This paper considers opportunistic spectrum access for Secondary Users (SUs) from an adaptive learning perspective. A SU dynamically hops over multiple idle frequency-slots of a licensed frequency band, each with an adaptive activity factor. Aiming to determine the optimal activity factors of SUs, an algorithm is developed, in which each SU independently adjusts its activity factors by learning other SUs' behavior from locally available information. Due to the error-prone learning procedure, the proposed algorithm is interpreted as a stochastic gradient descent method. In order to establish stochastic stability for the proposed algorithm, the convergence with probability of 1 and also convergence rate are investigated with analysis and simulation.