Novel Function Approximation Techniques For Large-Scale Reinforcement Learning