def __init__(self, m, means):
self.m = m
self.K = len(means)
self.mu_hat = np.zeros(self.K)
self.T = np.zeros(self.K)
self.means = means
self.bandit = BernoulliBandit(means=self.means)
def __init__(self, m, means, delta):
self.m = m
self.k = len(means)
self.mu_hat = np.zeros(self.k)
self.UCB = np.zeros(self.k)
self.T = np.zeros(self.k)
self.means = means
self.delta = delta
self.bandit = BernoulliBandit(means=self.means)
class KL_UCB(object):
def __init__(self, m, means):
self.m = m
self.k = len(means)
self.mu_hat = np.zeros(self.k)
self.estimate = np.zeros(self.k)
self.T = np.zeros(self.k)
self.means = means
self.bandit = BernoulliBandit(means=self.means)
def run_KL_UCB(self, n):
for t in range(n):
for i in range(self.k):
if self.T[i] == 0:
self.estimate[i] = 10**10
else:
thresh = np.log2(f(t)) / self.T[i]
self.estimate[i] = -1
for j in range(100):
mu_tilde = j / 100
tmp = d(self.mu_hat[i], mu_tilde)
if tmp <= thresh:
self.estimate[i] = mu_tilde
a_t = np.argmax(self.estimate)
reward_t = self.bandit.pull(a_t)
self.mu_hat[a_t] = (self.mu_hat[a_t] * self.T[a_t] +
reward_t) / (self.T[a_t] + 1)
self.T[a_t] += 1
def update_mean(self, a_t, reward_t):
self.mu_hat[a_t] = (self.mu_hat[a_t] * self.T[a_t] +
reward_t) / (self.T[a_t] + 1)
class Explore_then_commit(object):
def __init__(self, m, means):
self.m = m
self.K = len(means)
self.mu_hat = np.zeros(self.K)
self.T = np.zeros(self.K)
self.means = means
self.bandit = BernoulliBandit(means=self.means)
def run_ETC(self, n):
for t in range(n):
if t <= self.m * self.K:
a_t = (t % self.K)
else:
a_t = np.argmax(self.mu_hat)
reward_t = self.bandit.pull(a_t)
if t <= self.m * self.K:
self.mu_hat[a_t] = (self.mu_hat[a_t] * self.T[a_t] +
reward_t) / (self.T[a_t] + 1)
self.T[a_t] += 1
class Minimax_optimal_Strategy_in_stochastic_case(object):
def __init__(self, m, means):
self.m = m
self.K = len(means)
self.mu_hat = np.zeros(self.K)
self.estimate = np.zeros(self.K)
self.T = np.zeros(self.K)
self.means = means
self.bandit = BernoulliBandit(means=self.means)
def run_Moss(self, n):
for t in range(n):
for i in range(self.K):
if self.T[i] == 0:
self.estimate[i] = 10 ** 10
else:
self.estimate[i] = self.mu_hat[i] + np.sqrt(4*log_star(N/self.K * self.T[i])/self.T[i])
a_t = np.argmax(self.estimate)
reward_t = self.bandit.pull(a_t)
self.mu_hat[a_t] = (self.mu_hat[a_t] * self.T[a_t] + reward_t)/(self.T[a_t] + 1)
self.T[a_t] += 1
class Upper_Confidence_bound(object):
def __init__(self, m, means, delta):
self.m = m
self.k = len(means)
self.mu_hat = np.zeros(self.k)
self.UCB = np.zeros(self.k)
self.T = np.zeros(self.k)
self.means = means
self.delta = delta
self.bandit = BernoulliBandit(means=self.means)
def run_UCB(self, n):
for t in range(n):
a_t = np.argmax(self.UCB)
reward_t = self.bandit.pull(a_t)
for i in range(self.k):
if self.T[i] == 0:
self.UCB[i] = 10**10
else:
self.UCB[i] = self.mu_hat[i] + np.sqrt(
2 * np.log2(1 / self.delta) / self.T[i])
self.mu_hat[a_t] = (self.mu_hat[a_t] * self.T[a_t] +
reward_t) / (self.T[a_t] + 1)
self.T[a_t] += 1