Learning DNF by Decision Trees

We investigate the problem of learning D N F concepts f rom examples using decision trees as a concept description language. Due to the replication problem, D N F concepts do not always have a concise decision tree descrip­ tion when the tests at the nodes are l im­ ited to the in i t ia l at t r ibutes. However, the representational complexity may be overcome by using high level at tr ibutes as tests. We present a novel a lgor i thm that modifies the in i ­ t ial bias determined by the pr imi t ive attr ibutes by adaptively enlarging the at t r ibute set wi th high level at tr ibutes. We show empirically that this a lgor i thm outperforms a standard decision tree algor i thm for learning small random D N F wi th and wi thout noise, when the examples are drawn f rom the uni form dist r ibut ion.