in isoexp/mab/smab_algs.py [0:0]
def attacked_EXP3_P(T, MAB, target_arm, eta = None, gamma = None, delta=0.99, constant_attack=False):
K = len(MAB)
estimated_S = np.zeros((K,))
rewards = np.zeros((T,))
draws = 0 * rewards
sum_exp = K
exp_estimated_S = np.ones((K,))
arms = np.linspace(0, K - 1, K, dtype='int')
N = np.ones((K,)) # number of observations of each arm
S = np.zeros((K,))
beta = np.zeros((K,))
attacks = np.zeros((T,))
time_of_attacks = np.zeros((T,))
if eta is None and gamma is None:
eta = np.sqrt(np.log(K + 1) / (K * T))
gamma = 0
elif eta is None:
eta = np.sqrt(np.log(K + 1) / (K * T))
elif gamma is None:
gamma = 0
for t in range(T):
P = (1 - gamma) * exp_estimated_S / sum_exp + gamma/K*np.ones((K,))
if t < K:
action = t
attack_t = 0
else:
action = np.random.choice(arms, p=P)
if action != target_arm:
time_of_attacks[t] = 1
if constant_attack:
attack_t = - 2*np.maximum(0, MAB[action].mean - MAB[target_arm].mean)
else:
beta = np.sqrt(np.log(np.pi ** 2 * K * N ** 2 / (3 * delta)) / (2 * N))
attack_t = - np.maximum((S / N)[action] - (S / N)[target_arm] + beta[action] + beta[target_arm], 0)
else:
attack_t = 0
attacks[t] = attack_t
true_X = 1 * MAB[action].sample().squeeze()
X = true_X + attack_t
estimated_S = estimated_S + 1
estimated_S[action] = estimated_S[action] - (1 - X) /P[action]
exp_estimated_S = exp_estimated_S*np.exp(eta)
exp_estimated_S[action] = exp_estimated_S[action] * np.exp(eta * (- (1 - X) /P[action]))
sum_exp = np.sum(exp_estimated_S)
rewards[t] = true_X
draws[t] = action
N[action] += 1
S[action] += true_X
return rewards, draws, attacks, time_of_attacks