def run_mini_monster(Bconstructor,
Bval,
learning_algs,
rewards,
risks,
val,
T,
slate_len,
i,
start,
params=None):
for (k, v) in learning_algs.items():
if params is None:
B = Bconstructor()
M = Semibandits.MiniMonster(B,
learning_alg=v,
classification=False)
(r, reg, val_tmp) = M.play(T,
verbose=True,
params={'weight': np.ones(slate_len)},
validate=Bval)
rewards['mini_%s' % (k)].append(r)
print("Num Uniform Actions: %d" % (M.num_unif), flush=True)
np.savetxt(outdir + "mini_%s_rewards_%d.out" % (k, i), r)
if Bval is not None:
val['mini_%s' % (k)].append(val_tmp)
print('Mini %s Final Validation %0.2f' %
(k, val['mini_%s' % (k)][i - start][-1]),
flush=True)
np.savetxt(outdir + "mini_%s_validation_%d.out" % (k, i),
val_tmp)
if params is not None:
for a in params:
B = Bconstructor()
M = Semibandits.MiniMonster(B,
learning_alg=v,
classification=False)
(r, reg, val_tmp) = M.play(T,
verbose=True,
params={
'weight': np.ones(slate_len),
'mu': a
},
validate=Bval)
rewards['mini_%s_%0.3f' % (k, a)].append(r)
print("Num Uniform Actions: %d" % (M.num_unif), flush=True)
np.savetxt(outdir + "mini_%s_%0.3f_rewards_%d.out" % (k, a, i),
r)
if Bval is not None:
val['mini_%s_%0.3f' % (k, a)].append(val_tmp)
print('Mini %s %0.3f Final Validation %0.2f' %
(k, a, val['mini_%s_%0.3f' % (k, a)][i - start][-1]),
flush=True)
np.savetxt(
outdir + "mini_%s_%0.3f_validation_%d.out" % (k, a, i),
val_tmp)
return (rewards, risks, val)
def init_minimonster(self):
learning_alg = lambda: sklearn.linear_model.LinearRegression()
self.base_learner = Semibandits.MiniMonster(self.passthrough,
learning_alg=learning_alg,
classification=False)
self.base_learner.init(self.T, params={'mu': self.mu})
self.base_learner.verbose = False
def run_lin_ucb(Bconstructor,
Bval,
rewards,
risks,
val,
T,
slate_len,
i,
start,
params=None):
if params is None:
B = Bconstructor()
L = Semibandits.LinUCB(B)
(r, reg, val_tmp) = L.play(T, verbose=True, validate=Bval)
rewards['lin'].append(r)
np.savetxt(outdir + "lin_rewards_%d.out" % (i), r)
if Bval is not None:
np.savetxt(outdir + "lin_validation_%d.out" % (i), val_tmp)
val['lin'].append(val_tmp)
print('Lin Final Validation %0.2f' % (val['lin'][i - start][-1]),
flush=True)
if params is not None:
for a in params:
B = Bconstructor()
L = Semibandits.LinUCB(B)
(r, reg, val_tmp) = L.play(T,
verbose=True,
validate=Bval,
params={'delta': a})
rewards['lin_%0.5f' % (a)].append(r)
np.savetxt(outdir + "lin_%0.5f_rewards_%d.out" % (a, i), r)
if Bval is not None:
np.savetxt(outdir + "lin_%0.5f_validation_%d.out" % (a, i),
val_tmp)
val['lin_%0.5f' % (a)].append(val_tmp)
print('Lin %0.5f Final Validation %0.2f' %
(a, val['lin_%0.5f' % (a)][i - start][-1]),
flush=True)
return (rewards, risks, val)
def LinearExperiment(ks, ds, iters=5, T=100):
"""
Run LinUCB against MiniMonster on a synthetic LinearBandit problem.
"""
outs = {}
for k in ks:
outs[k] = {}
for d in ds:
outs[k][d] = []
for i in range(iters):
print("Starting simulation with k=%d d=%d i=%d" % (k, d, i),
flush=True)
B = Simulators.LinearBandit(d, 1, k, noise=True)
M = Semibandits.MiniMonster(
B, sklearn.linear_model.LinearRegression)
(r, reg) = M.play(T, verbose=False)
L = Semibandits.LinUCB(B)
(r2, reg2) = L.play(T, verbose=False)
print("MiniMonster: %0.2f LinUCB: %0.2f" %
(r[len(r) - 1], r2[len(r2) - 1]),
flush=True)
outs[k][d].append((r, r2))
return outs
def init_linucb(self):
self.base_learner = Semibandits.LinUCB(self.passthrough)
self.base_learner.init(self.T, params={'delta': self.delta})
if Args.param is not None:
Args.param = float(Args.param)
rewards = []
regrets = []
for i in range(Args.iters):
S = Simulators.LinearBandit(Args.d,
1,
Args.K,
noise=Args.noise,
seed=i,
pos=False,
low=Args.s)
if Args.alg == 'linucb':
Alg = Semibandits.LinUCB(S)
if Args.param is not None:
start = time.time()
(r, reg, val_tmp) = Alg.play(Args.T,
verbose=True,
params={
'delta': Args.param,
'schedule': 10
})
stop = time.time()
if Args.alg == 'limecb':
Alg = LimeCB(S)
if Args.param is not None:
start = time.time()
(r, reg, val_tmp) = Alg.play(Args.T,
verbose=True,
S = Simulators.SemiparametricBandit(Args.d,
1,
Args.K,
noise=Args.noise,
seed=i,
pos=True)
if Args.dataset == 'semiparametric' and Args.feat == 'sphere':
S = Simulators.SemiparametricBandit(Args.d,
1,
Args.K,
noise=Args.noise,
seed=i,
pos=False)
if Args.alg == 'linucb':
Alg = Semibandits.LinUCB(S)
if Args.param is not None:
start = time.time()
(r, reg, val_tmp) = Alg.play(Args.T,
verbose=True,
params={
'delta': Args.param,
'schedule': 1
})
stop = time.time()
if Args.alg == 'bose':
Alg = BOSE(S)
if Args.param is not None:
start = time.time()
(r, reg, val_tmp) = Alg.play(Args.T,
verbose=True,