A multi-weight self-adjusting comprehensive evaluation prediction system

Abstract

The present application relates to the technical field of prediction model, and disclose a kind of multi-weight self-adjusting comprehensive evaluation prediction system, including system module assembly, the system module assembly includes model training module and model prediction module, and model training module includes forward model set training module and reverse model set training module, model prediction module includes forward prediction module and reverse prediction module, the model training module is first collected in binary logic system when running, forward, reverse image sample set, and is merged and classified as S1 and S0;The model of the application contains two model sets of positive and negative, each model set contains several single models, each model has independent prediction weight, the weight is referenced linear rectification principle, can be constantly rolling update, form self-learning ability, on the basis of without frequently retraining to improve the learning ability and prediction accuracy of model.

Description

Technical Field

[0001] This invention relates to the field of prediction model technology, specifically to a multi-weight self-adjusting comprehensive evaluation prediction system. Background Technology

[0002] With the continuous progress of the times, AI underlying technologies based on neural networks and other algorithms have expanded their application scope from the initial digit recognition and speech extraction to a wider range of applications such as object recognition, contour recognition, and multi-target classification. This has improved the level of intelligence in fields such as smart communities, surveillance cameras, traffic violation recognition, and logistics sorting, and further promoted the development of various industries in society.

[0003] Predictive models are generally single-model structures, meaning they are built by creating a custom underlying algorithm (neural network or other algorithms) or by training an existing public model on your own sample data to form a model that can recognize specific sample data. This model is then used for predictive applications to achieve AI recognition effects. However, any model may have prediction errors due to factors such as sample data, training methods, and weights during training. If these errors are not corrected, they will persist and affect prediction accuracy unless the model is retrained by changing the sample data, adjusting the training weights and methods. This process requires time and other related resources, and even after retraining, the model cannot guarantee that the range of prediction errors will be further optimized. Therefore, a multi-weight self-adjusting comprehensive evaluation and prediction system is proposed. Summary of the Invention

[0004] (a) Technical problems to be solved To address the shortcomings of existing technologies, this invention provides a multi-weight self-adjusting comprehensive evaluation and prediction system. The model includes two model sets, one positive and one negative. Each model set contains several individual models, and each model has an independent prediction weight. This weight is based on the principle of linear rectification and can be continuously updated to form a self-learning capability. This improves the model's learning ability and prediction accuracy without the need for frequent retraining.

[0005] (II) Technical Solution To achieve the above objectives, the present invention provides the following technical solution: A multi-weight self-adjusting comprehensive evaluation and prediction system includes a system module assembly. The system module assembly includes a model training module and a model prediction module. The model training module includes a forward model set training module and a backward model set training module. The model prediction module includes a forward prediction module and a backward prediction module. During operation, the model training module first collects forward and backward image sample sets from a binary logic system and categorizes them as S1 and S0. Then, it obtains a common sample set different from the binary logic system from a network channel and labels it as S2. The forward model set training module includes the following steps during training: S1: For sample sets S1 and S2, a prediction model is built based on a neural network algorithm. This model learns from commonly used training parameters by adjusting sample data, training methods, and training weights, generating N positive prediction sub-models M1-M2. N N≥3; In S1, adjusting the sample data refers to changing the training samples, including but not limited to replacing samples or changing the number; adjusting the training method refers to changing the training method, including but not limited to changing the number of layers such as the input layer, hidden layer, and output layer in the algorithm; adjusting the training weights refers to changing the weights of each layer in the algorithm, including but not limited to the indicators of the regression function and the indicators of the fitted curve. S2: Assign initial weight factors F to the N positive prediction sub-models. i 1-F i N All of these values ​​are initialized to 1, i=0; S3: Determine the weight rectification function for the positive prediction sub-model: x refers to the judgment result of the current prediction sub-model, and d is the set judgment threshold. The significance of this rectification function is to directly discard judgment results below the set threshold and only retain judgment results equal to or higher than the set threshold. The reverse model set training module includes the following steps during training: S1: For sample sets S0 and S2, a prediction model is built based on a neural network algorithm. This model learns by adjusting commonly used training parameters such as sample data, training methods, and training weights, generating N inverse prediction sub-models RM1-RM2. N N≥3; S2: Assign initial weighting factors RF to the N backward prediction sub-models. i 1-RF i N All of these values ​​are initialized to 1, i=0; S3: Determine the weight rectification function for the inverse prediction sub-model: x refers to the judgment result of the current prediction sub-model, and d is the set judgment threshold.

[0006] As a further embodiment of the present invention, the model prediction module can be executed repeatedly.

[0007] Furthermore, the positive prediction module includes the following steps during prediction: S1: Based on N positive prediction sub-models M1-M N Make predictions and normalize the prediction result vector to ensure that its prediction probability distribution is in the interval [0,1]. S2: Calculate the maximum value P1-P among the prediction result vectors of N positive prediction sub-models. N ; S3: According to P1-P N Calculate the predicted average value Rp; S4: Based on the weight rectification function defined in the positive model set training module, substitute Rp as d, and apply the weight factor F1 to the N sub-models. i -F N i Rectification is performed to obtain a new weighting factor F1 i+1 -F N i+1 And based on the new non-zero weighting factors from P1-P N Select new effective prediction result vectors V1-V m , where (1≤m≤N); S5: Based on the new weighting factor F1 i+1 -F N i+1 For P1-P N A new prediction result P is obtained by performing a weighted average calculation. f ; S6: Utilizing V1-V m P f Calculate V1-V m For P f The variance of the standard variance is calculated, and the index of the smallest standard variance is taken as K, where K takes the value [1, m]. S7: Take P k This is the final result of this positive prediction.

[0008] Based on the aforementioned scheme, the reverse prediction module includes the following steps during prediction: S1: Based on N positive prediction sub-models RM1-RM N Make predictions and normalize the prediction result vector to ensure that its prediction probability distribution is in the interval [0,1]. S2: Calculate the maximum value RP1-RP among the prediction result vectors of N positive prediction sub-models. N ; S3: Based on RP1-RP N Calculate the predicted average value RRp; S4: Based on the weight rectification function defined in the reverse model set training module, substitute RRp as d, and apply the weight factor RF1 to the N sub-models. i -RF N i Rectification is performed to obtain a new weighting factor RF1 i+1 -RF N i+1 And based on the new non-zero weighting factor from RP1-RP N Select new effective prediction result vectors RV1-RV m , where (1≤m≤N); S5: Based on the new weighting factor RF1 i+1 -RF N i+1 For RP1-RP N A new prediction result RP is obtained by performing a weighted average calculation. f ; S6: Utilizing RV1-RV m RP f Calculate RV1-RV m For RP f The variance of the standard variance is calculated, and the index of the smallest standard variance is taken as K, where K takes the value [1, m]. S7: Get RP k This is the final result of this positive prediction.

[0009] Furthermore, the model prediction module also includes a comprehensive prediction and evaluation module, and the implementation steps of the comprehensive prediction and evaluation module include: S1: Calculate P k With RP k The difference is ∆P; S2: Based on the following formula, substitute ∆P into x to perform binary logic result judgment, d≥0.75, .

[0010] (III) Beneficial Effects Compared with the prior art, the present invention provides a multi-weight self-adjusting comprehensive evaluation and prediction system, which has the following beneficial effects: 1. In this invention, positive and negative samples of a binary logic system are trained separately to form a set of positive and negative models containing multiple sub-models. Through dual prediction in both positive and negative directions, and based on the difference between the positive and negative predictions, a comprehensive evaluation is performed to enhance the stability of the prediction model and improve the accuracy of the prediction.

[0011] 2. In this invention, multiple sub-prediction models in the positive and negative model set can be dynamically allocated and self-learned. Without the need for frequent retraining, the noise of the prediction model can be effectively reduced, the overall stability and prediction accuracy of the model can be improved, thereby overcoming the static error caused by a single model and the resources required for optimization error. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the modular framework of a multi-weight self-adjusting comprehensive evaluation and prediction system proposed in this invention. Detailed Implementation

[0013] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0014] Reference Figure 1 A multi-weight self-adjusting comprehensive evaluation and prediction system includes a system module assembly. The system module assembly includes a model training module and a model prediction module. The model training module includes a forward model set training module and a backward model set training module. The model prediction module includes a forward prediction module and a backward prediction module. The model training module generally only needs to be executed once, while the model prediction module can be executed repeatedly. During operation, the model training module first collects forward and backward image sample sets from the binary logic system and classifies them into S1 and S0. Then, it obtains a common sample set different from the binary logic system from network channels and labels it as S2. The forward model set training module includes the following steps during training: S1: For sample sets S1 and S2, a prediction model is built based on a neural network algorithm. This model learns from commonly used training parameters by adjusting sample data, training methods, and training weights, generating N positive prediction sub-models M1-M2. N N≥3; In S1, adjusting the sample data refers to changing the training samples, including but not limited to replacing samples or changing the number; adjusting the training method refers to changing the training method, including but not limited to changing the number of layers such as the input layer, hidden layer, and output layer in the algorithm; adjusting the training weights refers to changing the weights of each layer in the algorithm, including but not limited to the indicators of the regression function and the indicators of the fitted curve. S2: Assign initial weight factors F to the N positive prediction sub-models. i 1-F i N All of these values ​​are initialized to 1, i=0; S3: Determine the weight rectification function for the positive prediction sub-model: x refers to the judgment result of the current prediction sub-model, and d is the set judgment threshold. The significance of this rectification function is to directly discard judgment results below the set threshold and only retain judgment results equal to or higher than the set threshold. The reverse model set training module includes the following steps during training: S1: For sample sets S0 and S2, a prediction model is built based on a neural network algorithm. This model learns by adjusting commonly used training parameters such as sample data, training methods, and training weights, generating N inverse prediction sub-models RM1-RM2. N N≥3; S2: Assign initial weighting factors RF to the N backward prediction sub-models. i 1-RF i N All of these values ​​are initialized to 1, i=0; S3: Determine the weight rectification function for the inverse prediction sub-model: x refers to the judgment result of the current prediction sub-model, and d is the set judgment threshold.

[0015] The positive prediction module includes the following steps during prediction: S1: Based on N positive prediction sub-models M1-M N Make predictions and normalize the prediction result vector to ensure that its prediction probability distribution is in the interval [0,1]. S2: Calculate the maximum value P1-P among the prediction result vectors of N positive prediction sub-models. N ; S3: According to P1-P N Calculate the predicted average value Rp; S4: Based on the weight rectification function defined in the positive model set training module, substitute Rp as d, and apply the weight factor F1 to the N sub-models. i -F N i Rectification is performed to obtain a new weighting factor F1 i+1 -F N i+1 And based on the new non-zero weighting factors from P1-P N Select new effective prediction result vectors V1-V m , where (1≤m≤N); S5: Based on the new weighting factor F1 i+1 -F N i+1 For P1-P N A new prediction result P is obtained by performing a weighted average calculation.f ; S6: Utilizing V1-V m P f Calculate V1-V m For P f The variance of the standard variance is calculated, and the index of the smallest standard variance is taken as K, where K takes the value [1, m]. S7: Take P k This serves as the final result of this positive prediction; The reverse prediction module includes the following steps during prediction: S1: Based on N positive prediction sub-models RM1-RM N Make predictions and normalize the prediction result vector to ensure that its prediction probability distribution is in the interval [0,1]. S2: Calculate the maximum value RP1-RP among the prediction result vectors of N positive prediction sub-models. N ; S3: Based on RP1-RP N Calculate the predicted average value RRp; S4: Based on the weight rectification function defined in the reverse model set training module, substitute RRp as d, and apply the weight factor RF1 to the N sub-models. i -RF N i Rectification is performed to obtain a new weighting factor RF1 i+1 -RF N i+1 And based on the new non-zero weighting factor from RP1-RP N Select new effective prediction result vectors RV1-RV m , where (1≤m≤N); S5: Based on the new weighting factor RF1 i+1 -RF N i+1 For RP1-RP N A new prediction result RP is obtained by performing a weighted average calculation. f ; S6: Utilizing RV1-RV m RP f Calculate RV1-RV m For RP f The variance of the standard variance is calculated, and the index of the smallest standard variance is taken as K, where K takes the value [1, m]. S7: Get RP k This serves as the final result of this positive prediction; The model prediction module also includes a comprehensive prediction and evaluation module, and the implementation steps of the comprehensive prediction and evaluation module include: S1: Calculate P k With RPk The difference is ∆P; S2: Based on the following formula, substitute ∆P into x to perform binary logic result judgment, d≥0.75, .

[0016] In the description herein, it should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0017] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A multi-weight self-adjusting comprehensive evaluation and prediction system, comprising a system module assembly, characterized in that, The system module assembly includes a model training module and a model prediction module. The model training module includes a forward model set training module and a backward model set training module, and the model prediction module includes a forward prediction module and a backward prediction module. During operation, the model training module first collects forward and backward image sample sets from the binary logic system and categorizes them into S1 and S0. Then, it obtains a common sample set different from the binary logic system from network channels and labels it as S2. The forward model set training module includes the following steps during training: S1: For sample sets S1 and S2, a prediction model is built based on a neural network algorithm. This model learns from commonly used training parameters by adjusting sample data, training methods, and training weights, generating N positive prediction sub-models M1-M2. N N≥3; In S1, adjusting the sample data refers to changing the training samples, including but not limited to replacing samples or changing the number; adjusting the training method refers to changing the training method, including but not limited to changing the number of layers such as the input layer, hidden layer, and output layer in the algorithm; adjusting the training weights refers to changing the weights of each layer in the algorithm, including but not limited to the indicators of the regression function and the indicators of the fitted curve. S2: Assign initial weight factors F to the N positive prediction sub-models. i 1-F i N All of these values ​​are initialized to 1, i=0; S3: Determine the weight rectification function for the positive prediction sub-model: ， x refers to the judgment result of the current prediction sub-model, and d is the set judgment threshold. The significance of this rectification function is to directly discard judgment results that are lower than the set threshold and only retain judgment results that are equal to or higher than the set threshold. The reverse model set training module includes the following steps during training: S1: For sample sets S0 and S2, a prediction model is built based on a neural network algorithm. This model learns by adjusting commonly used training parameters such as sample data, training methods, and training weights, generating N inverse prediction sub-models RM1-RM2. N N≥3; S2: Assign initial weighting factors RF to the N backward prediction sub-models. i 1-RF i N All of these values ​​are initialized to 1, i=0; S3: Determine the weight rectification function for the inverse prediction sub-model: ， x refers to the judgment result of the current prediction sub-model, and d is the set judgment threshold.

2. The multi-weight self-adjusting comprehensive evaluation and prediction system according to claim 1, characterized in that, The model prediction module can be executed repeatedly.

3. The multi-weight self-adjusting comprehensive evaluation and prediction system according to claim 2, characterized in that, The positive prediction module includes the following steps during prediction: S1: Based on N positive prediction sub-models M1-M N Make predictions and normalize the prediction result vector to ensure that its prediction probability distribution is in the interval [0,1]. S2: Calculate the maximum value P1-P among the prediction result vectors of N positive prediction sub-models. N ; S3: According to P1-P N Calculate the predicted average value Rp; S4: Based on the weight rectification function defined in the positive model set training module, substitute Rp as d, and apply the weight factor F1 to the N sub-models. i -F N i Rectification is performed to obtain a new weighting factor F1 i+1 -F N i+1 And based on the new non-zero weighting factors from P1-P N Select new effective prediction result vectors V1-V m , where (1≤m≤N); S5: Based on the new weighting factor F1 i+1 -F N i+1 For P1-P N A new prediction result P is obtained by performing a weighted average calculation. f ; S6: Utilizing V1-V m P f Calculate V1-V m For P f The variance of the standard variance is calculated, and the index of the smallest standard variance is taken as K, where K takes the value [1, m]. S7: Take P k This is the final result of this positive prediction.

4. The multi-weight self-adjusting comprehensive evaluation and prediction system according to claim 3, characterized in that, The reverse prediction module includes the following steps during prediction: S1: Based on N positive prediction sub-models RM1-RM N Make predictions and normalize the prediction result vector to ensure that its prediction probability distribution is in the interval [0,1]. S2: Calculate the maximum value RP1-RP among the prediction result vectors of N positive prediction sub-models. N ; S3: Based on RP1-RP N Calculate the predicted average value RRp; S4: Based on the weight rectification function defined in the reverse model set training module, substitute RRp as d, and apply the weight factor RF1 to the N sub-models. i -RF N i Rectification is performed to obtain a new weighting factor RF1 i+1 -RF N i+1 And based on the new non-zero weighting factor from RP1-RP N Select new effective prediction result vectors RV1-RV m , where (1≤m≤N); S5: Based on the new weighting factor RF1 i+1 -RF N i+1 For RP1-RP N A new prediction result RP is obtained by performing a weighted average calculation. f ; S6: Utilizing RV1-RV m RP f Calculate RV1-RV m For RP f The variance of the standard variance is calculated, and the index of the smallest standard variance is taken as K, where K takes the value [1, m]. S7: Get RP k This is the final result of this positive prediction.

5. The multi-weight self-adjusting comprehensive evaluation and prediction system according to claim 1, characterized in that, The model prediction module also includes a comprehensive prediction and evaluation module, and the implementation steps of the comprehensive prediction and evaluation module include: S1: Calculate P k With RP k The difference is ∆P; S2: Based on the following formula, substitute ∆P into x and perform binary logic result judgment, d≥0.75, .

Images