Evaluating the relationship between regularity and learnability in recursive numeral systems using Reinforcement Learning