elo_sampling.py

import random

# Works, is pretty good
all_pairs: list[tuple[int, int]] = []
for i in range(100):
    for j in range(i+1, 100):
        i2, j2 = random.sample((i, j), 2)
        all_pairs.append((i2, j2))
random.shuffle(all_pairs)
selected_pairs = all_pairs[:600]

# But this is more sample efficient
all_pairs = []
indexes = list(range(100))
for K in range(4):
    random.shuffle(indexes)
    for i in range(100):
        all_pairs.append((indexes[i], indexes[(i+1) % 100]))
len(all_pairs) # 400

# ~400 samples with random cycles, resolves the ELOs as accurately as ~600 randomly sampled pairs. This makes synthetic training data 50% cheaper, which saves a lot when spending >$10k
# Image Showing Idea: https://private-user-images.githubusercontent.com/8939474/470457008-4731f849-e81f-40bd-a211-fd6f273f1f84.png?jwt=eyJhbGciOiJIUzI1NiIsInR5cCI6IkpXVCJ9.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.2xsCZ1cou_nRQRVZnBxxWzsUG21HrA6c6WMGYWPF1kU