def pick_action(self, observation):
"""Take random action prob epsilon, else be greedy."""
if np.random.rand() < self.epsilon:
action = np.random.randint(self.n_arm)
else:
posterior_means = self.get_posterior_mean()
action = random_argmax(posterior_means)
return action
def pick_action(self, observation):
"""Thompson sampling with Beta posterior for action selection."""
sampled_means = self.get_posterior_sample()
if random.random() < .2:
action = np.argmax(sampled_means[-1])
else:
action = random_argmax(sampled_means[:-1])
return action
def pick_action(self, observation):
"""Take random action prob epsilon, else be greedy."""
if np.random.rand() < self.epsilon:
action = np.random.randint(self.n_arm) # 从n个arm中随机选择一个
else: # 1-epsilon greedy
# 所谓reward, 就是success平均值
posterior_means = self.get_posterior_mean() # shape:[arm, 1], 从中选择一个reward最大的arm
action = random_argmax(posterior_means)
return action
def find_optimal_assortment(self, theta_hat):
'''finds the optimal assortment, given a sampled parameter.'''
# generating all possible assortments
assortment_tuples = list(itertools.product([0, 1], repeat=self.num_products))
total_profit = []
for assortment in assortment_tuples:
expected_demand = np.array(assortment)*np.exp(self.noise_var/2 +
theta_hat.dot(np.array(assortment)))
total_profit.append(expected_demand.dot(self.profits))
optimal_ind = random_argmax(np.array(total_profit))
return np.array(assortment_tuples[optimal_ind])
def _find_optimal_assortment(self):
'''finds the optimal assortment of the products.'''
# generating all possible assortments
assortment_tuples = list(
itertools.product([0, 1], repeat=self.num_products))
total_profit = []
for assortment in assortment_tuples:
expected_demand = np.array(assortment) * np.exp(
self.noise_var / 2 + self.theta.dot(np.array(assortment)))
total_profit.append(expected_demand.dot(self.profits))
optimal_ind = random_argmax(np.array(total_profit))
self.optimal_assortment = np.array(assortment_tuples[optimal_ind])
self.optimal_profit = total_profit[optimal_ind]
def pick_action(self, observation):
"""Thompson sampling with Beta posterior for action selection."""
sampled_means = self.get_posterior_sample()
action = random_argmax(sampled_means)
return action
def pick_action(self, observation): """Thompson sampling with Beta posterior for action selection.""" # 注意: 只有此处不一样, 即TS里是从后验分布中采样,而epsilon-greedy是计算期望 sampled_means = self.get_posterior_sample() # 每个arm都采样一个reward均值, [arm, 1] action = random_argmax(sampled_means) # 选择产生最大的均值的action return action