JaeDukSeo

# l softmax
l_pull_count   = np.zeros((num_ep,num_bandit))
l_estimation   = np.zeros((num_ep,num_bandit))
l_reward       = np.zeros((num_ep,num_iter))
l_optimal_pull = np.zeros((num_ep,num_iter))
l_regret_total = np.zeros((num_ep,num_iter))
                    
for eps in range(num_ep):
    temp_pull_count   = np.zeros(num_bandit)
    temp_estimation   = np.zeros(num_bandit) + 1/num_bandit

JaeDukSeo

# k neural network (with adam)
k_pull_count   = np.zeros((num_ep,num_bandit))
k_estimation   = np.zeros((num_ep,num_bandit))
k_reward       = np.zeros((num_ep,num_iter))
k_optimal_pull = np.zeros((num_ep,num_iter))
k_regret_total = np.zeros((num_ep,num_iter))

def sigmoid(x): return 1/(1+np.exp(-x))
def d_sigmoid(x): return sigmoid(x)*(1-sigmoid(x))

JaeDukSeo

# j Thompson Sampling (uniform) (slow)
j_pull_count   = np.zeros((num_ep,num_bandit))
j_estimation   = np.zeros((num_ep,num_bandit))
j_reward       = np.zeros((num_ep,num_iter))
j_optimal_pull = np.zeros((num_ep,num_iter))
j_regret_total = np.zeros((num_ep,num_iter))

for eps in range(num_ep):

    temp_pull_count   = np.zeros(num_bandit)

JaeDukSeo

# i Thompson Sampling (beta) (slow)
i_pull_count   = np.zeros((num_ep,num_bandit))
i_estimation   = np.zeros((num_ep,num_bandit))
i_reward       = np.zeros((num_ep,num_iter))
i_optimal_pull = np.zeros((num_ep,num_iter))
i_regret_total = np.zeros((num_ep,num_iter))

for eps in range(num_ep):

    temp_pull_count   = np.zeros(num_bandit)

JaeDukSeo

# h UBC 1 Tuned
h_pull_count   = np.zeros((num_ep,num_bandit))
h_estimation   = np.zeros((num_ep,num_bandit))
h_reward       = np.zeros((num_ep,num_iter))
h_optimal_pull = np.zeros((num_ep,num_iter))
h_regret_total = np.zeros((num_ep,num_iter))

for eps in range(num_ep):
    temp_pull_count   = np.zeros(num_bandit) 
    temp_estimation   = np.zeros(num_bandit) 

JaeDukSeo

# g UBC 1
g_pull_count   = np.zeros((num_ep,num_bandit))
g_estimation   = np.zeros((num_ep,num_bandit))
g_reward       = np.zeros((num_ep,num_iter))
g_optimal_pull = np.zeros((num_ep,num_iter))
g_regret_total = np.zeros((num_ep,num_iter))

for eps in range(num_ep):
    temp_pull_count   = np.zeros(num_bandit)
    temp_estimation   = np.zeros(num_bandit)

JaeDukSeo

# f Linear Reward Penalty
f_pull_count   = np.zeros((num_ep,num_bandit))
f_estimation   = np.zeros((num_ep,num_bandit))
f_reward       = np.zeros((num_ep,num_iter))
f_optimal_pull = np.zeros((num_ep,num_iter))
f_regret_total = np.zeros((num_ep,num_iter))
                    
for eps in range(num_ep):
    alpha = 0.01
    beta  = 0.001

JaeDukSeo

# e Linear Reward Inaction
e_pull_count   = np.zeros((num_ep,num_bandit))
e_estimation   = np.zeros((num_ep,num_bandit))
e_reward       = np.zeros((num_ep,num_iter))
e_optimal_pull = np.zeros((num_ep,num_iter))
e_regret_total = np.zeros((num_ep,num_iter))
      
for eps in range(num_ep):
    learning_rate = 0.1
    temp_pull_count   = np.zeros(num_bandit)

JaeDukSeo

# d decy e greedy 
d_pull_count   = np.zeros((num_ep,num_bandit))
d_estimation   = np.zeros((num_ep,num_bandit))
d_reward       = np.zeros((num_ep,num_iter))
d_optimal_pull = np.zeros((num_ep,num_iter))
d_regret_total = np.zeros((num_ep,num_iter))

for eps in range(num_ep):
    epsilon = 1.0
    temp_pull_count   = np.zeros(num_bandit)

JaeDukSeo

# c e greedy 
c_pull_count   = np.zeros((num_ep,num_bandit))
c_estimation   = np.zeros((num_ep,num_bandit))
c_reward       = np.zeros((num_ep,num_iter))
c_optimal_pull = np.zeros((num_ep,num_iter))
c_regret_total = np.zeros((num_ep,num_iter))
                    
for eps in range(num_ep):
    epsilon = np.random.uniform(0,1)
    temp_pull_count   = np.zeros(num_bandit)