Applying DDQN for Advanced Model Tuning: an Evaluation of Optimization Methods and Results

In deep learning, due to the large number of parameters, tuning the model is a crucial step in the training process. A well-tuned model typically exhibits better performance and higher accuracy compared to an untuned model. Currently, various optimization methods have emerged for model tuning, such as grid search, random search, and Bayesian optimization. Each of these methods has its own advantages in parameter tuning; however, they are limited to optimizing only the parameters and cannot address other influential factors in model training, such as the selection of the loss function. Typically, the loss function is fixed during model training and is chosen by the developer based on experience, making the training performance highly dependent on expertise. To address this limitation, this paper proposed a reinforcement learning-based model optimization method. This approach not only optimizes model parameters but also considers other factors that affect model accuracy. In the experiments, the proposed method was validated on an object recognition network by tuning parameters that influence accuracy, such as the learning rate, number of training epochs, and loss function. The experimental results show that the accuracy of the model tuned by this method improved by 1.3% compared to the best accuracy achieved by traditional tuning methods.