def offline_attack(self):
for setting in self.settings_to_run:
R_c, success_prob = setting[0], setting[1]
M_in = get_M(self.M_0[0], self.M_0[1], R_c, success_prob)
env_in = environment.Environment(M_in)
pool = teacher.generate_pool(M_in[0], M_in[1], M_in[2], M_in[3], self.target_pi)
for tchr in self.teachers_to_run:
target_pi = tchr[2]["target_pi"]
p = tchr[1]
epsilon = tchr[2]["epsilon"]
epsilon_p = tchr[2]["epsilon_p"]
teacher_type = tchr[0]
cost_p = tchr[2]["cost_p"]
teacher_obj = teacher.teacher(env=env_in, target_pi=target_pi, p=p, epsilon=epsilon, epsilon_p=epsilon_p, teacher_type=teacher_type, pool=pool) #Pool here
print("==================================================")
try:
M_out, feasible = teacher_obj.get_target_M(M_in)
except Exception as e:
print("--teacher_type={}--R_c={}--P_success={}".format(teacher_type, R_c, success_prob))
if not feasible:
print("--teacher_type={}--R_c={}--P_success={}".format(teacher_type, R_c, success_prob))
cost = self.max_cost_value_if_non_feasible(cost_p)
self.append_cost_to_accumulator(cost, teacher_type, p, cost_p, success_prob, R_c)
continue
else:
print("--teacher_type={}--R_c={}--P_success={}".format(teacher_type, R_c, success_prob))
env_out = environment.Environment(M_out)
_, pi_T, _ = MDPSolver.averaged_valueIteration(env_out, env_out.reward)
cost = teacher_obj.cost(M_in, M_out, cost_p)
self.append_cost_to_accumulator(cost, teacher_type, p, cost_p, success_prob, R_c)
return self.accumulator
def offline_attack(self):
dict = {}
for setting in self.settings_to_run:
for tchr in self.teachers_to_run:
R_c, success_prob = setting[0], setting[1]
M_in = get_M(self.M_0[0], self.M_0[1], R_c, success_prob)
env_in = environment.Environment(M_in)
target_pi = tchr[2]["target_pi"]
p = tchr[1]
epsilon = tchr[2]["epsilon"]
epsilon_p = tchr[2]["epsilon_p"]
teacher_type = tchr[0]
teacher_obj = teacher.teacher(env=env_in, target_pi=target_pi, p=p, epsilon=epsilon, epsilon_p=epsilon_p, teacher_type=teacher_type, pool=None) #Pool here
# print("==================================================")
try:
##########
time_start = time.time()
##########
# print(time_start)
if "general_attack_on_dynamics" == tchr[0]:
pool = teacher.generate_pool(M_in[0], M_in[1], M_in[2], M_in[3], self.target_pi)
else:
pool = None
teacher_obj.pool = pool
# print("time after pool=", time.time())
M_out, feasible = teacher_obj.get_target_M(M_in)
#######
end_time = time.time()-time_start
#######
if feasible:
dict["time_R_c={}_Teacher_type={}_P={}".format(R_c, teacher_type, p)] = end_time
else:
dict["time_R_c={}_Teacher_type={}_P={}".format(R_c, teacher_type, p)] = "NF"
# print(end_time)
# print("R_c={}, Teacher_type={} --- Runtime = {}".format(R_c, teacher_type, end_time))
except Exception as e:
print(e)
print("time_R_c={}_Teacher_type={}_P={}".format(R_c, teacher_type, p))
pass
return dict
def __init__(self, env, target_pi, epsilon, p, epsilon_p, T_UCRL,
alpha_UCRL, attackers_cost_p, teacher_type):
self.env = env
self.reward_no_attack = env.reward
_, self.pi_no_attack, _ = MDPSolver.averaged_valueIteration(
env, env.reward)
self.target_pi = target_pi
self.epsilon = epsilon
self.p = p
self.epsilon_p = epsilon_p
self.T_UCRL = T_UCRL
self.attackers_cost_p = attackers_cost_p
self.alpha_UCRL = alpha_UCRL
self.teacher_type = teacher_type
self.M_0 = env.get_M_0()
self.pool = teacher.generate_pool(self.M_0[0], self.M_0[1],
self.M_0[2], self.M_0[3], target_pi)
self.teacher = teacher.teacher(env, target_pi, epsilon, p, epsilon_p,
teacher_type, self.pool)
self.learner = None