A sample-based approach to lookup table construction by SVM active learning

In intelligent robot systems, lookup table is often used to avoid computationally expensive calculations. To minimize the computational cost for constructing a lookup table, the table should be learned from a minimum number of informative training data (examples). In this paper, we address the problem of constructing lookup tables, from a point of view of binary classification problem. If the lookup table can be viewed as a binary classifier, there exists an optimal active learning algorithm, called support vector machine (SVM) active learning, that can select most informative examples in an optimal manner. To utilize the SVM active learning techniques, we interpret typical general lookup tables as binary classifiers. The main point of our approach is to utilize the spatial continuity common in lookup tables. Then, we propose sample-based techniques for efficiently constructing lookup tables through SVM active learning