Architecture generation and hierarchical alignment precise aggregation method for heterogeneous federated learning
By using Markov decision processes and reinforcement learning to explore Pareto optimal architectures in heterogeneous federated learning, and combining beam search and gradient tracking mechanisms, the problems of architecture suboptimality and parameter mismatch among heterogeneous devices are solved, achieving efficient model training and aggregation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV
- Filing Date
- 2026-01-20
- Publication Date
- 2026-06-09
AI Technical Summary
Existing heterogeneous federated learning methods suffer from suboptimal and parameter space mismatch issues when generating architectures, resulting in performance limitations and aggregation interference, and failing to fully utilize the resources of heterogeneous devices.
We employ Markov decision processes and reinforcement learning agents to explore Pareto optimal architectures in the global supernet search space. By combining beam search algorithm and gradient tracking mechanism, we dynamically adjust the sub-network architecture and perform precise aggregation.
It enables efficient model training and aggregation across heterogeneous devices, improves system performance, eliminates interference caused by parameter space mismatch, and ensures the accuracy and stability of the model.
Smart Images
Figure CN122174924A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of distributed machine learning (federated learning), specifically to a method for architecture generation and hierarchical alignment-based precise aggregation for heterogeneous federated learning. Background Technology
[0002] Federated Learning (FL) has become a highly attractive distributed machine learning paradigm, allowing for the training of powerful models using distributed data while protecting user privacy. However, in practical deployments, significant system heterogeneity exists among client devices in terms of computing power, storage space, and network bandwidth, severely hindering the widespread adoption of federated learning. Traditional federated learning algorithms (such as FedAvg) typically require all clients to train models with the same structure. This one-size-fits-all design limits the overall system performance to the most resource-constrained devices, failing to fully utilize the resources of more powerful nodes in the network. To address this issue, Heterogeneous Federated Learning (HFL) has emerged. Its core idea is to dynamically allocate sub-network models of different sizes to clients based on their resource budgets, striving to achieve efficient model training and aggregation in heterogeneous environments.
[0003] In the field of Heuristic Networking Filtering (HFL), mainstream methods adopt the supernet-subnet paradigm, which has mainly gone through three stages of development: early static single-dimensional scaling (such as HeteroFL) only reduced the network width by a uniform ratio, often resulting in a thin and suboptimal structure; subsequent multi-dimensional heuristic scaling (such as ScaleFL) introduced a combined search of width, depth, and early retreat mechanisms to try to expand the adaptation space; recent research has turned to dynamic adaptive pruning (such as FlexFL, FedCDP), using data-driven metrics (such as APoZ, L1 norm) to dynamically adjust the network structure. Meanwhile, some works have attempted to introduce Network Assortment (NAS) into FL, aiming to automatically search for network architectures. However, existing federated NAS methods either focus on searching for a single globally optimal model while ignoring device differences, or pursue completely personalized architectures while losing the basis for collaboration between models, and often require searching on the client side, bringing huge computational overhead and privacy risks.
[0004] While the aforementioned methods alleviate resource constraints to some extent, they still suffer from two unavoidable deep-seated flaws in their theoretical assumptions and architectural design. First, there is the inherent suboptimal nature of heuristic architecture generation: existing HFL methods (whether scaling or pruning) primarily rely on pre-defined empirical rules (such as fixed grid search ratios) or simple heuristic strategies. These methods implicitly assume that a balanced architecture is optimal; however, this assumption often fails in complex heterogeneous environments. Experiments show that these subnets generated through heuristic rules systematically deviate from the Pareto front in the performance-efficiency space. This means that, with equivalent computational overhead, existing methods cannot uncover hidden high-performance non-uniform architectures, limiting the performance ceiling of federated systems. Second, there is the destructive interference caused by parameter space mismatch: this is the most fatal challenge faced by existing methods. The traditional FedAvg algorithm requires aggregation to be performed in a unified parameter space, while existing nested scaling or pruning strategies break this premise. In heterogeneous subnetworks generated by these methods, shared neurons (parameters) often exist in drastically different architectural contexts (e.g., the same convolutional kernel might be located in a 32-channel layer in one subnetwork and a 128-channel layer in another). This difference in context leads to the same parameters being trained on different clients to assume drastically different or even conflicting functional roles. When the server forces a weighted average of these functionally heterogeneous parameters, it generates severe and destructive parameter interference, directly causing the performance of the global model to collapse. Existing solutions typically rely on indirect alignment methods such as knowledge distillation (KD) for compensation, but this is only a stopgap approximation strategy and cannot fundamentally eliminate the accuracy loss and training instability caused by parameter space mismatch. Therefore, designing an adaptive federated learning method that can break the constraints of heuristic rules to intelligently explore Pareto optimal architectures and fundamentally eliminate parameter space mismatch between heterogeneous subnetworks to achieve lossless and accurate aggregation has become an urgent technical challenge. Summary of the Invention
[0005] The purpose of this invention is to provide a method for architecture generation and hierarchical alignment-accurate aggregation for heterogeneous federated learning, comprising the following steps:
[0006] Step 1) Initialize the global supernet and in the public proxy dataset Pre-training is performed on the network to obtain the trained global supernet weights;
[0007] Step 2) Establish a Markov decision process;
[0008] Step 3) The reinforcement learning-based agent explores the global supernet search space according to the Markov decision process until the number of Pareto optimal architectures exceeds a preset value, forming a candidate architecture pool. ;
[0009] Step 4) The server obtains the set of clients currently participating in training. And divide the client into based on the client hardware resource budget. An orderly resource hierarchy; ;
[0010] Step 5) Build a system for resource grading Select the optimal subnetwork architecture The optimization model;
[0011] Step 6) Solve the optimization model using the beam search algorithm, and select from the candidate architecture pool. The middle level is for each resource level. Select the optimal subnetwork architecture ;
[0012] Step 7) The selected optimal sub-network architecture and corresponding model parameters are sent to the client with the appropriate resource level. Then the client uses local private data to train the model, calculates and uploads gradient update parameters.
[0013] Step 8) The server aggregates the gradient update parameters and updates the global supernet weights;
[0014] Step 9) Determine whether the current communication round t has reached the preset maximum communication round. If yes, end the training and output the trained model. If no, set t = t + 1 and return to step 4.
[0015] Step 10) Use the trained model to perform the tasks uploaded by the customer.
[0016] Furthermore, the task is an image recognition task.
[0017] Furthermore, in step 1), the initialized global supernet contains the maximum search space. The architecture of the global supernet is determined by the maximum depth and maximum width allowed in the search space, which constitutes the upper limit of the weights of all possible subnets. This means that the global supernet must contain all possible levels and channel number configurations within the search space.
[0018] Furthermore, the Markov decision-making process includes states. ,action and rewards ;
[0019] Among them, state Low-dimensional random exploration signal; action = ,in It is a width coefficient vector obtained from Gaussian distribution sampling, used to define the number of channels in each layer; It is a depth index obtained through Softmax sampling, used to determine the early regression position of the network;
[0020] award As shown below:
[0021] (1)
[0022] In the formula, Basic reward; reward ; This is the currently generated sub-network architecture. It is a collection of historical schemas stored in a sliding window. It is a history structure within a sliding window; It is the weighted Euclidean distance in the architecture parameter space. This formula is used to calculate the current architecture. The minimum distance from all architectures in the historical window is used as the reward. This reward encourages the agent to explore novel architectures that differ from previous ones, preventing it from getting trapped in local optima or converging prematurely. , For weights.
[0023] Furthermore, basic rewards As shown below:
[0024] (2)
[0025]
[0026] In the formula, For proxy indicators; The variance of the width vector; , These are the maximum and minimum values of the width vector; , , As weight;
[0027] proxy indicators As shown below:
[0028] (3)
[0029] In the formula, For proxy data batches; To activate the feature map of the convolutional layer.
[0030] Furthermore, when exploring the search space of the global supernet, a regional search strategy, an adaptive exploration intensity strategy, and a periodic partial reset strategy are adopted.
[0031] Furthermore, the regional search strategy refers to dividing the search space of the global supernet into... Each region is a logical sub-region; during the adaptive phase, the historical innovation rate of each region is considered. Probabilistically select the current search area ;
[0032] Among them, select area As the current search area The probabilities are as follows:
[0033] (4)
[0034] In the formula This is the only architecture found in this region. The total number of architectures generated for this region. This is a smoothing factor.
[0035] Furthermore, an adaptive exploration intensity strategy is used to adjust the agent's exploration intensity;
[0036] The agent's exploration intensity in the next exploration As shown below:
[0037] (5)
[0038] In the formula, The uniqueness ratio of the generated architecture for the current batch; , The preset ratio threshold; Current exploration intensity; It is an increasing factor; To reduce factors.
[0039] Furthermore, the periodic partial reset strategy refers to: every preset batch, using a binary mask that follows a Bernoulli distribution. Randomly reset some parameters of the policy network In order to break the inertia of strategy;
[0040] Parameters during the next exploration As shown below:
[0041] (6)
[0042] In the formula, These are the initial parameters randomly generated when the network begins training; For policy networks In the Current real-time parameters for each training batch;
[0043] Furthermore, the optimized model is shown below:
[0044] (7)
[0045] (8)
[0046] (9)
[0047] (10)
[0048] (11)
[0049] in, It is the first The number of clients in the group and These represent computing costs and memory costs, respectively.
[0050] Furthermore, in step 8), the step of the server aggregating the gradient update parameters includes:
[0051] Step 8.1) For each client and architecture The server performs a virtual forward and backward propagation on the supernet, activating the server belonging to the supernet. The path generates a binary mask. This identifies the parameters trained by the client.
[0052] binary mask As shown below:
[0053] (12)
[0054] In the formula, For loss;
[0055] Step 8.1) Apply the generated mask to the global supernet weights After updating, we get:
[0056] (13)
[0057] In the formula, This is the learning rate.
[0058] The technical effects of this invention are undeniable. This invention solves the suboptimal performance problem caused by traditional heuristic architecture generation methods in heterogeneous device federated learning scenarios, as well as the parameter space mismatch and aggregation interference problems caused by incompatibility of heterogeneous subnet structures. Attached Figure Description
[0059] Figure 1 Flowchart for offline architecture generation and online dynamic hierarchical training. Detailed Implementation
[0060] The present invention will be further described below with reference to embodiments, but it should not be construed that the scope of the present invention is limited to the following embodiments. Various substitutions and modifications made based on ordinary technical knowledge and common practices in the art without departing from the above-described technical concept of the present invention should be included within the scope of protection of the present invention.
[0061] Example 1:
[0062] See Figure 1 A precise aggregation method for architecture generation and hierarchical alignment in heterogeneous federated learning includes the following steps:
[0063] Step 1) Initialize the global supernet and in the public proxy dataset Pre-training is performed on the network to obtain the trained global supernet weights;
[0064] Step 2) Establish a Markov decision process;
[0065] Step 3) The reinforcement learning-based agent explores the global supernet search space according to the Markov decision process until the number of Pareto optimal architectures exceeds a preset value, forming a candidate architecture pool. ;
[0066] Step 4) The server obtains the set of clients currently participating in training. And divide the client into based on the client hardware resource budget. An orderly resource hierarchy; ;
[0067] Step 5) Build a system for resource grading Select the optimal subnetwork architecture The optimization model;
[0068] Step 6) Solve the optimization model using the beam search algorithm, and select from the candidate architecture pool. The middle level is for each resource level. Select the optimal subnetwork architecture ;
[0069] Step 7) Send the selected optimal sub-network architecture and corresponding model parameters to the clients of the appropriate resource level;
[0070] The client uses local private data to train the model, calculates and uploads gradient update parameters;
[0071] Step 8) The server aggregates the gradient update parameters and updates the global supernet weights;
[0072] Step 9) Determine whether the current communication round t has reached the preset maximum communication round. If yes, end the training and output the trained model. If no, set t = t + 1 and return to step 4.
[0073] Step 10) Use the trained model to perform the tasks uploaded by the customer.
[0074] Example 2:
[0075] The architecture generation and hierarchical alignment precise aggregation method for heterogeneous federated learning is technically the same as in Example 1. Further, in step 1), the initialized global hypernet contains the maximum search space. The architecture of the global hypernet is determined by the maximum depth and maximum width allowed in the search space, which constitute the upper limit of the weights of all possible subnetworks. This means that the global hypernet must contain all possible hierarchical and channel number configurations within the search space.
[0076] Example 3:
[0077] A precise aggregation method for architecture generation and hierarchy alignment for heterogeneous federated learning, with technical content identical to any one of Examples 1-2, further includes a Markov decision process comprising states. ,action and rewards ;
[0078] Among them, state It is a low-dimensional random exploration signal;
[0079] action = ,in It is a width coefficient vector obtained from Gaussian distribution sampling, used to define the number of channels in each layer; It is a depth index obtained through Softmax sampling, used to determine the early regression position of the network;
[0080] award As shown below:
[0081] (1)
[0082] In the formula, Basic reward; reward ; This is the currently generated sub-network architecture. It is a collection of historical schemas stored in a sliding window. It is a history structure within a sliding window; It is the weighted Euclidean distance in the architecture parameter space. This formula is used to calculate the current architecture. The minimum distance from all architectures in the historical window is used as the reward. This reward encourages the agent to explore novel architectures that differ from previous ones, preventing it from getting trapped in local optima or converging prematurely. , For weights.
[0083] Example 4:
[0084] A precise aggregation method for architecture generation and hierarchy alignment for heterogeneous federated learning, with technical content identical to any one of Examples 1-3, further including basic rewards. As shown below:
[0085] (2)
[0086]
[0087] In the formula, For proxy indicators; The variance of the width vector; , These are the maximum and minimum values of the width vector; , , As weight;
[0088] Among them, the proxy indicator As shown below:
[0089] (3)
[0090] In the formula, For proxy data batches; To activate the feature map of the convolutional layer.
[0091] Example 5:
[0092] The architecture generation and hierarchical alignment precise aggregation method for heterogeneous federated learning has the same technical content as any one of Examples 1-4. Furthermore, when exploring the search space of the global supernet, a region search strategy, an adaptive exploration intensity strategy, and a periodic partial reset strategy are adopted.
[0093] Example 6:
[0094] The architecture generation and hierarchical alignment precise aggregation method for heterogeneous federated learning has the same technical content as any one of Examples 1-5. Furthermore, the region search strategy refers to dividing the search space of the global supernet into... Each region is a logical sub-region; during the adaptive phase, the historical innovation rate of each region is considered. Probabilistically select the current search area ;
[0095] Among them, select area As the current search area The probabilities are as follows:
[0096] (4)
[0097] In the formula This is the only architecture found in this region. The total number of architectures generated for this region. This is a smoothing factor.
[0098] Example 7:
[0099] An architecture generation and hierarchical alignment precise aggregation method for heterogeneous federated learning, with the same technical content as any one of Examples 1-6, further wherein an adaptive exploration intensity strategy is used to adjust the exploration intensity of the agent;
[0100] The agent's exploration intensity in the next exploration As shown below:
[0101] (5)
[0102] In the formula, The uniqueness ratio of the generated architecture for the current batch; , The preset ratio threshold; Current exploration intensity; It is an enhancement factor, a multiplicative coefficient used to rapidly increase exploration intensity. To reduce the factor, it is a subtraction value used to slowly search for exploration intensity; both serve the purpose of better exploring diverse architectures.
[0103] Example 8:
[0104] The architecture generation and hierarchical alignment precise aggregation method for heterogeneous federated learning has the same technical content as any one of Examples 1-7. Furthermore, the periodic partial reset strategy refers to: every preset batch, using a binary mask following a Bernoulli distribution... Randomly reset some parameters of the policy network In order to break the inertia of strategy;
[0105] Parameters during the next exploration As shown below:
[0106] (6)
[0107] In the formula, The initial parameters are randomly generated at the start of network training, representing the original random state without any training bias. Policy network In the The current real-time parameters for each training batch carry the "learned knowledge" accumulated by the agent after previous training and the resulting policy inertia. The system uses a mask... Will A small portion of the parameters were forced to roll back and replaced with This mixing of old and new states aims to break the established policy inertia of the model, thereby helping the agent escape deep local optima.
[0108] Example 9:
[0109] A precise aggregation method for architecture generation and hierarchical alignment for heterogeneous federated learning, with technical content identical to any one of Examples 1-8, further optimized as follows:
[0110] (7)
[0111] (8)
[0112] (9)
[0113] (10)
[0114] (11)
[0115] in, It is the first The number of clients in the group and These represent computing costs and memory costs, respectively.
[0116] Example 10:
[0117] A precise aggregation method for architecture generation and hierarchy alignment for heterogeneous federated learning, with technical content identical to any one of embodiments 1-9, further comprising, in step 8), the step of the server aggregating gradient update parameters including:
[0118] Step 8.1) For each client and architecture The server performs a virtual forward and backward propagation on the supernet, activating the server belonging to the supernet. The path generates a binary mask. This identifies the parameters trained by the client.
[0119] binary mask As shown below:
[0120] (12)
[0121] In the formula, For loss;
[0122] Step 8.1) Apply the generated mask to the global supernet weights After updating, we get:
[0123] (13)
[0124] In the formula, This is the learning rate.
[0125] Example 11:
[0126] This paper presents a method for architecture generation and hierarchical alignment-based precise aggregation for heterogeneous federated learning. The technical content is the same as any one of Examples 1-10. Furthermore, the task is an image recognition task, such as pedestrian detection in smart security and autonomous driving scenarios. In practical applications, this method primarily addresses the problem of terminals with extremely unbalanced computing power (such as high-performance servers and low-power IoT devices) being unable to collaboratively train the same visual model. For example, in smart security and autonomous driving scenarios, this solution allows high-performance vehicle chips and low-power roadside monitoring probes to jointly train a pedestrian detection model. Through hierarchical nesting and precise aggregation mechanisms, it ensures the effective fusion of feature maps with different resolutions and channel numbers. In smart healthcare scenarios, it can adapt to the performance differences of CT / MRI equipment in hospitals of different levels, jointly training lesion segmentation models.
[0127] Example 12:
[0128] An architecture generation and hierarchical alignment-based precise aggregation method for heterogeneous federated learning includes the following steps:
[0129] 1) Initialize the global supernet and pre-train it. Define a global supernet containing the maximum search space. In the public agent dataset Pre-training is performed on the supernet to obtain trained supernet weights, which serve as the basis for the weights of subsequent subnets.
[0130] 2) Next, to leverage reinforcement learning (RL)-based agents to explore the search space of the supernet and build a high-quality pool of candidate architectures, we construct a Markov Decision Process (MDP) comprising states, actions, and rewards:
[0131] state( ): Use low-dimensional random exploration signals as states After feature extraction, the data is input into the policy network.
[0132] action( ):action It is a composite tuple ,in It is a width coefficient vector obtained from Gaussian distribution sampling, used to define the number of channels in each layer; It is a depth index obtained through Softmax sampling, used to determine the early regression position (i.e., effective depth) of the network.
[0133] award( To balance performance and diversity in the architecture, a dual-objective reward function is employed. Total reward The calculation formula is:
[0134]
[0135] 2.1) First is the basic reward ( To evaluate the intrinsic quality of the architecture, and to avoid costly training and evaluation, we use the nuclear norm as a proxy metric for performance. This is combined with the variance and range of the width vector to encourage non-uniform distribution. The calculation formula is:
[0136]
[0137]
[0138] Among them, proxy indicators Defined as a subnetwork In proxy data batch Feature maps of all activated convolutional layers The sum of the nuclear norms:
[0139]
[0140] 3) To prevent the agent from getting trapped in local optima and to improve exploration efficiency, three progressive strategies are introduced during the exploration process in step 2):
[0141] 3.1) The first is Region-based Search (RS), where we divide the search space into... Each region is a logical sub-region. During the adaptive phase, the historical innovation rate of each region is considered. Probabilistically select the current search area Select area The probability is calculated as follows:
[0142]
[0143] in This is the only architecture found in this region. The total number of architectures generated for this region.
[0144] 3.2) Next is the Adaptive Exploration Strength (AES), the purpose of which is to dynamically adjust the agent's exploration strength (such as the entropy regularization coefficient). Based on the uniqueness ratio of the architecture generated in the current batch. Adjustments will be made:
[0145]
[0146] When the repetition rate is high ( Increase exploration intensity when it is increased, and decrease it when it is decreased.
[0147] 3.3) Finally, there is the Periodic Partial Reset (PPR), which serves as the ultimate fail-safe mechanism. At regular intervals, it uses a binary mask to perform a partial reset. (Following a Bernoulli distribution) Randomly reset some parameters of a policy network To break the inertia of strategy:
[0148]
[0149] 4) Repeat steps 2) and 3) until a sufficient number of Pareto optimal architectures are collected to form a candidate architecture pool. .
[0150] 5) Next, we will proceed with the online dynamic federated training process. Let's assume we are currently in the federated learning communication round. The server retrieves the set of clients currently participating in training. And based on their hardware resource budget (such as computational load FLOPS, parameter count Params), the clients are divided into An orderly resource hierarchy ,in Representing the group with the most limited resources, The group that represents the most abundant resources.
[0151] 6) The server needs to be from the architecture pool. The middle is for each resource level Choose an optimal subnetwork architecture To address the parameter space mismatch issue, we model model selection as a constrained optimization problem. The optimization objective is to maximize the expected proxy performance (nuclear norm) for all client groups. The sum of these factors must simultaneously satisfy strict hierarchical nesting constraints and the resource budgets of each group.
[0152]
[0153]
[0154]
[0155] in, It is the first The number of clients in the group and These represent computational cost and memory cost, respectively. Strict hierarchical constraints (… The definition of ) is to ensure mathematical accuracy during aggregation, for smaller models. It must be a large model A strict subset of, specifically satisfying two nesting conditions, the first being deep nesting: (The early exit point of the smaller model must precede or be the same as that of the larger model). The second is width prefix locking: for all shared levels. The width vector must be an exact match:
[0156]
[0157] 7) Next, we need to solve for the optimal nested architecture group, using the beam search algorithm to solve the above optimization problem. The search process adopts a strategy of ascending from the lowest resource level. Initially, prioritize selecting the optimal model for the most restricted client. To use it as a subsequent level The hard constraints are applied, and the search proceeds upwards level by level until an optimal nested architecture for this round is determined. .
[0158] 8) The server distributes the selected architecture and its corresponding model parameters to the clients with the appropriate resource levels. The clients then use their local private data for training and gradient updates.
[0159] 9) After the client uploads the update, the server performs an aggregation operation. To eliminate parameter interference between heterogeneous subnets, a gradient tracking mechanism is used for accurate aggregation:
[0160] 9.1) The first step is for each client and the architecture it uses The server performs a virtual forward and backward propagation on the supernet, activating only those belonging to... The path. A binary mask is generated by detecting whether the gradient is non-zero. This is used to precisely identify which parameters the client trained:
[0161]
[0162] 9.2) Next, the generated mask is used to apply the global supernet. The update is now in effect. The aggregation formula uses a normalized weighted average, where the denominator precisely counts how many clients each parameter has been trained on, thus eliminating interference from parameter mismatch.
[0163]
[0164] 10) Finally, the server will communicate the rounds (time steps). Increment by 1. Determine if the preset maximum number of communication rounds has been reached. If yes, end the training; otherwise, return to step 5) to perform a new dynamic model selection and training for the client distribution in the next round.
Claims
1. A precise aggregation method for architecture generation and hierarchical alignment in heterogeneous federated learning, characterized in that, Includes the following steps: Step 1) Initialize the global supernet and in the public proxy dataset Pre-training is performed on the network to obtain the trained global supernet weights; Step 2) Establish a Markov decision process; Step 3) The reinforcement learning-based agent explores the global supernet search space according to the Markov decision process until the number of Pareto optimal architectures exceeds a preset value, forming a candidate architecture pool. ; Step 4) The server obtains the set of clients currently participating in training. And divide the client into based on the client hardware resource budget. An orderly resource hierarchy; ; Step 5) Build for resource level Select the optimal subnetwork architecture The optimization model; Step 6) Solve the optimization model using the beam search algorithm, and select from the candidate architecture pool. The middle level is for each resource level. Select the optimal subnetwork architecture ; Step 7) Send the selected optimal sub-network architecture and corresponding model parameters to the clients of the appropriate resource level; The client uses local private data to train the model, calculates and uploads gradient update parameters; Step 8) The server aggregates the gradient update parameters and updates the global supernet weights; Step 9) Determine whether the current communication round t has reached the preset maximum communication round. If yes, end the training and output the trained model. If no, set t = t + 1 and return to step 4). Step 10) Use the trained model to perform the tasks uploaded by the customer.
2. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 1, characterized in that, The task is an image recognition task.
3. The method for architecture generation and hierarchical alignment-based precise aggregation for heterogeneous federated learning according to claim 1, characterized in that, In step 1), the initialized global hypernet contains the maximum search space, as well as the configuration of all possible levels and channels within the search space; the architecture of the global hypernet is determined by the maximum depth and maximum width allowed in the search space.
4. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 1, characterized in that, Markov decision processes include states ,action and rewards Among them, state Low-dimensional random exploration signal; action = ,in It is a width coefficient vector obtained from Gaussian distribution sampling, used to define the number of channels in each layer; It is a depth index obtained through Softmax sampling, used to determine the early regression position of the network; award As shown below: （1） In the formula, Basic reward; reward ; This is the currently generated sub-network architecture. It is a collection of historical schemas stored in a sliding window. It is a history structure within a sliding window; It is the weighted Euclidean distance in the architecture parameter space; , As weight; Basic Rewards As shown below: （2） In the formula, For proxy indicators; The variance of the width vector; , These are the maximum and minimum values of the width vector; , , As weight; proxy indicators As shown below: （3） In the formula, For proxy data batches; To activate the feature map of the convolutional layer.
5. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 1, characterized in that, When exploring the search space of the global supernet, a region search strategy, an adaptive exploration intensity strategy, and a periodic partial reset strategy are adopted.
6. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 5, characterized in that, The regional search strategy refers to dividing the search space of the global supernet into regions. Each region is a logical sub-region; during the adaptive phase, the historical innovation rate of each region is considered. Probabilistically select the current search area ; Among them, select area As the current search area The probabilities are as follows: （4） In the formula, This is the only architecture found in this region. The total number of architectures generated for this region; This is a smoothing factor.
7. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 5, characterized in that, The adaptive exploration intensity strategy is used to adjust the agent's exploration intensity; The agent's exploration intensity in the next exploration As shown below: （5） In the formula, The uniqueness ratio of the generated architecture for the current batch; , The preset ratio threshold; Current exploration intensity; It is an increasing factor; To reduce factors.
8. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 5, characterized in that, The periodic partial reset strategy refers to: every preset batch, using a binary mask that follows a Bernoulli distribution. Randomly reset some parameters of the policy network In order to break the inertia of strategy; Parameters during the next exploration As shown below: （6） In the formula, These are the initial parameters randomly generated when the network begins training; For policy networks In the The current real-time parameters for each training batch.
9. The method for architecture generation and hierarchical alignment precise aggregation for heterogeneous federated learning according to claim 1, characterized in that, The optimized model is shown below: （7） （8） （9） （10） （11） in, It is the first The number of clients in the group and These represent computing costs and memory costs, respectively.
10. The method for architecture generation and hierarchical alignment-based precise aggregation for heterogeneous federated learning according to claim 1, characterized in that, In step 8), the server's aggregation of gradient update parameters includes: Step 8.1) For each client and architecture The server performs a virtual forward and backward propagation on the supernet, activating the server belonging to the supernet. The path generates a binary mask. This identifies the parameters trained by the client. binary mask As shown below: （12） In the formula, For loss; Step 8.1) Apply the generated mask to the global supernet weights After updating, we get: （13） In the formula, This is the learning rate.