On a Stochastic Approximation Procedure Applied to the Bandit Problem