Departing from existing learning algorithms on this learning problem that require the convexity property of the underlying system, we develop an upper confidence bound (UCB)-type learning framework ...