Minimum Score Not Recorded

[pacowong]

Steps to reproduce

This is an issue discovered in my pull request (#902). To isolate the issue, I have created a smaller test case.

import nevergrad as ng
min_score = None

def fake_training(score: int) -> float:
    global min_score
    if min_score is None or min_score > score:
        min_score = score
    return score

def test(choice_size=1000, budget=300):
    global min_score
    min_score = None
    parametrization = ng.p.Instrumentation(score=ng.p.Choice(list(range(0, choice_size))))
    optimizer = ng.optimizers.OnePlusOne(parametrization=parametrization, budget=budget, num_workers=1)
    recommendation = optimizer.minimize(fake_training)
    print((min_score, recommendation.kwargs['score']))
    return min_score==recommendation.kwargs['score']

To run the test:

for i in range(0, 100):
    if not test(1000, 300):
         break

Observed Results

(0, 0)
(1, 1)
(0, 0)
(2, 2)
(10, 10)
(9, 21)

Expected Results

Print 100 tuples and the elements in each tuple should be the same.

Investigation

The _update_archive_and_bests function in nevergrad/optimization/base.py should be responsible for incorrect behaviour.
I have checked that the minimum loss has been passed into this function but the archiving process may have some problems:

1. the value of mvalue can be None in if mvalue is self.current_bests[name]
2. Assertion file if assert mvalue.parameter.loss == candidate.loss is added before if mvalue.parameter.loss > candidate.loss
   Not sure why this is the case because fake_training function is a deterministic function.

[jrapin]

Hi @pacowong
Thanks for pointing this out. This was indeed an expected behavior which aim was to deal with noisy data. In this case however it was confusing and definitely suboptimal so we have updated the optimizers code. It should be way better indeed.

[pacowong]

Thanks.