[gradys]

My intuition on this:

Maximum likelihood training -> faithfully represent training data

Reinforcement learning -> seek out the most preferred answer you can