Contributor(s)The Pennsylvania State University CiteSeerX Archives
Full recordShow full item record
AbstractIncorporating declarative bias or prior knowledge into learning is an active research topic in machine learning. Treestructured bias specifies the prior knowledge as a tree of &quot;relevance&quot; relationships between attributes. This paper presents a learning algorithm that implements treestructured bias, i.e., learns any target function probably approximately correctly from random examples and membership queries if it obeys a given tree-structured bias. The theoretical predictions of the paper are empirically validated. 1 Introduction Mitchell defined &quot;bias&quot; to be any prior knowledge that the learner has about the target function that enables it to generalize the examples beyond what it has already seen (Mitchell, 1980). Incorporating declarative bias into learning is an active research topic in machine learning, and has recently given rise to many useful algorithms, e.g., (Cohen, 1992), (Pazzani and Kibler, 1992), and (Shavlik and Towell, 1989). However, most of these algorithms are &quot;gene...