def rate_all():
games = Game.query.all()
users = sanitized_users(games)
g_vec = sanitized_games(games)
ratings = {u.id: u.last_rating() for u in users}
rating_prior = {u: v.rating if (v and v.rating) else 0 for u,v in ratings.items()}
neighbors = rm.neighbors(games)
neighbor_avgs = rm.compute_avgs(games, rating_prior)
t_min = min(games, key=lambda g: g.date_played or datetime.datetime.now()).date_played
t_max = max(games, key=lambda g: g.date_played or datetime.datetime.now()).date_played
iters = 100
lam = .37 # Weight for controlling 'pull' of neighborhood weighted average. Higher = stays in the neighborhood.
lrn = lambda i: ((1. + .1*iters)/(i + .1 * iters))**.6 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt) ** 2)
# Shuffle the vector of result-tuples
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[w]) * (rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (-1.0 * weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[b]) * (rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual) ** 2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k,v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max - r_min) * 40.0
neighbor_avgs = rm.compute_avgs(games, rating_prior)
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1-g[2]
wins[g[1]] = wins.get(g[1], 0) + 1-g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(Rating(user_id=k, rating=rating_prior[k]))
print("%d: %f (%d - %d)" % (k,rating_prior[k], wins.get(k,0), losses.get(k,0)) )
db.session.commit()
def test_compute_avgs(self):
ratings = {1:2, 2:4, 3:2, 4:3}
games = [Game(Player(1), Player(2), 'W+R', date(2000,1,1)),
Game(Player(1), Player(3), 'B+R', date(2000,1,1)),
Game(Player(1), Player(4), 'W+R', date(2000,1,1)),
Game(Player(3), Player(2), 'B+R', date(2000,1,1))]
averages = rm.compute_avgs(games, ratings)
self.assertEqual(averages, {1: 3.0, 2: 2.0, 3: 3.0, 4: 2.0})
def test_compute_avgs(self):
ratings = {1: 2, 2: 4, 3: 2, 4: 3}
games = [
Game(Player(1), Player(2), 'W+R', date(2000, 1, 1)),
Game(Player(1), Player(3), 'B+R', date(2000, 1, 1)),
Game(Player(1), Player(4), 'W+R', date(2000, 1, 1)),
Game(Player(3), Player(2), 'B+R', date(2000, 1, 1))
]
averages = rm.compute_avgs(games, ratings)
self.assertEqual(averages, {1: 3.0, 2: 2.0, 3: 3.0, 4: 2.0})
def test_compute_avgs(self):
ratings = {1:2, 2:4, 3:2, 4:3, 5:2}
games = [(1, 2, '1', date(2000,1,1).timestamp(), 0, 7),
(1, 3, '0', date(2000,1,2).timestamp(), 0, 7),
(1, 4, '1', date(2000,1,3).timestamp(), 0, 7),
(2, 5, '1', date(2000,1,3).timestamp(), 0, 7),
(3, 2, '0', date(2000,1,4).timestamp(), 0, 7)]
averages = rm.compute_avgs(games, ratings)
# 2 & 4 both only played people with the same rating,
# so they should have the same weighted average opponent rating
self.assertEqual(averages[2], averages[4])
# 1 faced lower average opposition than 3 did, so 3's average opponent rating should be higher.
self.assertLess(averages[1], averages[3])
# a clear strongest player, two equal players
ratings = {1:10, 2:3, 3:3}
games = [(1, 2, '1', date(2000,1,1).timestamp(), 0, 7),
(1, 3, '0', date(2000,1,3).timestamp(), 0, 7),
(2, 3, '1', date(2000,1,5).timestamp(), 0, 7)]
averages = rm.compute_avgs(games, ratings)
#the player who faced the strongest player most recently should have the higher average
self.assertLess(averages[2], averages[3])
# all the same strength, but 2 gave 3 a 3-handicap game.
ratings = {1:3, 2:3, 3:3}
games = [(1, 2, '1', date(2000,1,1).timestamp(), 0, 7),
(1, 3, '0', date(2000,1,1).timestamp(), 0, 7),
(2, 3, '1', date(2000,1,1).timestamp(), 3, 0)]
averages = rm.compute_avgs(games, ratings)
# The player who gave stones should have a higher average
self.assertGreater(averages[2], averages[3])
# the player who received them should have a lower one
self.assertGreater(averages[1], averages[3])
def test_compute_avgs(self):
ratings = {1: 2, 2: 4, 3: 2, 4: 3, 5: 2}
games = [(1, 2, '1', date(2000, 1, 1).timestamp(), 0, 7),
(1, 3, '0', date(2000, 1, 2).timestamp(), 0, 7),
(1, 4, '1', date(2000, 1, 3).timestamp(), 0, 7),
(2, 5, '1', date(2000, 1, 3).timestamp(), 0, 7),
(3, 2, '0', date(2000, 1, 4).timestamp(), 0, 7)]
averages = rm.compute_avgs(games, ratings)
# 2 & 4 both only played people with the same rating,
# so they should have the same weighted average opponent rating
self.assertEqual(averages[2], averages[4])
# 1 faced lower average opposition than 3 did, so 3's average opponent rating should be higher.
self.assertLess(averages[1], averages[3])
# a clear strongest player, two equal players
ratings = {1: 10, 2: 3, 3: 3}
games = [(1, 2, '1', date(2000, 1, 1).timestamp(), 0, 7),
(1, 3, '0', date(2000, 1, 3).timestamp(), 0, 7),
(2, 3, '1', date(2000, 1, 5).timestamp(), 0, 7)]
averages = rm.compute_avgs(games, ratings)
#the player who faced the strongest player most recently should have the higher average
self.assertLess(averages[2], averages[3])
# all the same strength, but 2 gave 3 a 3-handicap game.
ratings = {1: 3, 2: 3, 3: 3}
games = [(1, 2, '1', date(2000, 1, 1).timestamp(), 0, 7),
(1, 3, '0', date(2000, 1, 1).timestamp(), 0, 7),
(2, 3, '1', date(2000, 1, 1).timestamp(), 3, 0)]
averages = rm.compute_avgs(games, ratings)
# The player who gave stones should have a higher average
self.assertGreater(averages[2], averages[3])
# the player who received them should have a lower one
self.assertGreater(averages[1], averages[3])
def rate_all():
games = Game.query.all()
users = sanitized_users(games)
g_vec = sanitized_games(games)
ratings = {u.id: u.last_rating() for u in users}
rating_prior = {
u: v.rating if (v and v.rating) else 0
for u, v in ratings.items()
}
neighbors = rm.neighbors(games)
neighbor_avgs = rm.compute_avgs(games, rating_prior)
t_min = min(
games,
key=lambda g: g.date_played or datetime.datetime.now()).date_played
t_max = max(
games,
key=lambda g: g.date_played or datetime.datetime.now()).date_played
iters = 100
lam = .37 # Weight for controlling 'pull' of neighborhood weighted average. Higher = stays in the neighborhood.
lrn = lambda i: (
(1. + .1 * iters) /
(i + .1 * iters))**.6 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt)**2)
# Shuffle the vector of result-tuples
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (
weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[w]) *
(rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (
-1.0 * weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[b]) *
(rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual)**2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k, v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max -
r_min) * 40.0
neighbor_avgs = rm.compute_avgs(games, rating_prior)
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1 - g[2]
wins[g[1]] = wins.get(g[1], 0) + 1 - g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(Rating(user_id=k, rating=rating_prior[k]))
print("%d: %f (%d - %d)" %
(k, rating_prior[k], wins.get(k, 0), losses.get(k, 0)))
db.session.commit()
def rate_all(t_from=datetime.datetime.utcfromtimestamp(1.0),
t_to=datetime.datetime.now(),
iters=200, lam=.22):
"""
t_from -- datetime obj, rate all games after this
t_to -- datetime obj, rate all games up to this
iters -- number of iterations
lam -- 'neighborhood pull' parameter. higher = more error from moving a rank away from ranks in its neighborhood
"""
games = Game.query.filter(Game.date_played < t_to, Game.date_played > t_from)
g_vec = sanitized_games(games)
users = sanitized_users(g_vec)
print("found %d users with %d valid games" % (len(users), len(g_vec)) )
aga_ids_to_uids = dict([(int(u.aga_id), u.id) for u in users])
ratings = {int(u.aga_id): u.last_rating() for u in users}
rating_prior = {id: v.rating if (v and v.rating) else 20 for id,v in ratings.items()}
print ("%d users with no priors" % len(list(filter(lambda v: v == 20, rating_prior.values()))))
neighbors = rm.neighbors(g_vec)
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
t_min = min([g[3] for g in g_vec])
t_max = max([g[3] for g in g_vec])
lrn = lambda i: ((1. + .1*iters)/(i + .1 * iters))**.3 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt) ** 2)
# Shuffle the vector of result-tuples and step through them, accumulating error.
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[w]) * (rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (-1.0 * weight*(odds - actual)*odds*(1-odds) + (lam/len(neighbors[b]) * (rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual) ** 2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k,v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max - r_min) * 40.0
#update neighborhood averages?
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
if (i % 50 == 0):
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1-g[2]
wins[g[1]] = wins.get(g[1], 0) + 1-g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k])[-10:]:
print("%d (uid: %d): %f (%d - %d)" % (k, aga_ids_to_uids[k], rating_prior[k], wins.get(k,0), losses.get(k,0)) )
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(Rating(user_id=aga_ids_to_uids[k], rating=rating_prior[k], created=t_to))
db.session.commit()
def rate_all(t_from=datetime.datetime.utcfromtimestamp(1.0),
t_to=datetime.datetime.now(),
iters=200,
lam=.22):
"""
t_from -- datetime obj, rate all games after this
t_to -- datetime obj, rate all games up to this
iters -- number of iterations
lam -- 'neighborhood pull' parameter. higher = more error from moving a rank away from ranks in its neighborhood
"""
games = Game.query.filter(Game.date_played < t_to,
Game.date_played > t_from)
g_vec = sanitized_games(games)
users = sanitized_users(g_vec)
print("found %d users with %d valid games" % (len(users), len(g_vec)))
aga_ids_to_uids = dict([(int(u.aga_id), u.id) for u in users])
ratings = {int(u.aga_id): u.last_rating() for u in users}
rating_prior = {
id: v.rating if (v and v.rating) else 20
for id, v in ratings.items()
}
print("%d users with no priors" %
len(list(filter(lambda v: v == 20, rating_prior.values()))))
neighbors = rm.neighbors(g_vec)
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
t_min = min([g[3] for g in g_vec])
t_max = max([g[3] for g in g_vec])
lrn = lambda i: (
(1. + .1 * iters) /
(i + .1 * iters))**.3 #Control the learning rate over time.
for i in range(iters):
loss = 0
# Accumulate the neighborhood loss prior to changing the ratings around
for id, neighbor_wgt in neighbor_avgs.items():
loss += lam * ((rating_prior[id] - neighbor_wgt)**2)
# Shuffle the vector of result-tuples and step through them, accumulating error.
random.shuffle(g_vec)
for g in g_vec:
w, b, actual, t, handi, komi = g
odds = rm.expect(rating_prior[b], rating_prior[w], handi, komi)
weight = rm.time_weight(t, t_min, t_max)
rating_prior[w] -= lrn(i) * (
weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[w]) *
(rating_prior[w] - neighbor_avgs[w])))
rating_prior[b] -= lrn(i) * (
-1.0 * weight * (odds - actual) * odds * (1 - odds) +
(lam / len(neighbors[b]) *
(rating_prior[b] - neighbor_avgs[b])))
loss += weight * ((odds - actual)**2)
# Scale the ratings
r_min = min(rating_prior.values())
r_max = max(rating_prior.values())
if r_max != r_min:
for k, v in rating_prior.items():
rating_prior[k] = (rating_prior[k] - r_min) / (r_max -
r_min) * 40.0
#update neighborhood averages?
neighbor_avgs = rm.compute_avgs(g_vec, rating_prior)
if (i % 50 == 0):
print('%d : %.4f' % (i, loss))
# Update the ratings and show how we did.
wins, losses = {}, {}
for g in g_vec:
wins[g[0]] = wins.get(g[0], 0) + g[2]
losses[g[0]] = losses.get(g[0], 0) + 1 - g[2]
wins[g[1]] = wins.get(g[1], 0) + 1 - g[2]
losses[g[1]] = losses.get(g[1], 0) + g[2]
for k in sorted(rating_prior, key=lambda k: rating_prior[k])[-10:]:
print("%d (uid: %d): %f (%d - %d)" %
(k, aga_ids_to_uids[k], rating_prior[k], wins.get(
k, 0), losses.get(k, 0)))
for k in sorted(rating_prior, key=lambda k: rating_prior[k]):
db.session.add(
Rating(user_id=aga_ids_to_uids[k],
rating=rating_prior[k],
created=t_to))
db.session.commit()