Create an experiment to show the effects of learning p_non_comp to precision and recall

[meneguzzi]

Answering @scranefield question on how to test stuff, we need to decide:

Felipe which is the best experiment to run to evaluate the effects of learning p_non_comp on norm learning?

My hypothesis is that if we get p_non_comp wrong, it depends on the value of p_sanc, I can think of six cases:

1. p_sanc high, p_non_comp correct: ideal case, we should get high high-precision, high-recall
2. p_sanc low, p_non_comp correct: we may overestimate some norms, so recall should be OK
3. p_sanc high, p_non_comp low: we may get low recall (because we won't see much non-compliant behaviour)
4. p_sanc low, p_non_comp low: we may get low recall (because we won't see much non-compliant behaviour), and low precision because the ones we do get may not get sanctioned
5. p_sanc high, p_non_comp high: we should get low precision because we will mistakenly think a lot of the generated behaviour is non-compliant, even when we don't see sanctioning
6. p_sanc low, p_non_comp high: we may get really high low recall (because we will see non-compliant behaviour), but low precision because we will mistakenly think a lot of the correct behaviour is non-compliant.