MMSE-Based Optimized Transfer Strategy for Transfer Prediction of Parking Data

Abstract

To predict the spaces available of parking lots in a short term is a key technique in development of parking guidance systems. Time-series prediction is usually used to resolve this kind of problems, and work well after being trained with sufficient data. But as for city-wide parking guidance system (CPGS), we cannot acquire the data from every parking lots at the beginning, which lead to failure to train the specialized prediction model for them all individually. So, we introduce transfer learning to achieve the expected effect. There are so many papers to discuss how to use transfer learning in many specified tasks, as far as we know. So just applying it to another case is not innovative very much and we don’t want to be satisfied with that as well. In fact, how to transfer more effectively is novel topic and challenge to researchers. In this paper, we try to explain the mathematic meanings behind transfer learning for the first time. Then we build a LSTM-based network for transfer prediction, meanwhile propose an optimized transfer strategy, based on MMSE, to help build the target neural network more effectively and accurately. The following experiments show the strategy is very effective, although simple, it finds the most optimized transfer solution actively and directly. It is the first time we can build and evaluate the transfer learning solution at the design stage, rather than waiting passively till the moment when end results come out.