def _tc3():
env = loader.get_env("AntBulletEnv-v0", "ppo2_norm")
actor = loader.get_original_policy("AntBulletEnv-v0", "ppo2_norm")
obs = env.reset()
s_size = len(obs)
a_size = env.action_space.shape[0]
q_net = QNetwork(s_size, a_size)
sarsa = SARSA(env, actor, q_net, attack_type="obs")
sarsa.train(1000, 100, 5, 10)
def learn_q(env_name,
algo,
step_num,
rollout_num,
update_intv,
update_iter,
gamma=0.99,
attack_type="state",
optimizer=None):
env = loader.get_env(env_name, algo)
actor = loader.get_original_policy(env_name, algo)
if attack_type == "state":
env.reset()
s_size = len(env.unwrapped.state)
else:
s_size = env.observation_space.shape[0]
a_size = env.action_space.shape[0]
q_net = QNetwork(s_size, a_size)
sarsa = SARSA(env, actor, q_net, optimizer, attack_type=attack_type)
sarsa.train(step_num, rollout_num, update_intv, update_iter, gamma)
return q_net, sarsa.losses
def get_pi(env_name, algo):
return loader.get_original_policy(env_name, algo)
def _surrogate():
env = loader.get_env(env_name, algo, reward_type=None)
pi = loader.get_original_policy(env_name, algo)
return simulation_attack_obs(env, pi, pi_net, q_net, step_num,
rollout_num // thread_number, attack_fn,
attack_freq, attack_kwargs)
def _surrogate():
env = loader.get_env(env_name, algo)
pi = loader.get_original_policy(env_name, algo)
return simulation(env, pi, actor_net, step_num,
rollout_num // thread_number, attack_fn, attack_freq,
attack_kwargs)