A simulated annealing-based learning algorithm for Boolean DNF

We describe a stochastic algorithm learning Boolean functions from positive and negative examples. The Boolean functions are represented by disjunctive normal form formulas. Given a target DNF F depending on n variables and a set of uniformly distributed positive and negative examples, our algorithm computes a hypothesis H that rejects a given fraction of negative examples and has an epsilon -bounded error on positive examples. The stochastic algorithm utilises logarithmic cooling schedules for inhomogeneous Markov chains. The paper focuses on experimental results and comparisons with a previous approach where all negative examples have to be rejected [4]. The computational experiments provide evidence that a relatively high percentage of correct classifications on additionally presented examples can be achieved, even when misclassifications are allowed on negative examples. The detailed convergence analysis will be presented in a forthcoming paper [3].