Model parameter optimization method based on combined multi-armed bandit weighted fusion

Abstract

The application relates to a model parameter optimization method based on combined multi-arm slot machine weighted fusion, which comprises the following steps: determining an initial common parameter point; training a plurality of models under different data enhancement strategies according to the initial common parameter point; dynamically adjusting the weight of each model in final parameter fusion through a combined multi-arm slot machine algorithm; performing weighted average on the parameters of all models according to the weight calculated by the combined multi-arm slot machine algorithm to obtain a common parameter point for next iteration; and repeating the foregoing steps until a preset training round is reached to obtain an updated global model. The application has the beneficial effect that the combined multi-arm slot machine algorithm is used to intelligently adjust the weight of each model according to the performance of the model in the current training stage, so that effective fusion of the information of the plurality of models is realized, and the limitations of fixed weight or simple arithmetic average in traditional optimization methods are avoided.

Description

Technical Field

[0001] This invention relates to the field of model parameter optimization technology, and more specifically, to a model parameter optimization method based on combined multi-armed slot machine weighted fusion. Background Technology

[0002] With the rapid development of deep learning technology, optimizers, as key factors driving model training efficiency and performance, have become particularly important in design and improvement. Traditional optimizers such as Stochastic Gradient Descent (SGD), while simple and efficient, are prone to getting trapped in local optima when faced with complex data distributions and non-convex loss functions, resulting in poor model generalization ability. To address this issue, researchers have proposed various improvement methods, among which Weight Averaging (WA) has attracted much attention due to its ability to effectively smooth the loss surface and reduce the variance between model parameters, thereby improving model generalization ability. In recent years, researchers have proposed several WA-based optimizers, such as Lookahead and SWA. These optimizers achieve model parameter diversification and smoothing by continuously integrating the weights of multiple models during training, thus achieving significant results in improving model generalization ability.

[0003] However, existing data augmentation (WA) methods have certain limitations in application. On the one hand, many WA methods rely too heavily on fixed data augmentation or model diversity strategies, resulting in unsatisfactory performance under complex data distributions. On the other hand, some WA methods, while maintaining model diversity, are prone to violating the principle of locality, meaning that the range of model parameters changing too much during training affects the stability and convergence speed of the model.

[0004] In recent years, an optimizer called Lookaround has been proposed, achieving a balance between diversity and locality by iteratively performing a "around step" and an "averaging step." In the "averaging step," multiple models start from the same initial point and are trained on data processed with different data augmentation techniques, thus increasing model diversity. In the "averaging step," the parameters of these models are averaged to ensure parameter locality throughout the training process. However, although Lookaround achieves a balance between diversity and locality, its averaging strategy still uses a simple arithmetic average, which ignores the different performance characteristics that different models may exhibit during training, and its reliance on a limited variety of data augmentations is also limited. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a model parameter optimization method based on weighted fusion of combined multi-armed slot machines.

[0006] Firstly, a method for optimizing model parameters based on weighted fusion of combined multi-armed slot machines is provided, including:

[0007] Step 1: Determine the initial common parameter points. Multiple models are trained under different data augmentation strategies based on the initial common parameter points.

[0008] Step 2: By combining multi-armed slot machine algorithms, dynamically adjust the weight of each model in the final parameter fusion;

[0009] Step 3: Based on the weights calculated by the combined multi-armed slot machine algorithm, perform a weighted average of the parameters of all models to obtain the common parameter points for the next iteration;

[0010] Step 4: Repeat steps 1-3 until the preset number of training rounds is reached to obtain the updated global model.

[0011] Preferably, step 1 includes:

[0012] Step 1.1: Set up the optimizer and the objective function for evaluating model performance; randomly sample mini-batch data from the training dataset and set up various data augmentation strategies; obtain multiple models and initial common parameter points.

[0013] Step 1.2: Each model is trained under a specific data augmentation strategy, and the gradient is calculated and the model parameters are updated by the optimizer.

[0014] Preferably, step 2 includes:

[0015] Step 2.1: Configure the model weights in the optimizer as base arms, and multiple base arms form super arms. Maximize the cumulative revenue by selecting super arms.

[0016] Step 2.2: After each iteration, sample the original training set data without data augmentation, test the performance of each model, and collect the performance data of each model.

[0017] Step 2.3: The optimizer calculates dynamic weights for each superarm based on the collected performance data.

[0018] As a preferred option, it also includes:

[0019] Step 5: Verify the performance of the updated global model on the test set and draw charts to observe the training process.

[0020] Preferably, in steps 1-4, the same mixed data is used for training in all training rounds.

[0021] Preferably, in step 1, the data augmentation strategy includes: random horizontal flipping, random vertical flipping, rotation, cropping, random erasure, and unsupervised augmentation.

[0022] Secondly, a model parameter optimization system based on combined multi-armed slot machine weighted fusion is provided for performing any of the methods described in the first aspect, including:

[0023] The determination module is used to determine the initial common parameter points, and multiple models are trained under different data augmentation strategies based on the initial common parameter points.

[0024] The adjustment module is used to dynamically adjust the weight of each model in the final parameter fusion by combining the multi-armed slot machine algorithm;

[0025] The weighted average module is used to perform a weighted average of the parameters of all models based on the weights calculated by the combined multi-armed slot machine algorithm, so as to obtain the common parameter points for the next iteration.

[0026] The repeat module is used to make the determination module, adjustment module, and weighted average module work repeatedly until the preset number of training rounds are reached to obtain the updated global model.

[0027] Thirdly, a computer storage medium is provided, wherein a computer program is stored therein; when the computer program is run on a computer, the computer causes the computer to perform any of the methods described in the first aspect.

[0028] Fourthly, an electronic device is provided, comprising:

[0029] Memory, used to store computer programs;

[0030] A processor for executing the computer program to implement the method as described in any of the first aspects.

[0031] The beneficial effects of this invention are: during the training process, this invention maintains multiple models under different data augmentation, each model is updated under different data augmentation and training strategies, and by combining multi-armed slot machine algorithms, the weights of each model are intelligently adjusted according to the performance of each model in the current training stage, so as to achieve effective fusion of information from multiple models, avoiding the limitations of fixed weights or simple arithmetic averages in traditional optimization methods. Attached Figure Description

[0032] Figure 1 A flowchart illustrating a model parameter optimization method based on weighted fusion of combined multi-armed slot machines provided in an embodiment of the present invention;

[0033] Figure 2a This is a schematic diagram of the model using fixed-weight fusion on the loss basin.

[0034] Figure 2b This is a schematic diagram of the model using direct average fusion on the loss basin;

[0035] Figure 2c This is a schematic diagram of the model using weighted average fusion on the loss basin;

[0036] Figure 3 This is a schematic diagram of the structure of the model parameter optimization system based on combined multi-armed slot machine weighted fusion provided in an embodiment of the present invention. Detailed Implementation

[0037] The present invention will be further described below with reference to embodiments. The description of the embodiments below is only for the purpose of helping to understand the present invention. It should be noted that those skilled in the art can make several modifications to the present invention without departing from the principle of the present invention, and these improvements and modifications also fall within the protection scope of the claims of the present invention.

[0038] Example 1:

[0039] The Combinatorial Multi-Armed Bandit (CMAB) algorithm is a powerful tool for handling optimal choice problems in unknown environments, particularly adept at intelligently balancing exploration and exploitation among multiple options. Through continuous trial and error and learning, the CMAB algorithm can dynamically adjust its preferences for different model configurations under uncertainty, thereby finding the optimal or near-optimal model combination, which to some extent outperforms the arithmetic mean. Building upon the multi-armed bandit algorithm, the CMAB incorporates linear and nonlinear reward models to predict based on reward functions and can adapt to different feedback types, such as full-information feedback, semi-robber feedback, and robber feedback. Furthermore, common CMAB methods include those involving LLR, CUCB, and CMAB-T.

[0040] To address the problems of existing technologies, Embodiment 1 of this application combines the decision optimization mechanism of the multi-armed slot machine algorithm with the Lookaround optimizer, aiming to achieve faster convergence speed and higher stability by intelligently selecting and adjusting the weights of multiple fused models during the optimization process.

[0041] Specifically, Embodiment 1 of this application provides a model parameter optimization method based on weighted fusion of combined multi-armed slot machines, such as... Figure 1 As shown, it includes:

[0042] Step 1, the wraparound step: Initial common parameter points are determined. In each iteration, multiple models are trained using different data augmentation strategies based on these initial common parameter points. These models will exhibit different performance characteristics during training.

[0043] Step 1 includes:

[0044] Step 1.1: Set up the optimizer and the objective function for evaluating model performance; randomly sample mini-batch data from the training dataset and set up various data augmentation strategies; obtain multiple models and initial common parameter points.

[0045] For example, at the beginning of the loop step, it is necessary to set up optimizer A and objective function L for evaluating model performance, and randomly draw a mini-batch of data B from the training dataset D. To increase data diversity and the model's generalization ability, N different data augmentation strategies AUG = {AUG1,...,AUG} are set. d Subsequently, N different models were assigned, each starting from a common parameter point φ0.

[0046] Step 1.2: Each model is trained under a specific data augmentation strategy, and the gradient is calculated and the model parameters are updated by the optimizer.

[0047] In step 1.2, after initialization, the training phase begins. The key to this phase is using N different data augmentation strategies (AUG). Since each model is trained under its specific data augmentation strategy, they learn different data transformation patterns, which helps increase the model's robustness. During each batch step in the training process, the model's parameters need to be optimized and updated.

[0048] For example, for each model j (j = 1, ..., N) and each step k (k = 1, ..., k batch Using optimizer A and the current data augmentation strategy AUG j To calculate the gradient of the parameters, and update the model parameters θ based on the gradient. t,j,k Specifically, θ t,j,k θ represents the model parameters trained on the k-th batch of data for the j-th model at time step t. t,j,k-1 This represents the model parameters trained on the (k-1)th batch of data for the j-th model at time step t. The update formula is: θ t,j,k =θ i,j,k-1 +

[0049] A(L,θ t,j,k-1 AUG j (B))

[0050] Where A is the optimizer, L is the objective function, and AUG is the target function. j (B) represents the batch of data processed using the current data augmentation strategy. This allows for the gradual adjustment of the model's parameters to better adapt to the training data and move towards minimizing the objective function.

[0051] Step 2, CMAB weighting strategy: The weight of each model in the final parameter fusion is dynamically adjusted through the CMAB algorithm.

[0052] In step 2, to more intelligently integrate the parameters of these models, this invention defines a combined multi-armed slot machine problem in the optimizer and introduces the CMAB algorithm. The CMAB algorithm maximizes the cumulative reward by continuously selecting among multiple arms—that is, among models trained with specific data augmentation strategies. The CMAB algorithm dynamically adjusts the weight of each network in the final average parameters based on its historical performance.

[0053] Specifically, step 2 includes:

[0054] Step 2.1: Configure the model weights in the optimizer as base arms, and multiple base arms form super arms. Maximize the cumulative revenue by selecting super arms.

[0055] In step 2.1, this application first defines the combinatorial multi-armed slot machine problem: in an optimizer scenario, it is set on a discrete-time system where, at each time step t, the system generates a reward based on an unknown stochastic process. In this context, each arm can be analogized to different model weight configurations in an optimizer. Unlike the traditional multi-armed slot machine problem, in CMAB, this application no longer selects a single arm, but rather a combination of multiple base arms (i.e., a superarm). The decision-maker needs to maximize the long-term cumulative gain or performance through a series of trials. Since the performance of different superarms is often unknown and may be affected by various factors (such as data distribution, system state, etc.), the decision-maker needs to make a trade-off between exploring new combinations and utilizing known optimal combinations.

[0056] Specifically, the combinatorial multi-armed slot machine problem in the optimizer is defined as follows: Given a discrete-time system with N unknown stochastic processes X... i (t), 1≤i≤N, with time indexed by t. The system proceeds at each time step t according to an unknown stochastic process. Generate rewards, where N is the number of potential base arms, which in the optimizer problem can be viewed as the number of models under different data augmentations. Assume the values ​​of these random variables are normalized to X. i (t)∈[0,1], and it is not necessary to assume X i (t) are independent of each other with respect to i. The mean of this random process is set to μ. i =E(X) iThe action vector (superarm) is unknown to the decision-maker. At each time step t, an N-dimensional action vector (superarm) a(t) is selected, corresponding to a policy of π(n) in a finite set F. Each action a(t) consists of a set of indices of the base arms, i.e., a(t) = (a1(t), a2(t), ..., a...). N (t)), and a i (t)≥0.

[0057] When action a(t) is selected, only those corresponding to non-zero sub-actions (a i A random variable X (t) ≠ 0 i The value of (t) is observed. This application defines A. a(t) ={i:a i (t)≠0,1≤i≤N} represents the action a in action a. i The set of indices for (t)≠0.

[0058] The reward function can be defined as:

[0059]

[0060] The multi-armed slot machine problem typically uses regret to evaluate strategies. Regret is defined as the difference between the expected reward obtained by consistently choosing the optimal action and the expected reward obtained with a given strategy. Minimizing regret is equivalent to maximizing reward. Regret can be expressed as:

[0061]

[0062] in, μ i Let t be the average reward of the i-th model, and t be the current time step.

[0063] Step 2.2, Performance Data Collection: After each iteration, the original training set data without data augmentation is sampled to test the performance of each model and collect the performance data of each model.

[0064] In step 2.2, performance data is used to update the cumulative reward for each arm in the CMAB algorithm. Here, cross-entropy loss is used as the reward metric, expressed as:

[0065]

[0066] Among them, y i These are real tags, with values ​​of 0 or 1, p i Let be the probability of the i-th class in the dataset, and n be the total number of classes. Assume the total number of training epochs is T, and its range is L(y,p)∈[0,+∞), and the cross-entropy generally decreases with the number of training epochs. Here, the sigmoid function is used for scaling, with the addition of a negative sign to indicate incremental or decremental changes, expressed as:

[0067]

[0068] This causes it to increase with each training round.

[0069] Step 2.3, Dynamic Weight Calculation: The optimizer calculates dynamic weights for each superarm based on the collected performance data.

[0070] In step 2.3, the optimizer calculates a dynamic weight for each superarm based on the collected performance data. This weight reflects the expected benefit or performance of selecting that superarm in the current context. According to X... i (t) During the optimization process, the trend of first increasing and then converging with time step transforms the non-stationary combinatorial multi-armed slot machine problem into a piecewise stationary problem.

[0071] During the rising phase, n i (t) represents the total number of times arm i is pulled at time t. The reward for each arm can be estimated as:

[0072]

[0073] in, and Represent Estimates of the slope and intercept for linear modeling.

[0074] The upper confidence bound for each arm can be defined as:

[0075]

[0076] Where δ∈(0,1) are hyperparameters;

[0077] The first round is a warm-up phase, where each model participates in the fusion and relevant metrics are recorded and updated. Subsequent selections require determining the action a(t) for selecting a subset of models in the next round based on the following strategy, expressed as:

[0078]

[0079] Where L represents the maximum number of models selected to participate in model fusion. At each time step t, an N-dimensional action vector (superarm) a(t) is selected, corresponding to a policy of π(n) in a finite set F. Each action a(t) consists of a set of indices of the base arms, i.e., a(t) = (a1(t), a2(t), ..., a N (t)), and a i (t)≥0, while A a(t) ={i:a i (t)≠0,1≤i≤N} represents the action a in action a. iThe set of indices for (t)≠0.

[0080] For the current step, the weight is:

[0081]

[0082] Step 3, Weighted Average Step: Based on the weights calculated by the CMAB algorithm, the parameters of all models are weighted and averaged to obtain the common parameter points for the next iteration.

[0083] Specifically, at the end of each iteration, a weighted average is calculated for the parameters of all models based on the weights calculated by the CMAB algorithm. This weighted average can be expressed as:

[0084]

[0085] Where w i θ is the weight of the i-th network. t,i These are the parameters of the i-th network at the end of iteration t.

[0086] In addition, the weighted average parameter φ t This serves as the starting point for the next iteration, i.e., the new common parameter point. This common parameter point will be used in the loop step of the next iteration as the starting point for training all models. The parameters n are updated according to the current strategy. i (t).

[0087] Step 4: Repeat steps 1-3 until the preset number of training rounds is reached to obtain the updated global model.

[0088] like Figure 2a As shown, the models participating in weight fusion often have varying performance. For example... Figure 2b As shown, the result obtained by directly averaging the model weights is easily affected by poorly performing models, leading to a larger loss in the fused model. Figure 2c As shown, if the models involved in the fusion are weighted and averaged, the fused model is less affected by the poorly performing models, resulting in a smaller loss and achieving better performance and robustness.

[0089] Example 2:

[0090] Based on Example 1, Example 2 of this application provides a more specific model parameter optimization method based on combined multi-armed slot machine weighted fusion, which aims to dynamically adjust the weights of different networks in the final model aggregation by introducing a combined multi-armed slot machine algorithm, thereby achieving more efficient model integration.

[0091] Experiments were conducted on the CIFAR10 and CIFAR100 datasets. For each epoch in all experiments, training was performed using the same mixed data with six data augmentation options: random horizontal flip, random vertical flip, rotation, cropping, random erasure, and RandAugment. This ensured fair comparisons between different methods. For the training process, we decayed the initial learning rate by 0.1 at epochs 60, 120, and 160 using a discount factor of size 0.2.

[0092] Specifically, the above methods include:

[0093] Step 1: Initialize the network settings, including necessary preparations such as dividing the dataset and configuring the optimizer.

[0094] The specific steps in step 1 are as follows:

[0095] Step 1.1, Dataset Splitting: Both the CIFAR10 and CIFAR100 datasets contain 60,000 images, of which 50,000 are used for training and 10,000 for testing. 500 un-data-augmented images are stratified and sampled from the original data for validation.

[0096] Step 1.2, Optimizer Configuration: Select SGD as the base optimizer.

[0097] Step 1.3, Model Architecture Definition: Experiments validated different methods on multiple network architectures, such as VGG19, ResNet50, ResNet101, ResNet152, and ResNeXt50, and all methods were performed on well-validated parameters.

[0098] Step 1.4, Data Augmentation Method Setup: For each round in all experiments, we trained using the same mixed data and used six data augmentation options: random horizontal flip, random vertical flip, rotation, cropping, random erasure, and unsupervised augmentation (RandAugment). This ensured fair comparisons between different methods.

[0099] Step 2: Conduct a warm-up phase, test each model, and record the reward metrics.

[0100] The specific steps in step 2 are as follows:

[0101] Step 2.1: Train a model for each data augmentation method.

[0102] Step 2.2: Evaluate the performance of the initially trained model on the validation set and record the key reward metrics.

[0103] Step 2.3, Update n i(t) and weights, etc. Develop the next selection strategy a. Initialize N copies of the model parameters θ. t,j,0 , where j represents different data augmentation methods, t represents the current training round, and 0 represents the starting state of the training round.

[0104] Step 3: Perform a wraparound step, which involves reinforcing the training of multiple model versions with various datasets.

[0105] The specific steps in step 3 are as follows:

[0106] Step 3.1: Sample mini-batch data B from the training dataset Dt. t .

[0107] Step 3.2: For each enhanced version of model j, update its parameters θ using optimizer A. t,j,i , where i represents the current step, AUG j (B t ) indicates data B t Apply the j-th enhancement method.

[0108] Step 4: Calculate the weights, which means evaluating the model's performance on the validation set using a combined multi-armed slot machine method and assigning different weights.

[0109] The specific steps in step 4 are as follows:

[0110] Step 4.1: Using the combined multi-armed slot machine algorithm, calculate the expected return for each model version based on its performance on the validation set. From the validation dataset D... v Medium-sampled small batch data B v For each enhanced version of model j, use the validation set B. v Calculate its reward R j .

[0111] Step 4.2: Calculate the weight W based on the reward list R. i .

[0112] Step 5: Perform the weighted average step, calculate the weighted average of the parameters of multiple model versions based on the calculated weights, and update the global parameters.

[0113] The specific steps in step 5 are as follows:

[0114] Step 5.1: Based on the weights calculated in step (4), perform a weighted average on the parameters of each model version.

[0115] Step 5.2: Verify the performance of the updated global model on the test set and draw charts to observe the training process.

[0116] This application's embodiments conducted experiments on the CIFAR10 and CIFAR100 datasets, using SGDM as a benchmark, and compared the proposed method (Look&Select) with SWA, SWAD, Lookahead, and Lookaround. This application validated different methods on various network architectures, including VGG19, ResNet50, ResNet101, ResNet152, and ResNeXt50, etc., trained and compared using various optimizers under random initialization. The results are shown in Table 1. TOP1 accuracy refers to the proportion of the class with the highest predicted probability that matches the actual class; TOP5 accuracy refers to the proportion of the top five predicted probabilities that include the actual class; NLL loss measures the difference between the model's predicted probability distribution and the true distribution; a smaller value indicates better model performance; "#param." represents the number of model parameters, reflecting the model's complexity and capacity.

[0117] Table 1: Accuracy and NLL Loss of Randomly Initialized Models on the Test Set

[0118]

[0119] As shown in Table 1, Look&Select achieved the best results on both CIFAR10 and CIFAR100 data centers, as well as on model architectures such as VGG19, ResNet50, ResNet101, ResNet152, and ResNeXt50.

[0120] Furthermore, to explore the scalability of the Look&Select method, this application applies it to scenarios requiring pre-training, including pre-training scenarios with CNN or ViT architectures. The test set accuracy and NLL loss for all methods were obtained during training using pre-trained models. All models were fine-tuned for the same amount of time at a 224*224 resolution, and the results are shown in Table 2.

[0121] Table 2: Accuracy and NLL Loss of the Fine-Tuned Pre-Trained Model on the Test Set

[0122]

[0123] As shown in Table 2, the Look&Select method also performs well in model fine-tuning tasks, and its fine-tuning results on ResNet50 and ViT-B / 16 architectures are also better than those of SWA, Lookahead and other comparative methods.

[0124] In summary, the Look&Select optimization method combines the advantages of model selection and model fusion, and outperforms existing comparative methods in optimization tasks involving random initialization and fine-tuning.

[0125] It should be noted that the parts in this embodiment that are the same as or similar to those in Embodiment 1 can be referred to each other, and will not be repeated in this application.

[0126] Example 3:

[0127] Based on Example 1, Example 3 of this application provides a model parameter optimization system based on combined multi-armed slot machine weighted fusion, such as... Figure 3 As shown, it includes:

[0128] The determination module is used to determine the initial common parameter points, and multiple models are trained under different data augmentation strategies based on the initial common parameter points.

[0129] The adjustment module is used to dynamically adjust the weight of each model in the final parameter fusion by combining the multi-armed slot machine algorithm;

[0130] The weighted average module is used to perform a weighted average of the parameters of all models based on the weights calculated by the combined multi-armed slot machine algorithm, so as to obtain the common parameter points for the next iteration.

[0131] The repeat module is used to make the determination module, adjustment module, and weighted average module work repeatedly until the preset number of training rounds are reached to obtain the updated global model.

[0132] Specifically, the system provided in this embodiment is the same as the system corresponding to the method provided in embodiment 1. Therefore, the parts in this embodiment that are the same as or similar to those in embodiment 1 can be referred to each other and will not be described again in this application.

Claims

1. A model parameter optimization method based on weighted fusion of combined multi-armed slot machines, characterized in that, include: Step 1: Determine the initial common parameter points of the image classification models. Multiple image classification models are trained according to the initial common parameter points under different image data augmentation strategies. In step 1, the image data augmentation strategies include: random horizontal flipping, random vertical flipping, rotation, cropping, random erasing, and unsupervised augmentation; Step 1 includes: Step 1.1: Set up the optimizer and the objective function for evaluating the performance of the image classification model; randomly sample small batches of image data from the training dataset and set up various image data augmentation strategies; obtain multiple image classification models and initial common parameter points. Step 1.2: Each image classification model is trained under a specific image data augmentation strategy, and the gradient is calculated and the model parameters are updated through the optimizer; Step 2: By combining the multi-armed slot machine algorithm, dynamically adjust the weight of each image classification model in the final parameter fusion; Step 2 includes: Step 2.1: Configure the model weights in the optimizer as base arms, and multiple base arms form super arms. Maximize the cumulative revenue by selecting super arms. Step 2.2: After each iteration, sample the original training set data without data augmentation, test the performance of each model, and collect the performance data of each model. Step 2.3: The optimizer calculates dynamic weights for each superarm based on the collected performance data; Step 3: Based on the weights calculated by the combined multi-armed slot machine algorithm, perform a weighted average of the parameters of all image classification models to obtain the common parameter points for the next iteration; Step 4: Repeat steps 1-3 until the preset number of training rounds is reached to obtain the updated global image classification model.

2. The model parameter optimization method based on weighted fusion of combined multi-armed slot machines according to claim 1, characterized in that, Also includes: Step 5: Verify the performance of the updated global model on the test set and draw charts to observe the training process.

3. The model parameter optimization method based on weighted fusion of combined multi-armed slot machines according to claim 2, characterized in that, In steps 1-4, the same mixed data is used for training in all training rounds.

4. A model parameter optimization system based on weighted fusion of combined multi-armed slot machines, characterized in that, For performing the method according to any one of claims 1 to 3, comprising: The determination module is used to determine the initial common parameter points of the image classification models. Based on the initial common parameter points, multiple image classification models are trained under different image data augmentation strategies. The adjustment module is used to dynamically adjust the weights of each image classification model in the final parameter fusion by combining the multi-armed slot machine algorithm; The weighted average module is used to perform a weighted average of the parameters of all image classification models based on the weights calculated by the combined multi-armed slot machine algorithm, so as to obtain the common parameter points for the next iteration. The repeat module is used to make the determination module, adjustment module, and weighted average module work repeatedly until the preset number of training rounds are reached to obtain the updated global image classification model.

5. A computer storage medium, characterized in that, The computer storage medium stores a computer program; when the computer program is run on the computer, it causes the computer to perform the method described in any one of claims 1 to 3.

6. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the method as described in any one of claims 1 to 3.

Images