the data offloading behavior among heterogeneous users. By analyzing real-time spectrum information, wireless access node computing power allocation, and heterogeneous task requirements, the optimal data offloading strategy is formulated.

[0064] Step 5, when the multi-users in the heterogeneous user complete the above adjustment, the swarm intelligence algorithm completes an iteration, and when the multi-dimensional communication resources of all heterogeneous network users are reasonably allocated or the set number of iterations is reached, the process is ended, including:

[0065] Step 5.1, the mobile user observes the current state through channel sensing, and randomly selects the upper-layer channel access decision based on the probability ε and the lower-layer data offloading action

[0066] Step 5.2, after the user performs data offloading , the base station feeds back the reward to the user according to the data offloading situation and inputs the generated data into the neural network for fitting, stores the historical reward experience in the experience review pool, and obtains a new state O i,m (t+1);

[0067] Step 5.3, when k∈L high , the Q function of the upper-layer network is fitted when k∈L low , the Q function of the lower-layer network is fitted , and the expected return is calculated according to the reward, and the Q-mix network of the upper-layer and the lower-layer and the respective sub-network parameters are updated with the minimum loss;

[0068] Step 5.4 loop iteration until the maximum number of iterations is reached, and the optimal edge computing decision of the user in the heterogeneous network is obtained to meet the computing and offloading requirements of the respective user tasks.

[0069] Further, the specific process of fitting the neural network in step 5.3 is as follows:

[0070] The upper network is deployed on the AP, and the Q-Mix network architecture is used to manage and allocate channel resources for all APs in the HetMEC network; each AP is responsible for selecting the optimal channel access strategy for its associated MUs to avoid severe channel fading and dynamic interference attacks in the network, and to prevent frequency conflicts between MUs covered by the same AP or adjacent APs; the high-order Q-Mix hybrid network is deployed on the maximum AP, i.e. the BS, which collects the Q-network parameters of all APs to maximize the channel resource revenue; in the high-level network, each MU sends local observation values to its AP, and the AP selects a discrete channel access strategy for its MU according to the observation values DDQN uses the advantages of double neural networks in DQN to alleviate the overestimation problem in traditional methods; the Q-network is trained by minimizing the iterative loss function, which is expressed as:

[0071]

[0072] The lower Q-mix algorithm hybrid network is deployed on the AP to maximize the user satisfaction of all MUs associated with the AP;

[0073] For the lower network part, the MU i selects continuous parameters according to its new observation results:

[0074]

[0075] where is the channel access and device association strategy obtained from the upper network; the task information, location information, current channel information and are combined into a new state and input into the lower network; considering the dynamic channel access decision and local observation information, the lower network advantage function is used, and the hybrid decision of each agent is input to generate the lower network value of the MU i , and the update strategy is:

[0076]

[0077]

[0078] where represents the MU iThe lower layer advantage function; y low is a step target value obtained by the lower layer hybrid network.

[0079] Compared with the prior art, the present application has the following advantages: (1) the interaction restriction mechanism among task scheduling, unloading strategy, communication resource and computing power resource dynamic allocation is fully considered, a high-efficiency joint multi-dimensional network resource optimization and task scheduling problem modeling is provided, the optimization target of maximizing the network internal user demand satisfaction is realized, and the quantitative solution of the problem is realized. The task scheduling problem faced by the heterogeneous user in processing the heterogeneous task is accurately described and modeled; the more accurate and clear physical meaning heterogeneous network resource management and task scheduling model can better describe the user data unloading scene in the heterogeneous MEC network; (2) an anti-interference data unloading algorithm based on the task-driven hierarchical heterogeneous network resource management system combined with the dynamic spectrum access technology is provided, the independent user independently perceives the external channel environment, selects the safe data unloading decision and transmission channel, and introduces the related system overhead factor, so as to improve the stability of the multi-user channel access decision. BRIEF DESCRIPTION OF DRAWINGS

[0080] Figure 1 is the system model diagram of the heterogeneous network data unloading model provided by the present application for processing heterogeneous tasks;

[0081] Figure 2 is the algorithm flowchart of the task-driven hierarchical heterogeneous network resource management data unloading algorithm provided by the present application;

[0082] Figure 3 is the heterogeneous data packet queuing processing schematic diagram of the task-driven hierarchical heterogeneous network resource management data unloading algorithm provided by the present application;

[0083] Figure 4 is the performance comparison diagram of the task-driven hierarchical heterogeneous network resource management data unloading algorithm provided by the present application for resisting tracking interference and dynamic random period interference. DETAILED DESCRIPTION

[0084] In order to make the purpose, technical scheme and advantages of the present application more clear, the embodiments of the present application will be further described in detail below with reference to the drawings.

[0085] Firstly, the embodiments of the present application will be introduced below with reference to the drawings.

[0086] The heterogeneous network multi-dimensional resource management data offloading model for processing heterogeneous tasks aims to more accurately depict the multi-dimensional resource management and data offloading process in the heterogeneous multi-access MEC network of the user processing heterogeneous tasks. In the model, all heterogeneous, wireless access nodes and base stations in the heterogeneous multi-access MEC network are divided into transmitting-receiving pairs, and all heterogeneous can select different base stations or wireless access nodes for task offloading. In the heterogeneous network, there is an intelligent reactive jammer that can track and interfere with offloaded data according to the data offloading behavior of the heterogeneous user, but when the intelligent reactive jammer cannot perceive the user offloading behavior, the intelligent reactive jammer uses dynamic interference to hinder the data offloading of the heterogeneous user in the heterogeneous network. The interference effect of the intelligent reactive jammer is related to its perception ability, interference power and channel gain; there is mutual interference between the heterogeneous users in the heterogeneous network, and the size of the mutual interference is related to the mutual interference coefficient, the selected channel and the offloading power. The purpose of the present application is to provide a task-driven hierarchical heterogeneous network resource management data offloading method. Considering the specific needs of each heterogeneous user in terms of priority, security and transmission cost in processing heterogeneous tasks, by analyzing the mutual restriction and interaction mechanism between dynamic resource allocation, task scheduling and edge offloading strategy, an efficient multi-dimensional resource joint optimization and task scheduling scheme is designed to maximize the user demand satisfaction degree in the network as the optimization goal, and the heterogeneous task is solved. The method of the present application can meet the user safety and delay constraint, and jointly optimize the channel access, dynamic computing power scheduling, transmission power and task scheduling of multiple users, realize the reasonable allocation of multi-dimensional resources and efficient data offloading.

[0087] Figure 1 The system model diagram of the heterogeneous network data offloading model for processing heterogeneous tasks. In the system model, the heterogeneous mobile device end is equipped with an intelligent module, and the mobile device makes corresponding mobile offloading decisions by sensing external information and its own data offloading requirements, and offloads data to the base station for auxiliary calculation. The reactive jammer has the ability to eavesdrop on mobile terminal data information and interfere with data offloading. When the mobile terminal offloads data, the reactive jammer perceives the data offloading behavior of the user, and when the user offloading behavior is perceived by the reactive jammer, the reactive jammer transmits tracking interference to destroy the data offloading behavior. When the reactive jammer cannot perceive the task offloading behavior, the eavesdropper uses common dynamic interference (sweep frequency, double sweep frequency, comb-shaped dynamic combination) to hinder the task offloading of the heterogeneous user.

[0088] Figure 2is the algorithm flowchart of the task-driven hierarchical heterogeneous network resource management data offloading algorithm proposed in the application. The proposed algorithm design closely focuses on user demand and network condition. The core of the design is to integrate hierarchical optimization strategy and multi-agent game learning mechanism, which is convenient for cross-layer accurate management and dynamic adjustment of network resources. The upper algorithm module is deployed in the base station and wireless access node, supervises the global network resource allocation and matching, and the lower algorithm focuses on local optimization to meet the QoS requirements of heterogeneous tasks. Specifically, the upper algorithm module uses Q-mix architecture and matching game, mainly responsible for the division of heterogeneous MEC network spectrum resources and the reasonable allocation of computing power resources; the upper reinforcement learning algorithm uses the reinforcement learning mechanism to adaptively adjust the spectrum resource allocation, effectively avoiding dynamic interference. At the same time, the multi-to-multi fast non-replacement matching game algorithm quickly and accurately associates AP with Mus, ensuring efficient computing power allocation. The lower algorithm module is deployed on MUs, and uses a distributed heterogeneous multi-agent reinforcement learning mechanism to solve the data offloading behavior between heterogeneous MUs. Through analyzing real-time spectrum information, AP computing power allocation and heterogeneous task demand, the module formulates the optimal data offloading strategy.

[0089] Figure 3 is the data packet queuing system schematic diagram of the task-driven hierarchical heterogeneous network resource management data offloading algorithm proposed in the application. The task scheduling method proposed in the application is mainly based on the freshness and urgency of task information. For delay-sensitive applications, data freshness is crucial, which is quantified by information age (AoI). AoI refers to the time delay experienced by data packets from generation to task processing. The longer the delay, the lower the task freshness. Before processing each task, the user will evaluate the task queue state and information freshness to determine whether to prioritize transmission or discard the data in the queue. For compute-intensive applications, data is processed in the order of task arrival. Similarly, before processing each task, the user will also evaluate the task queue state and information freshness to determine whether to prioritize transmission or discard the data in the queue.

[0090] At each discrete time step, the user decides whether to prioritize transmission or discard the data in the task queue by observing the execution state of the task queue, information age, and external environment information. The environmental features involved here are described in the system model and channel model.

[0091] From an information theory perspective, the user needs to use network resources to provide deterministic services for more high-priority tasks while ensuring safe offloading. The application considers the priority of task transmission and uses the urgency of the task itself as a linear weight of its information age to achieve more reasonable and efficient task scheduling.

[0092] Figure 4is a performance comparison chart of the task-driven hierarchical network resource management data offloading algorithm for resisting intelligent reactive interference in the embodiment of the application.

[0093] The embodiments of the application are specifically described as follows, system simulation adopts python language, and is based on TensorFlow deep learning framework. The simulation scene mainly consists of 3 APs (one base station and two microcell nodes), 6 MUs and 2 IRJs. The service coverage radius of the BS is set to 250 m. The microcell nodes are symmetrically placed on both sides of the base station service area, and the service radius of each microcell is 80 m, ensuring that there is no service overlap between microcells. The number of available channels outside is set to 10, and the offloading bandwidth of the MUs is 4 MHz. The APs, MUs and IRJs perform task offloading and confrontation in the 40 MHz frequency band, allowing MUs to dynamically adjust channels, power and offloading ratios in each time slot. The mobile users move at a constant speed of 1 m / s away from or close to the base station. The proposed algorithm is trained under different learning rates and heterogeneous task Qos configurations, and these factors will significantly affect wireless communication when set differently. The transmission power ranges of the heterogeneous users are 7 dBm-20 dBm and 10 dBm-23 dBm respectively. The communication signal and interference signal both adopt cosine waveform. A user task generation model is introduced to generate computing tasks of 200Kb-600Kb (subject to uniform distribution), and the task arrival process conforms to Poisson distribution. The computing capacity (CPU cycle number) of the user is 1.5×10 8 cycles / s. The inherent computing power of the mobile user is 1.5×10 8 cycles / s, and the energy efficiency factor is 10 -27 J / cycle. The computing power that the base station can allocate to each MU is 1×10 9 cycles / s. The total computing power that the microcell node can provide to the accessed user is 3×10 9 cycles / s, and the users belonging to the microcell node provide service in the microcell node computing power.

[0094] The intelligent reactive jamming is composed of a perception module, a tracking jamming module and a conventional jamming module. In the conventional jamming module, a strategy combining dynamic jamming and conventional jamming modes is adopted to pollute the external spectrum environment. Specifically, the IRJ intelligent reactive jamming randomly selects comb jamming and sweep jamming. The time length of maintaining the current jamming type is randomly selected, and the time length selection range is 10s-20s. The comb jamming is divided into three fixed frequency bands: 4MHz-8MHz, 12MHz-16MHz, 20MHz-24MHz, 28MHz-32MHz. In order to verify the performance superiority and effectiveness of the algorithm, the network utility of the algorithm is compared with that of several other MEC anti-jamming algorithms in this chapter: In order to evaluate the effectiveness of the algorithm, the task processing success rate of the algorithm is compared with that of three other anti-jamming data offloading algorithms:

[0095] 1. MA-PDQN algorithm: The MA-PQN algorithm combines P-DQN algorithm and MADDPG to solve the problem of distributed mixed heterogeneous strategy.

[0096] 2. MADDPG algorithm: MADDPG is a MARL algorithm based on CTDE paradigm. The continuous action space is discretized into a discrete action space to realize channel access and task scheduling strategy.

[0097] 3. Q-mix algorithm: The traditional Q-mix algorithm is applied to HetMEC scene for data offloading. The algorithm does not use hierarchical architecture, and the offloading ratio and offloading power and other continuous decision dimensions are discretized.

[0098] Figure 4 is the performance comparison chart of the algorithm based on task scheduling and resource allocation against the intelligent reactive jammer in embodiment 1 of the application. Through Figure 4 It can be seen that the effectiveness and practicality of the proposed algorithm. Energy-efficient communication resource scheduling and data offloading are realized.

[0099] It can be seen that the effectiveness and practicality of the proposed algorithm. Energy-efficient communication resource scheduling and data offloading are realized. Figure 4 It can be seen that the effectiveness and practicality of the proposed algorithm. Energy-efficient communication resource scheduling and data offloading are realized.

[0100] To sum up, the application organically fuses advanced intelligent algorithms such as matching game, multi-agent reinforcement learning (MARL), federated learning, and the like, and proposes a task-driven hierarchical heterogeneous network resource management data offloading algorithm, aiming to effectively cope with and comprehensively consider the quality of service (QoS) requirements of time-varying heterogeneous tasks by jointly optimizing device association, network resource allocation, and task scheduling. The proposed task-driven hierarchical heterogeneous network resource management data offloading algorithm can effectively solve the NP-hard optimization problem in multi-dimensional resource management. Specifically, the algorithm architecture is divided into two layers: the upper layer focuses on the optimization of device association and spectrum access strategy, and can quickly respond to the dynamic changes of network topology structure and quickly realize the association matching between wireless access nodes / bases and mobile users.

[0101] The lower layer focuses on the fine adjustment of heterogeneous data offloading and task scheduling, and through the cooperation and competition mechanism between agents, the algorithm can explore the optimal strategy under the constraint of limited resources.

[0102] On the one hand, the algorithm can further improve the resource utilization, so that various resources in the network can be more fully and reasonably utilized, and resource waste can be avoided; on the other hand, the fairness of resource allocation can be guaranteed, and different users and devices can obtain a relatively fair share of resources in the network, so that the overall network performance is not affected by the excessive occupation or insufficient of resources by some users or devices. This has important significance for improving the running efficiency and stability of the heterogeneous multi-access MEC network in the intelligent reactive interference scene.

[0103] The above-mentioned embodiments of the application do not constitute a limitation on the protection scope of the application.

Claims

1. A task-driven hierarchical heterogeneous network resource management data offloading method, characterized in that, The method includes: Step 1: Model the task-driven multidimensional resource management and task scheduling problem in heterogeneous MEC networks. The model is described as a partially observable Markov game process, in which the participants are heterogeneous users in the heterogeneous network. Distributed data offloading decision optimization is performed on joint task scheduling, dynamic spectrum access, and dynamic computing resource management. Step 2: Heterogeneous users in the heterogeneous network design reward weighting values ​​based on the queuing information of their own cached tasks and task-related information, including latency requirements and information age, to determine the priority of task processing and use appropriate edge computing decisions for data processing. After detecting heterogeneous user data offloading behavior, the intelligent reactive jammer uses tracking jamming or dynamic jamming to hinder the data transmission offloading behavior; Step 3: The internal users of the heterogeneous network adjust their association matching relationship with the wireless access node / base station according to the current association status of the base station / wireless access node and the computing power requirements of their own processing tasks, so as to meet the dynamic computing power adaptation of heterogeneous users across the network in processing tasks. Mobile users (MUs) collect external Channel State Information (CSI) and construct their preferred AP association access priority ranking based on their own task processing capabilities and computing power requirements. Users need to reconstruct their priority list for APs based on externally perceived CSI and location information: mobile users whose task offloading requirements are not met will submit matching requests to their preferred radio access nodes, with the number of links not exceeding the number of MU associations they are allowed to equip; mobile users whose task offloading requirements have been met will reduce the number of other APs they can try to access. After receiving a matching request from a mobile user, the wireless access node (AP) will exchange information with the relevant AP and determine whether to accept the user's matching request based on specific criteria. If the change in the matching result can improve the overall satisfaction of the network, then the user's access request will be accepted. Step 4: Based on the interconnection of heterogeneous network devices, internal users of the heterogeneous network adjust their respective edge computing strategies, including channel access, data offload ratio, transmission power, and task scheduling, according to the perceived behavior of external interference devices and the data offload behavior of other users, in order to maximize the satisfaction of internal users' needs as the optimization goal. Step 5: After multiple users in the heterogeneous network have completed their adjustments, the swarm intelligence algorithm completes one iteration. The process ends when all multidimensional resources in the heterogeneous network are reasonably allocated, or when the set number of iterations is reached.

2. The method according to claim 1, characterized in that, Step 1 involves modeling the task-driven multidimensional resource management and task scheduling problem in heterogeneous MEC networks, including: Step 1.1: The task-driven multidimensional resource management and task scheduling problem in heterogeneous MEC networks is modeled as a multi-agent nonlinear mixed integer programming decision problem, and the local computing and mobile edge offloading problems of users within the heterogeneous network are analyzed. Step 1.2 involves modeling and analyzing the multidimensional resource management and task scheduling problems of heterogeneous network data offloading, analyzing the mobile edge computing problems of latency-sensitive and compute-intensive applications in heterogeneous networks, and modeling and describing latency-sensitive and compute-intensive applications.

3. The method according to claim 2, characterized in that, Step 1.1 models the task-driven multidimensional resource management and task scheduling problem in heterogeneous MEC networks as a multi-agent nonlinear mixed-integer programming decision problem, and analyzes the local computing and mobile edge offloading problems of users within the heterogeneous network, including: The multidimensional resource management, task scheduling, and task offloading problems of adversarial intelligent reactive systems are modeled as multi-objective optimization functions with common optimization objectives. A multi-access MEC model for wide-area heterogeneous networks is considered, which includes multiple mobile devices (MUs) of different types, a central base station (BS), multiple wireless access points (APs), and intelligent reactive jammers (IRJs). The edge computing servers deployed on the BS and AP nodes provide computing power support to the MUs. The edge computing servers are installed on the base station and wireless access nodes respectively to provide computing resource access services to the MUs. MU i The computing power This indicates that AP / BS is assigned to MU. i The computing power of the AP / BS is allocated based on the number of associated users. The allocation rule is to distribute the computing power of each AP / BS evenly based on the number of associated users. Regarding mobile offloading, the MUs has a built-in intelligent algorithm module that can formulate offloading strategies based on environmental spectrum information and its own data offloading needs, partially or completely offloading computing tasks to the BS / AP for auxiliary task processing. The set of available external channels is set as follows: and satisfy There are N mobile devices; for mobile user i, each task consists of three tuples (d... i (t),c i (t),τ i (t)), where d i (t) is the data size of the computation task; c i (t) represents the number of CPU cycles required for the computation task, τ i (t) gives the maximum tolerable delay time for the computation task; when the t-th computation task is generated or arrives at a mobile user terminal, the mobile user system must determine whether the task needs to be partially or completely offloaded to the base station, or only computed locally. For the t-th computation task, MUs determines whether to offload it to a specific AP or execute it locally based on task characteristics, spectrum resources, and computing power resources; i,m (t) represents AP m and MU i The correlation when processing task t; where o i,m (t) = 1, indicating that MU i With AP m Connected; otherwise, o i,m (t) = 0; During the task offloading process of the mobile user MU, the selected access channel (i.e., access frequency point), equipment, offloading ratio, transmission power, and MU are... i The uninstallation task is a key factor affecting users' ability to uninstall data; defined as follows: α i (α i ∈[0,1]), P i,m (P i,m ∈[P i,min ,P i,max ]) If the access channel quality is good, the MU further optimizes the offloading strategy by adjusting the offloading ratio α. i With transmission power P i,m .

4. The method according to claim 3, characterized in that, Step 1.2 involves modeling and analyzing the multidimensional resource management and task scheduling problems of heterogeneous network data offloading, discussing the mobile edge computing problems of latency-sensitive and compute-intensive applications within heterogeneous networks, and modeling and describing latency-sensitive and compute-intensive applications, including: Heterogeneous users need to offload data to APs based on task type, timeliness, and priority; all heterogeneous users are configured to periodically generate and process tasks with different QoS requirements; the communication services involved in the tasks are divided into two categories: latency-sensitive application services and compute-intensive application services. The specific modeling process is as follows: Delay-sensitive application model: adopting a six-dimensional tuple model ((d i (t),c i (t),θ i (t),τ i (t),ξ i (t),L i (t))) is used to characterize latency-sensitive application services. The tuple model consists of the following elements: data scale d i (t) Calculate the number of CPU cycles required, c i (t), maximum tolerable delay θ i (t), Maximum AoI Limit τ i (t), Task urgency index ξ i (t), and task type identifier L i (t); where L i (t) = 1 represents a time delay sensitive type, L i (t) = 0 represents computationally intensive; the local computation latency T for heterogeneous users processing the t-th computation task. i l (t) is shown below: MU i Tolerable delay θ i The constraint of (t) is Therefore, based on local computation, the energy consumption of the t-th computation task is... The expression is: in For MU i Effective capacitor switching parameters; heterogeneous users employ partial offload mode for time-sensitive applications; MU i Unload task to AP m The SINR utility function is defined as follows: Where P i MU i The transmission power, Represents the intelligent reactive expression J k Interference power, W i,m For MU i With AP m Channel gain of the transmission link between them For IRJJ k With AP m Channel gain of the interfering link, P n For the transmit power of other MUs, W n,m For other users and APs in the same frequency band m The link gain, where N0 is the AP m The noise spectral density; According to channel capacity theory, user MU i To base station AP m Offload transfer rate R i,m (t) is: Where λ m Indicates AP m The SINR threshold; therefore, for the t-th computational task, MU i The total unloading delay is expressed as: Transmission energy consumption of the t-th computation task for: Introducing energy consumption normalization metrics for heterogeneous users when handling latency-sensitive application tasks Normalized index of time delay The normalized index is calculated as follows: in Computation-intensive application model: For heterogeneous user MU i For example, computationally intensive tasks can be modeled using a single four-dimensional tuple. To model and represent: where d i (t) represents the data size, c i (t) represents the number of CPU cycles required for the calculation. This indicates the minimum transfer rate allowed when computationally intensive tasks are offloaded; therefore, MU i The transmission rate R allowed for computationally intensive tasks is... i,m (t) is denoted as: For compute-intensive applications, mobile users adopt a binary offload mode; for the t-th compute-intensive task, the normalized utility function and MU i It is closely related to the transmission throughput, and is defined as:

5. The method according to claim 4, characterized in that, The weighted reward values ​​in step 2 are shown below: Data freshness is quantified using AoI, which represents the time delay from data packet generation to task processing. The longer the delay, the lower the freshness of the task. Latency-sensitive applications adopt a mission-critical priority transmission processing mode. That is, before processing each latency-sensitive application task, the user will evaluate the status of its task queue and the freshness of the information to decide whether to prioritize the transmission or discard the data in the queue. Computation-intensive application data is processed according to the order of arrival, adopting a first-come, first-served processing mode. New data packets enter the user's buffer according to the average value λ of the Poisson distribution; data packets in the user's internal buffer are... Let ξ represent the urgency index and τ represent the maximum AoI limit; the urgency level of each packet is ξ∈{1,2,…,v}, where ξ=0 indicates that the task queue is empty; when MU i When the number of processed tasks exceeds the maximum AoI threshold, the corresponding information will be removed from the cache; when the number of user-processed tasks exceeds the maximum AoI threshold, the tasks will be discarded, and the relevant information will be removed from the emergency state vector; based on this, a weighted value for the relevant reward function is determined for the characteristics of latency-sensitive applications to encourage priority transmission behavior, where the reward weight value ι i Defined as: Where t age This indicates the waiting time before the task is transmitted.

6. The method according to claim 5, characterized in that, The association matching relationship in step 3 is determined by the following method: Define the correlation matching satisfaction function: Where f(MU) i AP m ) represents MU i With AP m The correlation between satisfaction levels ranges from 0 to 1; ζ is a constant ζ≥7; when AP m Assigned to MU i computing power δ i,m Exceeding the task's computing power requirement δ′ i,m When the satisfaction function is satisfied, the following conditions are met: The matching satisfaction function uses an S-shaped curve to represent the relationship between the computing power requirements of different tasks and the matching satisfaction; i,m This indicates the urgency of the unloading task and reflects the trend of the curve; lower ι... i,m A relatively flat utility curve slope indicates a lower computational demand on the task; conversely, ι i,m The higher the value, the steeper the curve, indicating a greater demand for computing resources from the task; the global optimization problem related to satisfaction is defined as: Among them, the overall network satisfaction is Λ; Formula (13) aims to maximize the overall network satisfaction by optimizing the association access strategy of all MUs; MU and AP have autonomous decision-making capabilities: MU needs to select a suitable AP and calculate whether the computing resources allocated to its device association meet the requirements; AP then judges whether to accept the current MU's access request based on the overall network satisfaction.

7. The method according to claim 6, characterized in that, In step 4, the user satisfaction function based on heterogeneous tasks is: Where μ and x i,m These represent the weight value 0≤μ≤1 and the secure transmission index x, respectively. i,m ={-1,0,1}; MU sets weight values ​​μ based on device performance and the current task's preference for latency and energy consumption; introduces a secure transmission index x. i,m To measure the security of task unloading; x i,m =1 indicates that the security of the uninstallation process is guaranteed, and the task calculation was successfully executed. i,m ≠f J , or x i,m =0 indicates that the security of the uninstallation process can be guaranteed, but the user task processing failed. i,m ≠f J , or x i,m =-1 indicates that the unloading task was interfered with by an attack; In the optimization problem, the goal is to find an optimal data offloading decision {P} for each user's computational task. i ,f i ,α i ,Y i }, to make the demand satisfaction level RS i Maximizing (t) means ensuring that the task aims to minimize the computational cost per user in terms of latency and energy consumption, while also ensuring the security of task offloading in the presence of eavesdropping jammers; the network optimization problem is expressed as: Where f, α represents heterogeneous users connected to the base station / wireless access node. Selected frequency, power, and offload ratio resource allocation matrix; Y i (t) represents MU i The transmission task selected at time t; constraints represent the scope of the multidimensional heterogeneous decision and task selected by MU.

8. The method according to claim 7, characterized in that, Step 5: After multiple users in the heterogeneous network have completed their adjustments, the swarm intelligence algorithm completes one iteration. The process ends when all multidimensional resources in the heterogeneous network are reasonably allocated, or when the set number of iterations is reached. This includes: Step 5.1: The mobile user observes the current state through channel awareness and randomly selects an upper-layer channel for access based on probability ε. and lower-level data unloading action Step 5.2, when the user performs data unloading Afterwards, the base station sends the feedback reward back to the user based on the data offloading situation and inputs the generated data into a neural network for fitting, incorporating historical reward experience. Store in the experience revisit pool and obtain new state O i,m (t+1); Step 5.3, when k∈L high Time-fitting the Q-function of the upper network When k∈L low Time-fitting the Q-function of the underlying network The expected return is calculated based on the reward, and the parameters of the upper and lower Q-mix networks and their respective sub-networks are updated with minimal loss. Step 5.4 Iterate repeatedly until the maximum number of iterations is reached to find the optimal edge computing decision for each user in the heterogeneous network to meet the computing and offloading requirements of their respective user tasks.

9. The method according to claim 8, characterized in that, Step 5.3 The specific process of the neural network performing the fitting is as follows: The upper-layer network is deployed on the APs, using the Q-Mix network architecture to manage and allocate channel resources for all APs in the HetMEC network; each AP is responsible for selecting the optimal channel access strategy for its associated MUs. The high-order Q-Mix hybrid network is deployed on the largest AP, i.e., the BS, and collects the Q-network parameters of all APs with the goal of maximizing channel resource benefits. In the higher layers of the network, each MU sends local observations to its AP, and the AP selects a discrete channel access strategy for its MU based on the observations. The Q-network is trained by minimizing an iterative loss function, expressed as: The hybrid network of the lower-level Q-mix algorithm is deployed on the AP, thereby maximizing the user satisfaction of all MUs associated with the AP; For the lower-layer network part, MU i Based on the new observations, continuous parameters were selected: in This is for channel access and device association strategies obtained from higher-level networks; combining task information, location information, current channel information, and... Combine into a new state The data is then input into the lower-level network; taking into account dynamic channel access decisions and local observation information, the lower-level network advantage function is used. The hybrid decision of each agent is used as input to generate the MU. i The lower-level network values ​​are updated using the following strategy: in MU i The lower-level dominance function; y low It is the target value obtained from the lower-level hybrid network.

Images