Value Function Approximation in Zero-Sum Markov Games