A generalized model for data science

Neural network is a powerful tool, which is often regarded as a black box. However, different task requires different parameters to be set up or couldn't work on, thus variable parameters might be a good solution for different tasks. We present a two-stage model based on Deep reinforcement learning as well as the pre-train method, this model could configure different parameters according to different data, improving and optimizing those parameters furthermore according to the returned loss value in each iteration. We apply this model to Boston housing pricing dataset, and it got a good result in restricted condition which was consistent with our expectations.