Gradient Reduction Convolutional Neural Network Policy for Financial Deep Reinforcement Learning