Dealing with users of different skills, and of variable capacity for learning and adapting over time, is a key issue in human-machine interaction, particularly in highly interactive applications such as computer games. Indeed, a recognized major concern for the game developers' community is to provide mechanisms to dynamically balance the difficulty level of the games in order to keep the user interested in playing. This work presents an innovative use of reinforcement learning techniques to build intelligent agents that adapt their behavior in order to provide dynamic game balancing. The idea is to couple learning with an action selection mechanism which depends on the evaluation of the current user's skills. To validate our approach, we applied it to a real-time fighting game, obtaining good results, as the adaptive agent is able to quickly play at the same level as opponents with different skills.