def epsilon_greedy(self,sensation,applicable_actions):
"""
Given self.epsilon() and self.Q(), return a distribution over
applicable_actions as an array where each element contains the
a probability mass for the corresponding action. I.e. The
action with the highest Q gets p = self.epsilon() and the
others get the remainder of the mass, uniformly distributed.
"""
Q = array([self.Q(sensation,action) for action in applicable_actions])
# simple epsilon-greedy policy
# get a vector with a 1 where each max element is, zero elsewhere
mask = (Q == mmax(Q))
num_maxes = len(nonzero(mask))
num_others = len(mask) - num_maxes
if num_others == 0: return mask
e0 = self.epsilon()/num_maxes
e1 = self.epsilon()/num_others
result = zeros(len(mask))+0.0
putmask(result,mask,1-e0)
putmask(result,mask==0,e1)
return result
def epsilon_greedy(self, sensation, applicable_actions):
"""
Given self.epsilon() and self.Q(), return a distribution over
applicable_actions as an array where each element contains the
a probability mass for the corresponding action. I.e. The
action with the highest Q gets p = self.epsilon() and the
others get the remainder of the mass, uniformly distributed.
"""
Q = array([self.Q(sensation, action) for action in applicable_actions])
# simple epsilon-greedy policy
# get a vector with a 1 where each max element is, zero elsewhere
mask = (Q == mmax(Q))
num_maxes = len(nonzero(mask))
num_others = len(mask) - num_maxes
if num_others == 0: return mask
e0 = self.epsilon() / num_maxes
e1 = self.epsilon() / num_others
result = zeros(len(mask)) + 0.0
putmask(result, mask, 1 - e0)
putmask(result, mask == 0, e1)
return result
def softmax(ar,temp):
"""
Given an array and a temperature, return the Boltzman distribution
over that array.
For an array X, and temp T returns a new array containing:
exp(Xi/T)/sum_j(exp(Xj/T) for all Xi.
If temp == 0 or any value in the array is inf, the function
returns the limit value as T -> 0.
"""
if temp == 0 or inf in ar:
v = (ar == mmax(ar))
return v/float(sum(v))
else:
numer = Numeric.exp(ar/float(temp))
denom = Numeric.sum(numer)
return numer/denom
def softmax(ar, temp):
"""
Given an array and a temperature, return the Boltzman distribution
over that array.
For an array X, and temp T returns a new array containing:
exp(Xi/T)/sum_j(exp(Xj/T) for all Xi.
If temp == 0 or any value in the array is inf, the function
returns the limit value as T -> 0.
"""
if temp == 0 or inf in ar:
v = (ar == mmax(ar))
return v / float(sum(v))
else:
numer = Numeric.exp(ar / float(temp))
denom = Numeric.sum(numer)
return numer / denom