Adapting the Function Approximation Architecture in Online Reinforcement Learning