Designing agents whose behavior challenges human players adequately is a key issue in computer games development. This work presents a novel technique, based on reinforcement learning (RL), to automatically control the game level, adapting it to the human player skills in order to guarantee a good game balance. RL has commonly been used in competitive environments, in which the agent must perform as well as possible to beat its opponent. The innovative use of RL proposed here makes use of a challenge function, which estimates the current player's level, as well as changes on the action selection mechanism of the RL framework. The technique is applied to a fighting game, Knock'em, to provide empirical validation of the approach.