Learning from triggers

In this article we provide a refined analysis of learning in finite parameter spaces using the Triggering Learning Algorithm (TLA) of Gibson and Wexler (1994). We show that the behavior of the TLA can be modeled exactly as a Markov chain. This Markov model allows us to (1) describe formally the conditions for learnability in such spaces, (2) uncover problematic states in addition to the local maxima described by Gibson and Wexler, and (3) characterize convergence times for the learning algorithms quantitatively. In addition, we present arguments questioning the psychological plausibility of the TLA as a learning algorithm