An Improved Neural Q-learning Approach for Dynamic Path Planning of Mobile Robots