Improving control-knowledge acquisition for planning by active learning

Automatically acquiring control-knowledge for planning, as it is the case for Machine Learning in general, strongly depends on the training examples. In the case of planning, examples are usually extracted from the search tree generated when solving problems. Therefore, examples depend on the problems used for training. Traditionally, these problems are randomly generated by selecting some difficulty parameters. In this paper, we discuss several active learning schemes that improve the relationship between the number of problems generated and planning results in another test set of problems. Results show that these schemes are quite useful for increasing the number of solved problems.