Gradient Ascent for Active Exploration in Bandit Problems