The model outputs probabilities, which you have to sample randomly. Choosing the "highest" probability every time leads to poor results in practice, such as th

The model outputs probabilities, which you have to sample randomly. Choosing the "highest" probability every time leads to poor results in practice, such as the model tending to repeat itself. It's a sort of Monte-Carlo approach.