Datasets:
Modalities:
Text
Formats:
json
Languages:
English
Size:
< 1K
Tags:
reinforcement-learning
policy-optimization
klpo
exact-moment-replay
score-centering
stratified-sampling
License:
File size: 3,488 Bytes
f95cbd4 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 | import math
import numpy as np
import pytest
from celestis_rl.sequential_audit import MetricSpec,SequentialParetoAudit,confidence_radius
from celestis_rl.certificates import ConfidenceLedger
def test_radius_formula_and_range():
n=1000;v=.01;width=2.;alpha=.01
r,b=confidence_radius(n,v,width,alpha,'bernstein')
assert r==pytest.approx(math.sqrt(2*v*math.log(2/alpha)/n)+7*width*math.log(2/alpha)/(3*(n-1)))
r,b=confidence_radius(n,v,width,alpha,'hybrid')
h,_=confidence_radius(n,v,width,alpha/2,'hoeffding')
e,_=confidence_radius(n,v,width,alpha/2,'bernstein')
assert r==min(h,e)
def test_low_variance_faster_acceptance():
looks=(64,128,256,512,1024,2048,4096,8192,16384)
audits=[SequentialParetoAudit([MetricSpec('reward')],ConfidenceLedger(),candidate_id='v1',looks=looks,method=m) for m in ['hybrid','hoeffding']]
for audit in audits:audit.add(np.full(16384,.55),np.full(16384,.5),candidate_id='v1')
assert audits[0].result.accepted and audits[1].result.accepted
assert audits[0].n<audits[1].n
def test_multi_metric_rejects_hidden_cost_regression():
specs=[MetricSpec('reward'),MetricSpec('cost',direction=-1)]
audit=SequentialParetoAudit(specs,ConfidenceLedger(),candidate_id='bad',looks=(64,2048))
result=audit.add(np.tile([.9,.7],(2048,1)),np.tile([.5,.5],(2048,1)),candidate_id='bad')
assert not result.accepted and result.lower_gain_bounds['cost']<0
assert audit.finished
with pytest.raises(RuntimeError):audit.add([[.9,.7]],[[.5,.5]],candidate_id='bad')
def test_batch_partition_and_resume(tmp_path):
kwargs=dict(candidate_id='frozen',looks=(64,128,256,512))
a=SequentialParetoAudit([MetricSpec('value',min_gain=.9)],ConfidenceLedger(),**kwargs)
b=SequentialParetoAudit([MetricSpec('value',min_gain=.9)],ConfidenceLedger(),**kwargs)
rng=np.random.default_rng(22);c=rng.uniform(.2,.8,512);d=rng.uniform(.2,.8,512)
a.add(c,d,candidate_id='frozen')
b.add(c[:90],d[:90],candidate_id='frozen');b.save(tmp_path/'audit.json')
b=SequentialParetoAudit.load(tmp_path/'audit.json');b.add(c[90:],d[90:],candidate_id='frozen')
assert a.n==b.n and a.look_index==b.look_index
assert np.allclose(a.mean,b.mean,atol=1e-15)
assert np.allclose(a.m2,b.m2,atol=1e-15)
def test_irreversible_ledger_and_proposal_guard(tmp_path):
ledger=ConfidenceLedger(delta=.01)
a=SequentialParetoAudit([MetricSpec('r')],ledger,candidate_id='a',ledger_path=tmp_path/'ledger.json')
assert ConfidenceLedger.load(tmp_path/'ledger.json').attempts==1
b=SequentialParetoAudit([MetricSpec('r')],ledger,candidate_id='b')
assert b.attempt_delta==a.attempt_delta/4
with pytest.raises(ValueError):a.add([.5],[.2],candidate_id='b')
with pytest.raises(ValueError):a.add([1.1],[.2],candidate_id='a')
assert a.n==0
@pytest.mark.parametrize('looks',[(1,),(),(32,16),(True,),(32,32)])
def test_invalid_looks(looks):
ledger=ConfidenceLedger()
with pytest.raises(ValueError):SequentialParetoAudit([MetricSpec('r')],ledger,candidate_id='a',looks=looks)
assert ledger.attempts==0
@pytest.mark.parametrize('seed',range(5))
def test_no_acceptance_deterministic_negative_margin(seed):
rng=np.random.default_rng(seed)
base=rng.uniform(.3,.7,1024);candidate=base-.02
audit=SequentialParetoAudit([MetricSpec('r')],ConfidenceLedger(),candidate_id='negative',looks=(32,128,1024))
res=audit.add(candidate,base,candidate_id='negative')
assert not res.accepted
|