In this paper, we apply a probably approximately correct (PAC) learning algorithm for multiplicity automata which can generate a quantitative model of target system behaviors with a statistical guarantee. By using the generated multiplicity automata model, we apply two analysis algorithms to estimate the minimum, maximum and average values of system behaviors. Also, we demonstrate how to apply the learning algorithm when the alphabet symbol size is not fixed. The result of the experiment is encouraging; Our approach made the estimation which is as precise as the exact reference answer obtains by a brute force enumeration.