Near Optimal On-Policy Control