How to discretize continuous state-action spaces in Q-learning: A symbolic control approach