Learning Approximate and Exact Numeral Systems via Reinforcement Learning