Learning k-variable pattern languages efficiently stochastically finite on average from positive data

Abstract

The present paper presents a new approach of how to convert Gold-style [4] learning in the limit into stochastically finite learning with high confidence. We illustrate this approach on the concept class of all pattern languages. The transformation of learning in the limit into stochastically finite learning with high confidence is achieved by first analyzing the Lange-Wiehagen [7] algorithm with respect to its average-case time behavior until convergence. This algorithm learns the class of all pattern languages in the limit from positive data. The expectation of the total learning time is analyzed and exponentially small tail bounds are established for a large class of probability distributions. For patterns containing k different variables Lange and Wiehagen's algorithm possesses an expected total learning time of {ie13-01}, where α and Β are two easily computable parameters from the underlying probability distribution, and E[λ] is the expected example string length.

Finally, we show how to arrive at stochastically finite learning with high confidence.