RoboRACE-Bench / tests /test_validation.py
Nabidnur's picture
RoboRACE: code, cached agent outputs, results, media, SO-101 renders (part 5)
99e41a8 verified
Raw History Blame Contribute Delete
2.16 kB
"""Validation against references and theory (not just behaviour tests)."""
import itertools
import numpy as np
import pytest
from robotics.bench import trackA, trackB
from robotics.swarm import fusion
def test_our_dawid_skene_matches_crowdkit_reference():
pytest.importorskip("crowdkit")
from robotics.swarm.crowd import crowdkit
(Ph, yh, _), (P, y, _) = trackA.synthetic(n=400, seed=3)
ours = fusion.dawid_skene(np.vstack([Ph, P]), P)[0] > 0 # same data as the reference (hist + eval)
ref = crowdkit(Ph, P, "DawidSkene")[0] > 0
assert np.mean(ours == ref) >= 0.90
assert abs(np.mean(ours == y) - np.mean(ref == y)) < 0.03
@pytest.mark.parametrize("attack", ["bias", "collude", "noise", "two_faced", "drift"])
@pytest.mark.parametrize("F", [1, 2, 3])
def test_wmsr_safety_on_complete_graph(attack, F):
"""W-MSR theorem (LeBlanc et al. 2013): with F-local adversaries on a sufficiently robust graph
(here complete, n=12 >= 3F+1), every honest state stays inside the convex hull (per coordinate)
of the honest initial states, whatever the Byzantine robots send."""
w = trackB.World(n=12, k=11, F=F, attack=attack, seed=F)
w.A = ~np.eye(12, dtype=bool)
target = np.array([5.0, 5.0])
m = w.measure(target)
honest = ~w.byz
lo, hi = m[honest].min(0), m[honest].max(0)
x = m.copy()
for rnd in range(15):
new = x.copy()
for i in np.flatnonzero(honest):
nb = np.flatnonzero(w.A[i])
vals = np.array([w.lie(j, i, x[j], target, "eval", rnd) if w.byz[j] else x[j] for j in nb])
new[i] = trackB._wmsr(x[i], vals, F)
x = new
assert np.all(x[honest] >= lo - 1e-9) and np.all(x[honest] <= hi + 1e-9)
def test_race_s_is_invariant_to_monotone_miscalibration():
"""Rank-Gauss makes RACE-S invariant to any strictly increasing distortion of an agent's scores."""
(Ph, yh, _), (P, y, _) = trackA.synthetic(n=300, seed=4)
a = fusion.race_s(Ph, P, 0)[0] > 0
Ph2, P2 = Ph.copy(), P.copy()
Ph2[:, 2], P2[:, 2] = Ph[:, 2] ** 3, P[:, 2] ** 3
b = fusion.race_s(Ph2, P2, 0)[0] > 0
assert np.mean(a == b) >= 0.98