def oneLR(mumax, mu):
""" One term of the Lai & Robbins lower bound for DiscreteArm arms: (mumax - mu) / KL(mu, mumax).
.. warning:: FIXME this is not correctly defined, except for the special case of having **only** 2 values, a ``DiscreteArm`` is NOT a one-dimensional distribution, and so the kl between two distributions is NOT a function of their mean!
"""
print(
"WARNING: DiscreteArm.oneLR({:.3g}, {:.3g}) is not defined, klBern is used but this is WRONG."
.format(mumax, mu)) # DEBUG
return (mumax - mu) / klBern(mu, mumax)
def kl(x, y):
""" The kl(x, y) to use for this arm.
.. warning:: FIXME this is not correctly defined, except for the special case of having **only** 2 values, a ``DiscreteArm`` is NOT a one-dimensional distribution, and so the kl between two distributions is NOT a function of their mean!
"""
print(
"WARNING: DiscreteArm.kl({:.3g}, {:.3g}) is not defined, klBern is used but this is WRONG."
.format(x, y)) # DEBUG
return klBern(x, y)
def kl(x, y):
""" The kl(x, y) to use for this arm."""
return klBern(x, y)
def oneLR(mumax, mu):
""" One term of the Lai & Robbins lower bound for UniformArm arms: (mumax - mu) / KL(mu, mumax). """
return (mumax - mu) / klBern(mu, mumax)