Internet of vehicles federal learning excitation method and system based on Nash game and coalition game

Abstract

The invention discloses an Internet of Vehicles federated learning excitation method and system based on a Nash game and a coalition game, and the method comprises the steps: initializing a vehicle client set in an Internet of Vehicles environment, and enabling a server to broadcast global model parameters to all clients; a two-stage Nash game model is established, the first stage is used for determining an alliance set in the vehicle client set, and the second stage is used for bargaining to optimize resource contribution; negotiating and determining an optimal data contribution ratio and an optimal resource contribution ratio based on the server side and the alliance set by using the two-stage Nash game model; and the server allocates rewards to the alliance set according to the optimal data contribution proportion and the optimal resource contribution proportion. The method can adapt to high dynamics of the Internet of Vehicles, and fair and efficient resource allocation is realized.

Description

Technical Field

[0001] The present invention belongs to the field of vehicle networking and distributed machine learning technology, and in particular relates to a vehicle networking federated learning incentive method and system based on Nash game and alliance game. Background Art

[0002] With the rapid development of intelligent transportation systems, advanced sensors in modern vehicles generate massive amounts of data. This data contains valuable information that can be used to improve traffic management, enhance autonomous driving capabilities, and analyze road safety. Federated Learning (FL) offers an attractive solution for the connected vehicle environment by supporting distributed model training while maintaining data locality and privacy.

[0003] However, deploying federated learning in the IoV environment faces numerous challenges. First, the high mobility of vehicles leads to frequent changes in network topology and unpredictable participation patterns, severely disrupting the training process. While existing research recognizes the potential impact of vehicle mobility on FL performance, it fails to fully consider its impact and is unable to maintain effective incentives when participants frequently join and leave the network. Second, data heterogeneity weakens the effectiveness of incentive mechanisms. Data in IoV exhibits non-independent and identically distributed (non-IID) characteristics, resulting in suboptimal convergence speed. Existing methods, such as introducing proximal regularization and training GAN networks, can alleviate the non-IID problem, but they suffer from instability and overhead constraints in IoV environments, and fail to establish fair compensation mechanisms that consider the quality and quantity of data contributions. Finally, the strict constraints on vehicle performance and energy, as well as their significant heterogeneity, make efficient incentive optimization imperative. Some studies have explored client selection strategies based on static resource allocation, but these methods fail to account for the cumulative impact of high mobility and time-varying data quality. Traditional federated learning incentive mechanisms, such as FedAvg and FedProx, primarily focus on static environments and are difficult to adapt to the dynamic nature of IoV. Stackelberg game-based approaches require multiple iterations to reach equilibrium, making them inefficient in high-mobility scenarios. While auction mechanisms account for resource heterogeneity, they can easily lead to participant loss in frequently changing network conditions. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a federated learning incentive method and system for the Internet of Vehicles based on Nash game and coalition game to adapt to the high dynamics of the Internet of Vehicles and achieve fair and efficient resource allocation.

[0005] To achieve the above objectives, the present invention provides a method for incentivizing federated learning in Internet of Vehicles (IoV) based on Nash game and coalition game, comprising:

[0006] Initialize the vehicle client set in the Internet of Vehicles environment, and the server broadcasts the global model parameters to all clients;

[0007] A two-stage Nash game model is established, wherein the first stage is used to determine the alliance set in the vehicle client set, and the second stage is used to bargain and optimize resource contribution;

[0008] Based on the two-stage Nash game model, the server and the alliance set negotiate to determine the optimal data contribution ratio and the optimal resource contribution ratio;

[0009] The server distributes rewards to the alliance set according to the optimal data contribution ratio and the optimal resource contribution ratio.

[0010] Optionally, before establishing the two-stage Nash game model, the following steps are further included:

[0011] Construct a stable alliance structure based on a multi-dimensional scoring function of data quality, communication efficiency, and mobility cost;

[0012] Within each alliance in the stable alliance structure, a cluster-head and a backup cluster-head are selected based on a three-dimensional score of server communication efficiency, location centrality, and link stability.

[0013] Optionally, the multidimensional scoring function is:

[0014]

[0015] Among them, P m (M(t)) is the multidimensional scoring function, Q data (M(t)) is the data quality score, Q comm (M(t)) is the communication efficiency score, is the mobility cost score, μ1 and μ2 are normalization factors.

[0016] Optionally, building the stable alliance structure includes: a merging module, a splitting module, and a switching module;

[0017] The merging module is:

[0018] P(M i (t)∪M j (t))≥P(M i (t))+P(M j (t))+τ merege and|M i (t)|+|M j (t)|≤

[0019] l limit ,

[0020] Among them, M i (t) is alliance i, M j (t) is alliance j, P(M i (t)) is the multi-dimensional scoring function value of alliance i, P(M j (t)) is the multi-dimensional scoring function value of alliance j, τ merge is the threshold of the merge operation, l limit Set a size limit for the number of alliances;

[0021] The splitting modules are:

[0022] P(M i (t))<τ split or |M i (t)|>l limit ,

[0023] And the alliance subset after the split {M i1 ,M i2 ,…M ik The following conditions must be met:

[0024]

[0025] Among them, τ split is the splitting threshold, τ gain_split is the splitting profit threshold, M ij (t) is the jth alliance in the alliance subset that alliance i splits, P(M ij (t)) is the multidimensional scoring function value of the j-th alliance in the alliance subset, j is the subscript of the j-th alliance in the alliance subset after the split, and k is the number of sets in the alliance subset after the split;

[0026] The switching module is:

[0027] P m (M j (t)∪{v m})≥P m (M i (t)\{v m})+τ switch ,

[0028] Among them, τ switch is the switching benefit threshold, v m For the vehicle set v = {v1,v2,…,v N} in the mth client.

[0029] Optionally, the cluster-head and backup cluster-head are selected based on a three-dimensional score of server communication efficiency, location centrality, and link stability.

[0030] Optionally, the two-stage Nash game model is:

[0031]

[0032] Among them, V(y(t),z(t)) is the server utility function, U i (y j (t),z j (t)) is the alliance utility function, y(t) is the optimal alliance structure Π solved by Nash game * The solution vector is composed of the optimal data contribution ratio of each alliance in (t), and z(t) is the optimal alliance structure Π solved by the Nash game * The solution vector composed of the optimal calculation contribution ratio of each alliance in (t), y j (t) is the optimal alliance structure Π obtained by solving the Nash game * The optimal data contribution ratio of the jth alliance in (t), z j (t) is the optimal alliance structure Π obtained by solving the Nash game * The optimal computing resource contribution ratio of the jth alliance in (t), Π * (t) is the optimal alliance structure.

[0033] Optionally, determining the optimal data contribution ratio and the optimal resource contribution ratio through negotiation includes: using a multi-starting point sequential quadratic programming algorithm to determine the optimal data contribution ratio and the optimal resource contribution ratio through negotiation;

[0034] The method of using a multi-starting point sequence quadratic programming algorithm to negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio includes: using Latin hypercube sampling to generate a preset number of initial points, executing the quadratic programming algorithm in parallel for each of the initial points to obtain a local optimal solution, and selecting the solution with the largest objective function value among the local optimal solutions as the Nash bargaining solution;

[0035] Based on the Nash bargaining solution, the optimal data contribution ratio and the optimal resource contribution ratio are obtained.

[0036] Optionally, the server allocating rewards to the alliance set according to the optimal data contribution ratio and the optimal resource contribution ratio includes:

[0037] According to the optimal data contribution ratio and resource contribution ratio, a stochastic gradient descent algorithm is used to update the local model parameters;

[0038] The updated local model parameters are sent to the cluster head for local aggregation, and the central aggregation server is used to perform weighted averaging of the aggregated local model parameters to update the global model parameters;

[0039] Rewards are distributed to the alliance set according to the updated local model parameters and global model parameters.

[0040] The present invention also provides a vehicle network federated learning incentive system based on Nash game and alliance game, including: a data initialization module, an alliance formation module, a cluster management module, a bargaining optimization module and a global aggregation module;

[0041] The data initialization module is used to initialize the vehicle client set in the Internet of Vehicles environment, and the server broadcasts global model parameters to all clients;

[0042] The alliance formation module is used to build a stable alliance structure based on a multi-dimensional scoring function of data quality, communication efficiency and mobility cost;

[0043] The cluster management module is configured to select a cluster head and a backup cluster head within each alliance in the stable alliance structure based on a three-dimensional score of server communication efficiency, location centrality, and link stability;

[0044] The bargaining optimization module is used to establish a two-stage Nash game model, wherein the first stage is used to determine the alliance set in the vehicle client set, and the second stage is used to negotiate and optimize resource contribution; based on the two-stage Nash game model, the server and the alliance set negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio;

[0045] The global aggregation module is used for the server to distribute rewards to the alliance set according to the optimal data contribution ratio and the optimal resource contribution ratio.

[0046] The present invention also provides a storage medium on which a computer program is stored. When the computer program is running, it executes a federated learning incentive method for Internet of Vehicles based on Nash game and alliance game.

[0047] Compared with the prior art, the present invention has the following advantages and technical effects:

[0048] This invention incorporates alliance game theory to self-organize vehicles into a stable learning alliance, effectively mitigating the instability caused by individual vehicles frequently entering and exiting the network. It employs a Nash bargaining framework to achieve fair distribution of resource contributions. Compared to traditional fixed reward mechanisms, this approach can dynamically adjust incentive strategies based on real-time network conditions and data quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The accompanying drawings, which constitute part of this application, are intended to provide a further understanding of this application. The exemplary embodiments and descriptions of this application are intended to explain this application and do not constitute an improper limitation on this application. In the accompanying drawings:

[0050] Figure 1 This is a flow chart of a method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments in this application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0052] It should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and that, although a logical order is shown in the flowcharts, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0053] Example 1:

[0054] This embodiment proposes a method for encouraging federated learning in Internet of Vehicles based on Nash game and alliance game. Figure 1 As shown, the specific steps include:

[0055] Initialize the vehicle client set in the Internet of Vehicles environment, and the server broadcasts the global model parameters to all clients;

[0056] A two-stage Nash game model is established, where the first stage is used to determine the alliance set in the vehicle client set, and the second stage is used to bargain and optimize resource contribution;

[0057] Based on the two-stage Nash game model, the server and the alliance set negotiate to determine the optimal data contribution ratio and the optimal resource contribution ratio;

[0058] The server distributes rewards to the alliance set based on the optimal data contribution ratio and the optimal resource contribution ratio.

[0059] Furthermore, before establishing the two-stage Nash game model, the following steps are also included:

[0060] Construct a stable alliance structure based on a multi-dimensional scoring function of data quality, communication efficiency, and mobility cost;

[0061] Within each alliance in the stable alliance structure, cluster-head and backup cluster-head are selected based on the three-dimensional scores of server communication efficiency, location centrality and link stability.

[0062] Specifically, alliance formation based on mobility awareness: In the Internet of Vehicles environment, consider an edge network consisting of a server and a set of N heterogeneous vehicle nodes V = {v1, v2, ..., v NThe IoV is modeled as a federated learning system, where the central server acts as a global model aggregator and the vehicles act as distributed learning clients. The learning process consists of T global iterations, expressed as To achieve effective and resource-constrained incentive mechanism deployment, this embodiment proposes an adaptive alliance formation mechanism based on game theory. First, a comprehensive evaluation framework is established. For the client v after joining the alliance M(t), m , define the multidimensional scoring function:

[0063]

[0064] Among them, P m (M(t)) is the multidimensional scoring function, Q data (M(t)) is the data quality score, Q comm (M(t)) is the communication efficiency score, is the mobility cost score, μ1 and μ2 are normalization factors.

[0065] More specifically, 0<μ i <1, each sub-item is normalized as follows:

[0066] Data quality rating: Use Generalized Earth Mover Distance (GEMD) to measure client v m Improvement of data heterogeneity after joining the alliance.

[0067] Communication efficiency score: Assess communication capabilities relative to the league average.

[0068] Mobility Cost Score: Quantify the difference between client speed and league average speed.

[0069] Among them, α is the adjustment parameter, r m For client v m The transmission rate, is the average client transmission rate in alliance i, s m For client v m speed, is the average client speed in alliance i, s max is the maximum client speed in alliance i.

[0070] Based on a scoring mechanism, the alliance formation process is modeled as a hedonic alliance game, in which clients make autonomous decisions to maximize their individual utility. Client preference relationships and Pareto order are defined, and dynamic adaptation of the alliance is achieved through three basic operations: merging, splitting, and switching.

[0071] Furthermore, building a stable alliance structure includes: merging module, splitting module and switching module;

[0072] The merged modules are:

[0073] P(M i (t)∪M j (t))≥P(M i (t))+p(M j (t))+τ merge and|M i (t)|+|M j (t)|≤

[0074] l limit ,

[0075] Among them, M i (t) is alliance i, M j (t) is alliance j, P(M i (t)) is the multi-dimensional scoring function value of alliance i, P(M j (t)) is the multi-dimensional scoring function value of alliance j, τ merge is the threshold of the merge operation, l limit Set a size limit for the number of alliances;

[0076] The split modules are:

[0077] P(M i (t))<τ split or |M i (t)|>l limit ,

[0078] And the alliance subset after the split {M i1 ,M i2 ,…M ik The following conditions must be met:

[0079]

[0080] Among them, τ split is the splitting threshold, τ gain_split is the splitting profit threshold, M ij (t) is the jth alliance in the alliance subset that alliance i splits, P(M ij (t)) is the multidimensional scoring function value of the j-th alliance in the alliance subset, j is the subscript of the j-th alliance in the alliance subset after the split, and k is the number of sets in the alliance subset after the split;

[0081] The switching module is:

[0082] P m (M j (t)∪{v m})≥P m (M i(t)\{v m})+τ switch ,

[0083] Among them, τ switch is the switching benefit threshold, v m Assemble for vehicles The mth client in .

[0084] Furthermore, the cluster-head and backup cluster-head are selected based on the three-dimensional scores of server communication efficiency, location centrality, and link stability.

[0085] Specifically, alliance cluster head selection: After establishing a federated learning alliance for the Internet of Vehicles, cluster head selection is crucial to optimizing communication efficiency and model performance. This embodiment proposes a cluster head selection algorithm based on multi-dimensional scoring, which evaluates candidate nodes through a three-dimensional indicator system:

[0086]

[0087]

[0088] Scoring function g(v m ;M i (t)) quantifies the adaptability of each node as a cluster head:

[0089]

[0090] in, Represents the weight of the kth indicator. The three indicators are:

[0091] Server communication efficiency indicators:

[0092] Location centrality index:

[0093] Link stability indicators:

[0094] in, is the optimal cluster head in alliance i in round t, is the weight of the kth indicator, g k is the kth index, For client v m With Alliance M i The average distance to other clients in (t), ω1 is the normalization factor, and ω2 is the normalization factor.

[0095]

[0096] Where, Δs=|s m -s j |For client vm and client v j The absolute speed difference, ε is the speed difference threshold. j,m It is a variable with a value of 0 or 1, which is used to indicate the relative movement direction between two clients. A value of 1 indicates the same direction. trans is the maximum distance range that clients can communicate with each other. d(m,j) is the distance between clients v m and client v j The distance between them.

[0097] Dynamic alliance selection: After establishing a stable and efficient alliance network, a dynamic alliance selection strategy is introduced to effectively adapt to real-time network stability and evaluate the value of each alliance based on multiple factors. First, define the alliance selection ratio: r(t) = ζ·r base ·R stab Among them, r base As the basis, select the ratio. is the network stability index. Based on the selection ratio r(t), the number of alliances selected in round t is calculated as: Then, define the comprehensive score of the alliance as:

[0098] Score(M i (t))=ζ1·Q data (M i (t))+ζ2·Q his (M i (t))

[0099] Among them, ζ is the impact factor, λ is the attenuation coefficient, σ is the speed standard deviation of the vehicle client set v, is the average speed of the vehicle client set, I(t) is the number of alliances obtained by the alliance formation algorithm, ζ1 is the impact factor, and ζ2 is the impact factor.

[0100] A candidate list L(t) is formed based on descending order of scores, and the top τ(t) coalitions are selected to participate in the FL task.

[0101] Furthermore, the two-stage Nash game model is:

[0102]

[0103] Among them, V(y(t),z(t)) is the server utility function, U i (y j (t),z j (t)) is the alliance utility function, y(t) is the optimal alliance structure Π solved by Nash game * The solution vector is composed of the optimal data contribution ratio of each alliance in (t), and z(t) is the optimal alliance structure Π solved by the Nash game* The solution vector composed of the optimal calculation contribution ratio of each alliance in (t), y j (t) is the optimal alliance structure Π obtained by solving the Nash game * The optimal data contribution ratio of the jth alliance in (t), z j (t) is the optimal alliance structure Π obtained by solving the Nash game * The optimal computing resource contribution ratio of the jth alliance in (t), Π * (t) is the optimal alliance structure.

[0104] Specifically, Nash bargaining game: in forming the optimal alliance structure Π * After (t), the cluster head of each selected coalition engages in concurrent bargaining negotiations with the server to determine the optimal contribution ratio and available reward for each coalition. A multi-party Nash bargaining solution (NBS) is constructed to optimize resource allocation in round t. The payoffs of both bargaining parties at the divergence point are zero, so the divergence point is defined as the origin s0 = 0. According to the Nash theorem, the optimal protocol requires maximizing the weighted Nash product.

[0105] Furthermore, negotiating and determining the optimal data contribution ratio and the optimal resource contribution ratio includes: using a multi-starting point sequential quadratic programming algorithm to negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio;

[0106] The multi-starting point sequential quadratic programming algorithm is used to negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio. The method includes: using Latin hypercube sampling to generate preset initial points, executing the quadratic programming algorithm in parallel for each initial point to obtain a local optimal solution, and selecting the solution with the largest objective function value among the local optimal solutions as the Nash bargaining solution;

[0107] Based on the Nash bargaining solution, the optimal data contribution ratio and the optimal resource contribution ratio are obtained.

[0108] Furthermore, the server distributes rewards to the alliance set based on the optimal data contribution ratio and the optimal resource contribution ratio, including:

[0109] Based on the optimal data contribution ratio and resource contribution ratio, the stochastic gradient descent algorithm is used to update the local model parameters;

[0110] The updated local model parameters are sent to the cluster head for local aggregation, and the central aggregation server is used to perform weighted averaging of the aggregated local model parameters to update the global model parameters;

[0111] Distribute rewards to the coalition ensemble based on the updated local and global model parameters.

[0112] Specifically, local training and aggregation: each client v m ∈M i(t) Perform H steps of local iterations on the private dataset using the Stochastic Gradient Descent (SGD) algorithm according to the training strategy of its alliance:

[0113]

[0114] where h = {1, 2…, H-1}, η is the learning rate, is the mini-batch data for the hth iteration. After local training, the updated local model parameters are sent to the cluster head for local aggregation. The aggregated model parameters of each alliance are efficiently transmitted to the central aggregation server through quantization and compression techniques, and a weighted average is performed to update the global model.

[0115] Incentive Distribution: To ensure long-term and efficient participation, alliances that complete training will receive rewards based on Nash bargaining results. The reward scheme satisfies the Incentive Compatibility (IC) and Individual Rationality (IR) constraints:

[0116]

[0117] Among them, γ1 and γ2 are the reward values ​​of unit effective computing contribution and unit effective data contribution respectively, C i (t) and D i (t) are the effective computing contribution and effective data contribution of the alliance respectively.

[0118] The following is a detailed description of Example 1:

[0119] like Figure 1 As shown, the specific implementation steps of Example 1 are as follows:

[0120] 1) System Initialization and Parameter Setup: A 15km two-way highway scenario was simulated, with three lanes in each direction and a server located 7.5km away. Vehicle speeds in lanes 0-2 were 40-60, 60-80, and 80-100km / h, and speeds in lanes 3-5 (reverse direction) were -40-(-60), -60-(-80), and -80-(-100)km / h. 100 heterogeneous clients were set up, divided into three categories:

[0121] High-end devices (30%): 1.5-3.0 GFLOPS (computing power), 15-25 MHz (bandwidth), 80-120 mW (transmission power);

[0122] Mid-range devices (50%): 0.8-1.5 GFLOPS, 8-15 MHz, 50-80 mW;

[0123] Low-end devices (20%): 0.3-0.8 GFLOPS, 3-8 MHz, 20-50 mW;

[0124] 2) Alliance formation process: At initialization, each vehicle node forms a separate alliance V = {{v1}, {v2}, …, {v N}}. Calculate the initial preference P m (M(t)), set invalid operation count no_ops_count = 0. Iterate and perform the following operations until convergence (no_ops_count < 3):

[0125] Merge operation: traverse all alliance pairs (M i (t), M(t)), merge if the following conditions are met:

[0126] Collective benefits exceed the threshold: P(M i (t)∪M j (t))≥P(M i (t))+P(M j (t))+τ merge ;

[0127] Scale constraints are satisfied: |M i (t)|+|M j (t)|≤l limit ;

[0128] The merged partitions satisfy the Pareto order:

[0129] Split Operation: Alliance M i (t), if P(M i (t))<τ split or |M i (t)|>l limit , generate all possible partitioning schemes and select the one that satisfies:

[0130]

[0131] Conditional program execution.

[0132] Switching operation: For v m ∈M i (t) Switch to M j (t), checking individual payoff improvements and target coalition capacity constraints.

[0133] 3) Cluster head selection and alliance selection: For each formed alliance M i (t)∈Π ini (t), calculate the three-dimensional scores of all nodes:

[0134] Calculate server communication efficiency g1, calculate location centrality g2, and calculate link stability g3;

[0135] Select the node with the highest score as the cluster head The second highest cluster head serves as the backup cluster head According to the network stability R stab Calculate the selection ratio r(t) and determine the number of participating alliances τ(t). Calculate the comprehensive score Score(M i (t)), select the first τ(t) coalitions to form Π * (t).

[0136] 4) Nash bargaining solution: Constructing an optimization problem:

[0137]

[0138] Generated using Latin Hypercube Sampling (LHS) Initial points. Execute the SLSQP algorithm in parallel for each initial point to obtain the local optimal solution. The solution with the largest objective function value is selected as the final Nash bargaining solution (y * (t),z * (t)).

[0139] 5) Federated learning training execution: The server sends the global model ω t The cluster heads then broadcast the update to all members of the alliance. Each client performs H = 5 rounds of local training with a batch size of B = 32 and a learning rate of η = 0.01. The cluster heads collect the model updates within the alliance and perform a weighted average:

[0140]

[0141] Server aggregation alliance model:

[0142]

[0143] 6) Incentive allocation and history update: According to the bargaining results (y * (t),z * (t)) Calculate the rewards for each alliance and update the alliance's historical performance set The profit records of the last five rounds are retained for subsequent evaluation.

[0144] The algorithm is shown in Table 1:

[0145] Table 1

[0146]

[0147]

[0148] Example 2:

[0149] This embodiment also provides a federated learning incentive system for Internet of Vehicles based on Nash game and alliance game, including: a data initialization module, an alliance formation module, a cluster management module, a bargaining optimization module and a global aggregation module;

[0150] The data initialization module is used to initialize the vehicle client set in the Internet of Vehicles environment, and the server broadcasts the global model parameters to all clients;

[0151] Alliance formation module, used to build a stable alliance structure based on a multi-dimensional scoring function of data quality, communication efficiency and mobility cost;

[0152] A cluster management module is used to select cluster heads and backup cluster heads within each alliance in a stable alliance structure based on the three-dimensional scores of server communication efficiency, location centrality, and link stability;

[0153] The bargaining optimization module is used to establish a two-stage Nash game model. The first stage is used to determine the alliance set in the vehicle client set, and the second stage is used to negotiate and optimize resource contribution. Based on the two-stage Nash game model, the server and the alliance set negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio.

[0154] The global aggregation module is used by the server to distribute rewards to the alliance set based on the optimal data contribution ratio and the optimal resource contribution ratio.

[0155] Example 3:

[0156] This embodiment also provides an electronic device, including: a memory and a processor, wherein the memory stores a computer program executed by the processor, and when the computer program is executed by the processor, it executes the above-mentioned vehicle network federated learning incentive method based on Nash game and alliance game.

[0157] Example 4:

[0158] This embodiment also provides a storage medium, on which a computer program is stored. When the computer program is run, it executes the above-mentioned vehicle network federated learning incentive method based on Nash game and alliance game.

[0159] The above are merely preferred embodiments of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for incentivizing federated learning in Internet of Vehicles based on Nash game and coalition game, characterized in that: include: Initialize the vehicle client set in the Internet of Vehicles environment, and the server broadcasts the global model parameters to all clients; A two-stage Nash game model is established, wherein the first stage is used to determine the alliance set in the vehicle client set, and the second stage is used to bargain and optimize resource contribution; Based on the two-stage Nash game model, the server and the alliance set negotiate to determine the optimal data contribution ratio and the optimal resource contribution ratio; The server distributes rewards to the alliance set according to the optimal data contribution ratio and the optimal resource contribution ratio.

2. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 1, characterized in that: Before establishing the two-stage Nash game model, the following steps are also included: Construct a stable alliance structure based on a multi-dimensional scoring function of data quality, communication efficiency, and mobility cost; Within each alliance in the stable alliance structure, a cluster-head and a backup cluster-head are selected based on a three-dimensional score of server communication efficiency, location centrality, and link stability.

3. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 2, characterized in that: The multidimensional scoring function is: Among them, P m (M(t)) is the multidimensional scoring function, Q data (M(t)) is the data quality score, Q comm (M(t)) is the communication efficiency score, is the mobility cost score, μ1 and μ2 are normalization factors.

4. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 2, characterized in that: Constructing the stable alliance structure includes: merging module, splitting module and switching module; The merging module is: P(M i (t)∪M j (t))≥P(M i (t))+P(M j (t))+τ merge and Among them, M i (t) is alliance i, M j (t) is alliance j, P(M i (t)) is the multi-dimensional scoring function value of alliance i, P(M j (t)) is the multi-dimensional scoring function value of alliance j, τ merge is the threshold for the merge operation, Set a size limit for the number of alliances; The splitting modules are: P(M i (t))<τ split or And the alliance subset after the split {M i1 ,M i2 ,…,M ik The following conditions must be met: Among them, τ split is the splitting threshold, τ gain_split is the splitting profit threshold, M ij (t) is the jth alliance in the alliance subset that alliance i splits, P(M ij (t)) is the multidimensional scoring function value of the j-th alliance in the alliance subset, j is the subscript of the j-th alliance in the alliance subset after the split, and k is the number of sets in the alliance subset after the split; The switching module is: P m (M j (t)∪{v m })≥P m (M i (t)\{v m })+τ switch , Among them, τ switch is the switching benefit threshold, v m For the vehicle set V = {v1,v2,…,v N } in the mth client.

5. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 1, characterized in that: The cluster-head and backup cluster-head are selected based on the three-dimensional scores of server communication efficiency, location centrality and link stability.

6. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 1, characterized in that: The two-stage Nash game model is: Among them, V(y(t),z(t)) is the server utility function, U i (y j (t),z j (t)) is the alliance utility function, y(t) is the optimal alliance structure Π solved by Nash game * The solution vector is composed of the optimal data contribution ratio of each alliance in (t), and z(t) is the optimal alliance structure Π solved by the Nash game * The solution vector composed of the optimal calculation contribution ratio of each alliance in (t), y j (t) is the optimal alliance structure Π obtained by solving the Nash game * The optimal data contribution ratio of the jth alliance in (t), z j (t) is the optimal alliance structure Π obtained by solving the Nash game * The optimal computing resource contribution ratio of the jth alliance in (t), Π * (t) is the optimal alliance structure.

7. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 1, characterized in that: Negotiating and determining the optimal data contribution ratio and the optimal resource contribution ratio includes: adopting a multi-starting point sequence quadratic programming algorithm to negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio; The method of using a multi-starting point sequence quadratic programming algorithm to negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio includes: using Latin hypercube sampling to generate a preset number of initial points, executing the quadratic programming algorithm in parallel for each of the initial points to obtain a local optimal solution, and selecting the solution with the largest objective function value among the local optimal solutions as the Nash bargaining solution; Based on the Nash bargaining solution, the optimal data contribution ratio and the optimal resource contribution ratio are obtained.

8. The method for incentivizing federated learning in Internet of Vehicles based on Nash game and alliance game according to claim 1, characterized in that: The server distributing rewards to the alliance set according to the optimal data contribution ratio and the optimal resource contribution ratio includes: According to the optimal data contribution ratio and resource contribution ratio, a stochastic gradient descent algorithm is used to update the local model parameters; The updated local model parameters are sent to the cluster head for local aggregation, and the central aggregation server is used to perform weighted averaging of the aggregated local model parameters to update the global model parameters; Rewards are distributed to the alliance set according to the updated local model parameters and global model parameters.

9. A federated learning incentive system for Internet of Vehicles based on Nash game and coalition game, characterized by: include: Data initialization module, alliance formation module, cluster management module, bargaining optimization module and global aggregation module; The data initialization module is used to initialize the vehicle client set in the Internet of Vehicles environment, and the server broadcasts global model parameters to all clients; The alliance formation module is used to build a stable alliance structure based on a multi-dimensional scoring function of data quality, communication efficiency and mobility cost; The cluster management module is configured to select a cluster head and a backup cluster head within each alliance in the stable alliance structure based on a three-dimensional score of server communication efficiency, location centrality, and link stability; The bargaining optimization module is used to establish a two-stage Nash game model, wherein the first stage is used to determine the alliance set in the vehicle client set, and the second stage is used to negotiate and optimize resource contribution; based on the two-stage Nash game model, the server and the alliance set negotiate and determine the optimal data contribution ratio and the optimal resource contribution ratio; The global aggregation module is used for the server to distribute rewards to the alliance set according to the optimal data contribution ratio and the optimal resource contribution ratio.

10. A storage medium, characterized in that: The storage medium stores a computer program, which, when running, executes the method according to any one of claims 1 to 8.

Images