基于进化算法的自动调参方法

Whether in the field of machine learning or non machine learning, the selection of model parameters is the key factor of the model result. Optimization of super parameter for discontinuous or differentiable functions is usually based on three methods: grid search, random search and Bayesian optimization. Among them, random search and Bayesian optimization are the most widely used.However, in some situation(especially when the model is receiving data online), we need to quickly adjust the parameters in a short time so that the model can continue to work. However, random search sometimes can not find the qualified value in a short time, and the Bayesian search can not iterate fully, which leads to the poor effect of the parameters obtained in a short time. In order to get the right parameters as soon as possible in a short time, this paper uses multi-threaded Bayesian to adjust parameters in parallel. Then, based on evolutionary algorithm, we use the idea of "survival of the fittest", set checkpointsin these iteration threads. When the checkpoints are reached and the optimization process with poor optimization effect gives up its own parameters, copies the process parameters with better result, and adds noise to continue the iterationto concentrate thread resourceson better parameters.The research work of this paper obtains better parameters in the same time, which has a certain reference value for the direction of automatic parameter optimization.