def set_state(self, results):
N, self.pdf = LLRcalc.results_to_pdf(results)
if self.elo_model == "normalized":
mu, var = LLRcalc.stats(self.pdf) # code duplication with LLRcalc
if len(results) == 5:
self.sigma_pg = (2 * var)**0.5
elif len(results) == 3:
self.sigma_pg = var**0.5
else:
assert False
self.s0, self.s1 = [
self.elo_to_score(elo) for elo in (self.elo0, self.elo1)
]
mu_LLR, var_LLR = self.LLR_drift_variance(self.pdf, self.s0, self.s1,
None)
# llr estimate
self.llr = N * mu_LLR
self.T = N
# now normalize llr (if llr is not legal then the implications
# of this are unclear)
slope = self.llr / N
if self.llr > 1.03 * self.b or self.llr < 1.03 * self.a:
self.clamped = True
if self.llr < self.a:
self.T = self.a / slope
self.llr = self.a
elif self.llr > self.b:
self.T = self.b / slope
self.llr = self.b
def SPRT_elo(R, alpha=0.05, beta=0.05, p=0.05, elo0=None, elo1=None, elo_model=None):
"""
Calculate an elo estimate from an SPRT test."""
assert elo_model in ["BayesElo", "logistic"]
# Estimate drawelo out of sample
R3 = LLRcalc.regularize([R["losses"], R["draws"], R["wins"]])
drawelo = draw_elo_calc(R3)
# Convert the bounds to logistic elo if necessary
if elo_model == "BayesElo":
lelo0, lelo1 = [bayeselo_to_elo(elo_, drawelo) for elo_ in (elo0, elo1)]
else:
lelo0, lelo1 = elo0, elo1
# Make the elo estimation object
sp = sprt.sprt(alpha=alpha, beta=beta, elo0=lelo0, elo1=lelo1)
# Feed the results
if "pentanomial" in R.keys():
R_ = R["pentanomial"]
else:
R_ = R3
sp.set_state(R_)
# Get the elo estimates
a = sp.analytics(p)
# Override the LLR approximation with the exact one
a["LLR"] = LLRcalc.LLR_logistic(lelo0, lelo1, R_)
del a["clamped"]
# Now return the estimates
return a
def __init__(self, alpha=0.05, beta=0.05, elo0=0, elo1=5):
self.a = math.log(beta / (1 - alpha))
self.b = math.log((1 - beta) / alpha)
self.elo0 = elo0
self.elo1 = elo1
self.s0 = LLRcalc.L_(elo0)
self.s1 = LLRcalc.L_(elo1)
self.clamped = False
self.LLR_drift_variance = LLRcalc.LLR_drift_variance_alt2
def elo_to_score(self, elo):
"""
"elo" is expressed in our current elo_model."""
if self.elo_model == "normalized":
nt = elo / LLRcalc.nelo_divided_by_nt
return nt * self.sigma_pg + 0.5
else:
return LLRcalc.L_(elo)
def lelo_to_elo(self, lelo):
"""
For external use. "elo" is expressed in our current elo_model.
"lelo" is logistic."""
if self.elo_model == "logistic":
return lelo
score = LLRcalc.L_(lelo)
nt = (score - 0.5) / self.sigma_pg
return nt * LLRcalc.nelo_divided_by_nt
def outcome_prob(self, elo):
"""
The probability of a test with the given elo with worse outcome
(faster fail, slower pass or a pass changed into a fail)."""
s = LLRcalc.L_(elo)
mu_LLR, var_LLR = self.LLR_drift_variance(self.pdf, self.s0, self.s1,
s)
sigma_LLR = math.sqrt(var_LLR)
return Brownian(a=self.a, b=self.b, mu=mu_LLR,
sigma=sigma_LLR).outcome_cdf(T=self.T, y=self.llr)
def get_elo(results):
"""
"results" is an array of length 2*n+1 with aggregated frequences
for n games."""
results = LLRcalc.regularize(results)
games, mu, var = stats(results)
stdev = math.sqrt(var)
# 95% confidence interval for mu
mu_min = mu + Phi_inv(0.025) * stdev / math.sqrt(games)
mu_max = mu + Phi_inv(0.975) * stdev / math.sqrt(games)
el = elo(mu)
elo95 = (elo(mu_max) - elo(mu_min)) / 2.0
los = Phi((mu - 0.5) / (stdev / math.sqrt(games)))
return el, elo95, los
def set_state(self, results):
N, self.pdf = LLRcalc.results_to_pdf(results)
mu_LLR, var_LLR = self.LLR_drift_variance(self.pdf, self.s0, self.s1, None)
# llr estimate
self.llr = N * mu_LLR
self.T = N
# now normalize llr (if llr is not legal then the implications
# of this are unclear)
slope = self.llr / N
if self.llr > 1.03 * self.b or self.llr < 1.03 * self.a:
self.clamped = True
if self.llr < self.a:
self.T = self.a / slope
self.llr = self.a
elif self.llr > self.b:
self.T = self.b / slope
self.llr = self.b
def update_SPRT(R, sprt):
"""Sequential Probability Ratio Test
sprt is a dictionary with fixed fields
'elo0', 'alpha', 'elo1', 'beta', 'elo_model', 'lower_bound', 'upper_bound'.
It also has the following fields
'llr', 'state', 'overshoot'
which are updated by this function.
Normally this function should be called after each finished game (trinomial) or
game pair (pentanomial) but it is safe to call it multiple times with the same parameters.
Skipped updates are also handled sensibly.
The meaning of the inputs and the fields is as follows.
H0: elo = elo0
H1: elo = elo1
alpha = max typeI error (reached on elo = elo0)
beta = max typeII error for elo >= elo1 (reached on elo = elo1)
'overshoot' is a dictionary with data for dynamic overshoot
estimation. The theoretical basis for this is: Siegmund - Sequential
Analysis - Corollary 8.33. The correctness can be verified by
simulation
https://github.com/vdbergh/simul
R['wins'], R['losses'], R['draws'] contains the number of wins, losses and draws
R['pentanomial'] contains the pentanomial frequencies
elo_model can be either 'BayesElo' or 'logistic'
"""
# the next two lines are superfluous, but necessary for backward compatibility
sprt['lower_bound'] = math.log(sprt['beta'] / (1 - sprt['alpha']))
sprt['upper_bound'] = math.log((1 - sprt['beta']) / sprt['alpha'])
elo_model = sprt.get('elo_model', 'BayesElo')
assert (elo_model in ['BayesElo', 'logistic'])
elo0 = sprt['elo0']
elo1 = sprt['elo1']
# first deal with the legacy BayesElo/trinomial models
R3 = LLRcalc.regularize([R['losses'], R['draws'], R['wins']])
if elo_model == 'BayesElo':
# estimate drawelo out of sample
drawelo = draw_elo_calc(R3)
# conversion of bounds to logistic elo
lelo0, lelo1 = [bayeselo_to_elo(elo, drawelo) for elo in (elo0, elo1)]
else:
lelo0, lelo1 = elo0, elo1
# Log-Likelihood Ratio
R_ = R.get('pentanomial', R3)
sprt['llr'] = LLRcalc.LLR_logistic(lelo0, lelo1, R_)
# update the overshoot data
if 'overshoot' in sprt:
LLR_ = sprt['llr']
o = sprt['overshoot']
num_samples = sum(R_)
if num_samples < o['last_update']: # purge?
sprt['lost_samples'] = o['last_update'] - num_samples # audit
del sprt['overshoot'] # the contract is violated
else:
if num_samples == o['last_update']: # same data
pass
elif num_samples == o['last_update'] + 1: # the normal case
if LLR_ < o['ref0']:
delta = LLR_ - o['ref0']
o['m0'] += delta
o['sq0'] += delta**2
o['ref0'] = LLR_
if LLR_ > o['ref1']:
delta = LLR_ - o['ref1']
o['m1'] += delta
o['sq1'] += delta**2
o['ref1'] = LLR_
else:
# Be robust if some updates are lost: reset data collection.
# This should not be needed anymore, but just in case...
o['ref0'] = LLR_
o['ref1'] = LLR_
o['skipped_updates'] += (num_samples -
o['last_update']) - 1 # audit
o['last_update'] = num_samples
o0 = 0
o1 = 0
if 'overshoot' in sprt:
o = sprt['overshoot']
o0 = -o['sq0'] / o['m0'] / 2 if o['m0'] != 0 else 0
o1 = o['sq1'] / o['m1'] / 2 if o['m1'] != 0 else 0
# now check the stop condition
sprt['state'] = ''
if sprt['llr'] < sprt['lower_bound'] + o0:
sprt['state'] = 'rejected'
elif sprt['llr'] > sprt['upper_bound'] - o1:
sprt['state'] = 'accepted'
def update_SPRT(R, sprt):
"""Sequential Probability Ratio Test
sprt is a dictionary with fixed fields
'elo0', 'alpha', 'elo1', 'beta', 'elo_model', 'lower_bound', 'upper_bound', 'batch_size'
It also has the following fields
'llr', 'state', 'overshoot'
which are updated by this function.
Normally this function should be called each time 'batch_size' games (trinomial) or
game pairs (pentanomial) have been completed but it is safe to call it multiple times
with the same parameters. The main purpose of this is to be able to recalculate
the LLR for old tests.
In the unlikely event of a server crash it is possible that some updates may be missed
but this situation is also handled sensibly.
The meaning of the other inputs and the fields is as follows.
H0: elo = elo0
H1: elo = elo1
alpha = max typeI error (reached on elo = elo0)
beta = max typeII error for elo >= elo1 (reached on elo = elo1)
'overshoot' is a dictionary with data for dynamic overshoot
estimation. The theoretical basis for this is: Siegmund - Sequential
Analysis - Corollary 8.33. The correctness can be verified by
simulation
https://github.com/vdbergh/simul
R['wins'], R['losses'], R['draws'] contains the number of wins, losses and draws
R['pentanomial'] contains the pentanomial frequencies
elo_model can be either 'BayesElo', 'logistic' or 'normalized'"""
# the next two lines are superfluous, but unfortunately necessary for backward
# compatibility with old tests
sprt["lower_bound"] = math.log(sprt["beta"] / (1 - sprt["alpha"]))
sprt["upper_bound"] = math.log((1 - sprt["beta"]) / sprt["alpha"])
elo_model = sprt.get("elo_model", "BayesElo")
assert elo_model in ["BayesElo", "logistic", "normalized"]
elo0 = sprt["elo0"]
elo1 = sprt["elo1"]
# first deal with the legacy BayesElo/trinomial models
R3 = [R.get("losses", 0), R.get("draws", 0), R.get("wins", 0)]
if elo_model == "BayesElo":
# estimate drawelo out of sample
R3_ = LLRcalc.regularize(R3)
drawelo = draw_elo_calc(R3_)
# conversion of bounds to logistic elo
elo0, elo1 = [bayeselo_to_elo(elo, drawelo) for elo in (elo0, elo1)]
elo_model = "logistic"
R_ = R.get("pentanomial", R3)
batch_size = sprt.get("batch_size", 1)
# sanity check on batch_size
if sum(R_) % batch_size != 0:
sprt["illegal_update"] = sum(R_) # audit
if "overshoot" in sprt:
del sprt["overshoot"] # the contract is violated
# Log-Likelihood Ratio
assert elo_model in ["logistic", "normalized"]
if elo_model == "logistic":
sprt["llr"] = LLRcalc.LLR_logistic(elo0, elo1, R_)
else:
sprt["llr"] = LLRcalc.LLR_normalized(elo0, elo1, R_)
# update the overshoot data
if "overshoot" in sprt:
LLR_ = sprt["llr"]
o = sprt["overshoot"]
num_samples = sum(R_)
if num_samples < o["last_update"]: # purge?
sprt["lost_samples"] = o["last_update"] - num_samples # audit
del sprt["overshoot"] # the contract is violated
else:
if num_samples == o["last_update"]: # same data
pass
elif num_samples == o[
"last_update"] + batch_size: # the normal case
if LLR_ < o["ref0"]:
delta = LLR_ - o["ref0"]
o["m0"] += delta
o["sq0"] += delta**2
o["ref0"] = LLR_
if LLR_ > o["ref1"]:
delta = LLR_ - o["ref1"]
o["m1"] += delta
o["sq1"] += delta**2
o["ref1"] = LLR_
else:
# Be robust if some updates are lost: reset data collection.
# This should not be needed anymore, but just in case...
o["ref0"] = LLR_
o["ref1"] = LLR_
o["skipped_updates"] += (num_samples -
o["last_update"]) - 1 # audit
o["last_update"] = num_samples
o0 = 0
o1 = 0
if "overshoot" in sprt:
o = sprt["overshoot"]
o0 = -o["sq0"] / o["m0"] / 2 if o["m0"] != 0 else 0
o1 = o["sq1"] / o["m1"] / 2 if o["m1"] != 0 else 0
# now check the stop condition
sprt["state"] = ""
if sprt["llr"] < sprt["lower_bound"] + o0:
sprt["state"] = "rejected"
elif sprt["llr"] > sprt["upper_bound"] - o1:
sprt["state"] = "accepted"
