def PCRL(S, A, H, d, L, eps):
# Make a very simple prior
mu = 0.
n_mu = 1.
tau = 1.
n_tau = 1.
prior_ng = posterior_sampling.convert_prior(mu, n_mu, tau, n_tau)
rew = 0
av_rew = []
c1 = len(S)
prior_dir = np.ones(c1)
R = {}
P = {}
time = range(H)
for l in xrange(L):
Rl = {}
Pl = {}
for t in time:
for s in S:
for a in A:
if (t, s, a) not in R:
R[(t, s, a)] = []
P[(t, s, a)] = np.array([0 for i in xrange(c1)])
if len(R[(t, s, a)]) == 0:
Rpost = prior_ng
Ppost = prior_dir
else:
data = np.array(R[(t, s, a)])
counts = P[(t, s, a)]
Rpost = posterior_sampling.update_normal_ig(
prior_ng, data)
Ppost = posterior_sampling.update_dirichlet(
prior_dir, counts)
Rl[(t, s,
a)] = posterior_sampling.sample_normal_ig(Rpost)[0]
Pl[(t, s, a)] = posterior_sampling.sample_dirichlet(Ppost)
mu = policy(R, P, Rl, Pl, S, A, H)
rew += play(mu, H, R, P, A, eps)
if (l + 1) % 10 == 0:
av_rew.append(rew / float(10))
rew = 0
return av_rew
def PCRL(S,A,H,d,L,eps):
# Make a very simple prior
mu = 0.
n_mu = 1.
tau = 1.
n_tau = 1.
prior_ng = posterior_sampling.convert_prior(mu, n_mu, tau, n_tau)
rew=0
av_rew=[]
c1 = len(S)
prior_dir = np.ones(c1)
R = {}
P = {}
time = range(H);
for l in xrange(L):
Rl = {}
Pl = {}
for t in time:
for s in S:
for a in A:
if (t,s,a) not in R:
R[(t,s,a)]=[]
P[(t,s,a)]=np.array([0 for i in xrange(c1)])
if len(R[(t,s,a)]) == 0:
Rpost = prior_ng
Ppost = prior_dir
else:
data = np.array(R[(t,s,a)])
counts = P[(t,s,a)]
Rpost = posterior_sampling.update_normal_ig(prior_ng, data)
Ppost = posterior_sampling.update_dirichlet(prior_dir, counts)
Rl[(t,s,a)]= posterior_sampling.sample_normal_ig(Rpost)[0]
Pl[(t,s,a)]= posterior_sampling.sample_dirichlet(Ppost)
mu = policy(R,P,Rl,Pl,S,A,H)
rew += play(mu,H,R,P,A,eps)
if (l+1)%10==0:
av_rew.append(rew/float(10))
rew = 0
return av_rew
'''
import numpy as np
import posterior_sampling
#-----------------------------------------------------------------------
# Updating rewards (normal gamma)
# Make a very simple prior
mu = 0.
n_mu = 1.
tau = 1.
n_tau = 1.
# Convert it to our nice format
prior_ng = posterior_sampling.convert_prior(mu, n_mu, tau, n_tau)
# Generate some real data
real_mu = 1.
real_prec = 4.
n_data = 100
data = np.zeros(n_data)
for i in range(n_data):
data[i] = np.random.normal(real_mu, np.sqrt(1. / real_prec))
print 'True Normal distribution: ' + str((real_mu, real_prec)) + '\n'
# Sampled data from the posterior
posterior_ng = posterior_sampling.update_normal_ig(prior_ng, data)
n_samp = 10
def PSRL(S,A,H,L):
"""
Computes the number of episodes it takes for PSRL to experience a positive
reward.
IN
S: list
States
A: list
Actions
H: int
Number of states and time frame
L: int
Number of episodes
OUT
success: int
Number of episodes before UCRL experiences a positive reward.
"""
# Make a very simple prior
mu = 0.
n_mu = 1.
tau = 1.
n_tau = 1.
prior_ng = posterior_sampling.convert_prior(mu, n_mu, tau, n_tau)
c1 = len(S)
prior_dir = np.ones(c1)
R = {}
P = {}
time = range(H);
av_rew = []
rew = 0
for l in xrange(L):
Rl = {}
Pl = {}
for t in time:
for s in S:
for a in A:
if (t,s,a) not in R:
R[(t,s,a)]=[]
P[(t,s,a)]=np.array([0 for i in xrange(c1)])
# If we have not visited (t,s,a) we don't update our prior
if len(R[(t,s,a)]) == 0:
Rpost = prior_ng
Ppost = prior_dir
else:
data = np.array(R[(t,s,a)])
counts = P[(t,s,a)]
# Posterior updating
Rpost = posterior_sampling.update_normal_ig(prior_ng, data)
Ppost = posterior_sampling.update_dirichlet(prior_dir, counts)
# Posterior sampling
Rl[(t,s,a)]= posterior_sampling.sample_normal_ig(Rpost)[0]
Pl[(t,s,a)]= posterior_sampling.sample_dirichlet(Ppost)
# Optimal policy
mu = policy(R,P,Rl,Pl,S,A,H)
#Episode
rew += play(mu,H,R,P)
if (l+1)%10==0:
print rew/float(10)
av_rew.append(rew/float(10))
rew = 0
return av_rew