add minimum number of points parameter to AverageLearner

[basnijholt]

The AverageLearner samples a random

Like in #143 we should have an argument

    min_seeds_per_point : int, default 3
        Minimum amount of values per point. This means that the
        standard error of a point is infinity until there are
        'min_seeds_per_point' for a point.

Right now, it can happen that 2 consecutive sampling events happen to be the same up to some very small precision, resulting in a converged learner.

For example, in my case I sampled 2000 AverageLearners until convergence, most of them sampled ~3000 times except for this one:

The problem can be demonstrated using the following code.

Find a seed that produces two numbers that are close

for i in range(10_000_000):
    random.seed(i)
    r1 = random.random()
    r2 = random.random()
    if abs(r1 - r2) < 1e-7:
        break

random.seed(i)
print(i)

3317884

Then

random.seed(3317884)

def f(_):
    return random.random()

learner = adaptive.AverageLearner(f, rtol=0.01, atol=0.01)
adaptive.runner.simple(learner, goal=lambda l: l.loss() < 1)
print(learner.npoints, learner.data)

2 {0: 0.6232189044314053, 1: 0.6232188298033593}