Research On CNN Hyperparameter Quality Evaluation Method Based On Evidence Reasoning Rules

In recent years,Convolutional Neural Networks(CNN)have made great progress in various tasks in artificial intelligence fields,such as image recognition,semantic segmentation,and text classification.In the model training of CNN,the setting of hyperparameters depends on the long-term experience and subjective judgment of parameter adjusters.Moreover,the quality of hyperparameters determines the final effect of the model.Choosing poor quality hyperparameters means the failure of the model,which will greatly increase the time spent.Therefore,the hyperparameter optimization problem of convolutional neural networks is an urgent problem to be solved at present.In order to solve these problems,this paper deeply studies the CNN hyperparameter quality evaluation method of Evidential Reasoning Rule(ER rule).The main research contents include:By analyzing and researching the relationship between CNN hyperparameter quality and model quality,a CNN hyperparameter quality assessment method based on ER rule is proposed.First,the method regards hyperparameters as indicators to be evaluated,and discusses the selection method of indicators and the way of transforming confidence.Then,a dynamic update method of indicator weights is constructed according to the relationship between hyperparameters and models.The reliability of the indicators is constructed based on expert knowledge and historical data.Through the evidence inference rule algorithm,all the indicators are integrated and evaluated to obtain the prediction result.Finally,the effectiveness of the method is verified by the flower image recognition experiment.The experimental results show that the method can establish a mapping relationship with the effect of the model by effectively evaluating the quality of hyperparameters.This provides effective hyperparameter filtering for parameter adjusters.Through the research on the confidence of hyperparameter transformation of qualitative data type,it is found that this type of hyperparameter is difficult to express all the information contained in it with an accurate probability,and there is ambiguity.Therefore,a CNN hyperparameter quality evaluation method based on IER rule is proposed.The method first uses interval probability to express the uncertainty involved in transforming the confidence of the hyperparameters of qualitative data types.Then by reasonably constructing the weights and reliability of the indicators.By combining the ER rule with the interval measurement theory,the IER rule algorithm is used to fuse all the indicators to obtain the evaluation results.Finally,the experimental results show that the method has higher evaluation accuracy and better interpretability of qualitative hyperparameters.In order to further apply the CNN hyperparameter evaluation methods based on ER rule and IER rule to practical tasks,this paper designs and implements a prototype of a CNN hyperparameter quality evaluation system.In this system,the interface and evaluation entry for interacting with the user are provided through the human-computer interaction module.The confidence conversion method of ER rule and IER rule is provided through the data processing module.Finally,the evaluation methods of ER rule and IER rule are provided through the evaluation algorithm module.And passed the actual evaluation test,showing the effectiveness of the system.