Research And Implementation Of Algorithm And Hyperparameter Recommendation Platform Based On Reinforcement Learning

In the context of increasing problem scale,the application threshold of machine learn-ing and deep learning has become higher and higher,and it requires highly experienced manual intervention.However,manual intervention often requires a lot of time and com-putational cost.To better solve the above limitation,this thesis conducts in-depth research on the hyperparameter optimization and model selection in the machine learning and deep learning,and implements an efficient algorithm based on reinforcement learning.For the hyperparameter optimization problem,this thesis uses reinforcement learning as the technical framework to realize a hyperparameter optimization method.This method selects each hyperparameter by sequence,which can reduce the search space of the opti-mization,reduce the difficulty of exploration,and improve the optimization efficiency.At the same time,in order to solve the problem of time-consuming evaluation in traditional hyperparameter optimization methods,this thesis directly evaluates the hyperparameter configuration by using a predictive model.Moreover,in order to prevent the prediction model from introducing excessive errors and leading to sub-optimal policy,this method calculates the distance between the policy before and after the prediction model is used to dynamically control the number of uses of the model.In order to further improve the time efficiency,this thesis uses the historical optimization experience,which includes the historical experience of other optimization tasks and the optimization experience of the current task.In addition,compared to traditional optimization methods,this method im-proves the accuracy of the model by optimizing hyperparameters,while also reducing the latency of the model in the actual environment.For the dual optimization problem of model and hyperparameters,based on the re-inforcement learning framework,this thesis implements a multi-branch structure of the agent,which includes an algorithm controller and multiple hyperparameter optimization controllers.Specifically,the algorithm is selected by the algorithm controller and the corresponding hyperparameter optimization branch is selected accordingly.At the same time,the prediction model and usage method in the hyperparameter optimization method are also used in the solution of this problem to further improve the time efficiency.Finally,this thesis compares the current advanced optimization methods on multiple datasets,and the experimental results verify that the proposed method is feasible and effec-tive in terms of time efficiency and optimization results.At the same time,this thesis also conducted reasonable and fair ablation experiments to illustrate the effectiveness of each component.Combining all the experimental results,it can be shown that the proposed method can outperform other optimization methods in multiple performance indicators.