Convergence of Expected Utility for Universal AI