Tracking value function dynamics to improve reinforcement learning with piecewise linear function approximation