Online Learning And Its Implementation In Intelligent Traffic And Finantial Market

As one of the most hottest topic in the field of machine learning for recent decades, online learning problem has drawn more and more attention from researchers, mainly due to its high cosistency with the real world applications. It has its unique intrinsic advantage to deal with the explosive upcoming data combined with the redundant outdated data. Nowadays, plenty of work has been focused on its theoretical aspect, involving solving a problem of online convex optimization, and prooving the worst-case upper bound of an algorithm, etc. In our thesis, we give it a shot on several practical subjects by modifying existing algorithm to suit them well, and evaluated the performance of these improved algorithms by sufficient experiments. The subjects are:1) Kernal methods in online algorithms and how to avoid the infinite growth of the kernal matrix; 2) The performance and advantages of online regression algorithms, espically that of online Boosting regression algorithms. And their implementation on the problem of precise train stopping; 3) How to device an online portfolio stratety from an online algorithm, and how to improve the reward by utilizing the fluctuation of market prices. Our major work includes:(?) We surveyed several popular kernal based Perceptron algorithms. And proposed an im-proved version of Projectron, which is highly performed among these algorithms, in-spiring by the concept known as soft margin from Support Vector Machine. SMProj, the improved version, has a higher accuracy rate and smaller support set. It would be meaningful for those looking for an feasible algorithm for practical use.(?) Simplizing the model of train stopping, we introdeced and evaluated several regression algorithms on the model. We also proposed a method to transform GB Boosting to an online version. According to our experiment, online boosting regression algorithm has its feability and advantage in the problem of train stopping.(?) We studied deeply on Anticor algorithm, trying to figure out its motivation and proof its coi-ectness. Inspiring by the way to predict the trends of transfering captical between stocks in a market introduced in Anticor algorithm, we proposed our CorrReg algorithm, uti-lizing its trend prediction, meanwhile quatilizing the among to transfer with the notation of internal regret in online learning. Our algorithm is more reasonable and flexible.