Automatic Machine Learning Model Search Based On Intelligent Computing

Automatic machine learning refers to the complete design process of machine learning automatically completed by algorithms without any manual intervention.The core of the machine design is to implement automation design including model selection,feature selection and parameter tuning.Automated machine learning can bring researchers new research ideas on existing problems and lower the threshold for applying artificial intelligence technology in other industries.It has become one of the most advanced research directions of artificial intelligence.In this paper,we mainly study two sub-areas under the direction of architecture search in automated machine learning,namely machine learning model structure search and neural network architecture search.Machine learning model structure search is a way to study how to automatically design machine learning model integration.Each model is based on independent training of data stream input.By automatically trying to integrate the advantages of different models,an integrated model superior to single model is obtained.Neural network architecture search is to study how to automatically design a neural network architecture and train to find a suitable network for task problems by automatically generating new neural network architectures.In the machine learning model structure search,we define a graph-based model integration expression,and propose a method of splitting the graph into layer units.Then the graph structure is combined with the genetic programming algorithm to define the random operator,crossover operator and the mutation operator.We use the machine learning model as the node in the graph structure,the data is transmitted between the nodes in the graph structure,and the genetic algorithm is used to search the graph structure to find the integrated structure of the machine learning model.We apply the model search on fifteen classification tasks of PMLB and compare several other methods simultaneously.The experimental results show that the proposed method is better than traditional machine learning methods and other automated machine learning methods.At the same time,it shows the changes brought by the model structure when the accuracy of the optimal model is improved during the search process.It is found that the regression model has the effect of improving the accuracy in the integrated structure.In the neural network architecture search,we conduct research based on the Bayesian optimized neural network architecture search method,and improve the algorithm for optimizing the acquisition function.In this paper,Monte Carlo tree search is used as the search strategy to optimize the acquisition function.When searching,the width and depth of the tree are controlled.The relationship between exploration and utilization is optimized to optimize the search results,and the search time is optimized based on the algorithm principle.We experimented with neural network architecture search on two classified image datasets and compared it with simulated-annealing based method(Auto-Keras).The final experimental results show that our proposed search strategy is more efficient and accuracy rate is higher.At the same time,we show the neural network architecture that we searched with the comparison algorithm,and found that our model is less complex and less likely to overfit.