حداقل مربعات کاهش یافته مستقل پیوسته رگرسیون بردار پشتیبانی

In this paper, an online algorithm, viz. online independent reduced least squares support vector regression (OIRLSSVR), is proposed based on the linear independence and the reduced technique. As opposed to some offline algorithms, OIRLSSVR takes the realtime advantage, which is confirmed using benchmark data sets. In comparison with online algorithm, the realtime of OIRLSSVR is also favorable. As for this point, it is tested with experiments on the benchmark data sets and a more realistic scenario namely a diesel engine example. All in all, OIRLSSVR can enhance the modeling realtime, especially for the case where the samples enter in a flow mode.

algorithms does not usually satisfy our need. To enhance the modeling realtime, the online learning algorithms arise naturally. In this paper, this hot issue is discussed and developed. As we know, reduced least squares support vector regression can cut down the number of support vectors. However, it selects the support vectors randomly, maybe causing some subtle samples to be chosen as support vectors, which is unreasonable and unwise. Meantime, it alludes to the longer prediction time and worse realtime. Hence, the recursive reduced least squares support vector regression is developed. During each update, the sample making the most contribution to the modeling target is selected as support vector, thus obviously curtailing the number of support vectors and boosting the realtime. Unfortunately, RR-LSSVR is an offline algorithm. When the samples enter in a flow mode, it seems helpless. In this situation, an online learning algorithm is developed, viz. OIRLSSVR in which the linear independence is utilized to select the basis vectors as support vectors. For the state-of-the-art Gaussian kernel most used in kernel learning community, it is proofed theoretically that the number of support vectors does not grow infinitely, which guarantees the realtime of OIRLSSVR. Although OIRLSSVR needs more support vectors than RR-LSSVR, the incrementally updated structure helps OIRLSSVR obtain less rebuilding time. In consequence, the realtime of OIRLSSVR is superior to that of RR-LSSVR. To further show the efficacy and feasibility of the proposed OIRLSSVR, a lot of experiments are implemented. In contrast to the offline algorithms namely RLSSVR and RR-LSSVR, OIRLSSVR has the realtime advantage over both. In comparison with the online IncrLSSVR, benchmarks are also favorable for the presented OIRLSSVR. From the experiment on the more realistic scenario, the testing results are also consistent with those on the benchmarks. In a nutshell, when the samples enter in a flow mode, the online OIRLSSVR are favorable. Although OIRLSSVR is proposed and developed in the regression situation, it may be extended to classification problems.