Determination of the optimal batch size in incremental approaches: an application to tornado detection

Computing time and memory space limitations in applying support vector machines (SVMs) for large scale problems are recognized as critical limiting factors. Incremental approaches have serve as a remedy for large scale problems. However, determination of the appropriate batch size for incremental approaches has been explored rarely. In this study, the optimal batch size is defined as tradeoff between computing time and generalization error rate. Experiments for the determination of the optimal batch size, based on the mixture ratio of tornado and non-tornado data and a comparison between fixed batch size and knowledge based batch size, are performed. Preliminary results suggest that the knowledge based batch learning has the lowest generalization error rate.