Efficient Policy Evaluation with Safety Constraint for Reinforcement Learning