def test_prioritized_buffer_flood(
capacity, wait_priority_after_sampling, initial_priority, uniform_ratio
):
buf = prioritized.PrioritizedBuffer(
capacity=capacity, wait_priority_after_sampling=wait_priority_after_sampling,
)
for _ in range(100):
for x in range(capacity + 1):
if wait_priority_after_sampling:
buf.append(x)
else:
buf.append(x, priority=initial_priority)
for _ in range(5):
n = random.randrange(1, capacity + 1)
buf.sample(n, uniform_ratio=uniform_ratio)
if wait_priority_after_sampling:
buf.set_last_priority([1.0] * n)
def test_prioritized_buffer_convergence(uniform_ratio):
expected_corr_range = {
0: (0.9, 1),
0.7: (0.5, 0.85),
1: (-0.3, 0.3)
}[uniform_ratio]
size = 100
buf = prioritized.PrioritizedBuffer(capacity=size)
for x in range(size):
buf.append(x)
priority_init = list([(i + 1) / size for i in range(size)])
random.shuffle(priority_init)
count_sampled = [0] * size
def priority(x, n):
if n == 0:
return 1.0
else:
return priority_init[x] / count_sampled[x]
for _ in range(200):
sampled, probabilities, _ = buf.sample(16, uniform_ratio=uniform_ratio)
priority_old = [priority(x, count_sampled[x]) for x in sampled]
if uniform_ratio == 0:
# assert: probabilities \propto priority_old
qs = [x / y for x, y in zip(probabilities, priority_old)]
for q in qs:
np.testing.assert_allclose(q, qs[0])
elif uniform_ratio == 1:
# assert: uniform
for p in probabilities:
np.testing.assert_allclose(p, probabilities[0])
for x in sampled:
count_sampled[x] += 1
priority_new = [priority(x, count_sampled[x]) for x in sampled]
buf.set_last_priority(priority_new)
for cnt in count_sampled:
assert cnt >= 1
corr = np.corrcoef(np.array([priority_init, count_sampled]))[0, 1]
corr_lb, corr_ub = expected_corr_range
assert corr > corr_lb
assert corr < corr_ub