An earlier work (Herrera et al.: Earth Planets Space, 58, 973-979, 2006) introduced two new methods for seismic hazard evaluation in a geographic area with distinct, but related, seismogenic regions. These two methods are based on modeling the transition probabilities of states, i.e. patterns of presence or absence of large earthquakes, in the regions, as a Markov chain. This modeling is, in turn, based on a straightforward counting of observed transitions between states. The direct method obtains transition probabilities among states that include events with magnitudes M ≥ M r , where M r is a specified threshold magnitude. Both methods gave very good results when applied to the Japan area, with the mixed method giving the best results and an improved magnitude range. In the work presented here, we propose other methods that use the learning capacity of an elementary neuronal network (perceptron) to characterize the Markovian behavior of the system; these neuronal methods, direct and mixed, gave results ∼7 and ∼6% better than the counting methods, respectively. Method performance is measured using grading functions that evaluate a tradeoff between positive and negative aspects of performance. This procedure results in a normalized grade being assigned that allows comparisons among different models and methods.