Vision-based reinforcement learning: The role of transfer learning and action space design in symmetric reward financial markets