def setupSelfPlayZero(self):
#r = random.randint(0,self.iterNo-1)
r = self.iterNo-1
if self.ishumanFirstPlayer and self.ishumanCut:
self.model = deepq.load_act("model/selfPlayZero/shannon_switching_train_{}_{}_{}.pkl".format(1, 1, r))
elif self.ishumanFirstPlayer and not self.ishumanCut:
self.model = deepq.load_act("model/selfPlayZero/shannon_switching_train_{}_{}_{}.pkl".format(1, 0, r))
elif not self.ishumanFirstPlayer and self.ishumanCut:
self.model = deepq.load_act("model/selfPlayZero/shannon_switching_train_{}_{}_{}.pkl".format(0, 1, r))
else:
self.model = deepq.load_act("model/selfPlayZero/shannon_switching_train_{}_{}_{}.pkl".format(0, 0, r))
print("Self play Zero set up")
def main():
# setup environment
computerType = sys.argv[1]
ishumanFirstPlayer = int(sys.argv[2])
ishumanCut = int(sys.argv[3])
fileName = sys.argv[4]
env = gym.make('shannon_switching-v0')
env.configureEnvironment(computerType=computerType,
ishumanFirstPlayer=ishumanFirstPlayer,
ishumanCut=ishumanCut,
iterNo=20,
epsilon=0.2)
print("ishumanFirstPlayer ", ishumanFirstPlayer)
print("ishumanCut", ishumanCut)
model = deepq.load_act(fileName)
totalIterations = 2000
totalWins = 0
for i in range(totalIterations):
print(i)
state = env.reset()
while True:
state, reward, isOver, __ = env.step(model.step(state)[0][0])
if isOver != 0:
break
if reward == 1000:
totalWins += 1
print("Accuracy: ", totalWins / totalIterations)
def main():
##np.seterr(all='raise')
env = gym.make("apl-v0")
#act = deepq.learn(
#env,
#network='mlp',
#lr=1e-3,
#checkpoint_freq=None,
#total_timesteps=int(1e5),
#buffer_size=50000,
#exploration_fraction=0.1,
#exploration_final_eps=0.02,
#print_freq=10,
#load_path="./models/apl-v0-dqn-20181003-151750",
#callback=callback
#)
#timestr = time.strftime("%Y%m%d-%H%M%S")
#act.save("./models/apl-v0-dqn-" + timestr)
act = deepq.load_act("./models/apl-v0-dqn-20181003-152611.pickle")
while True:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
print("Episode reward", episode_rew)
def load_agent_from_file(name):
"""Loads a pickled RL agent from file"""
from baselines.deepq import load_act
# needed to get the unpickling to work since the pickling is done
# from a __name__=="__main__"
# pylint: disable=unused-import
from q_network import EdgeQNetwork
act = load_act(name)
return act
def main():
print('-*-*-*- enjoy worker -*-*-*-')
# tf.graph().as_default()
# tf.reset_default_graph()
env = gym.make("CartPole-v0")
act = deepq.load_act("model.pkl")
max_episodes = 5
while max_episodes > 0:
obs, done = env.reset(), False
episode_rew = 0
while not done:
env.render()
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
print("Episode reward", episode_rew)
max_episodes = max_episodes - 1
def test(env, load_path, num_episodes=1000):
act = deepq.load_act(load_path + ".pkl")
# success_count=0.0
test_render_file = open(load_path + ".txt", "w")
best_obs = np.ones(env.n * env.m, dtype=int)
best_episode_rew = -1 * env.n
for i in range(num_episodes):
obs, done = env.reset(), False
episode_rew = 0.0
while not done:
render_string = env.render(mode='ansi') + "\n"
test_render_file.write(render_string)
obs, rew, done, _ = env.step(act(obs[None])[0])
episode_rew += rew
if episode_rew < best_episode_rew:
best_episode_rew = episode_rew
best_obs = obs
render_string = env.render(mode='ansi') + "\n"
test_render_file.write(render_string)
test_render_file.write("Episode reward " + str(episode_rew) + "\n")
test_render_file.close()
obs_pm1 = best_obs * 2 - 1
state_pm1 = np.reshape(obs_pm1, [env.m, env.n])
print('State')
print(best_obs)
print(state_pm1)
print()
pairs = np.array(list(itertools.combinations(range(env.m), 2)))
first_i = pairs[:, 0]
second_i = pairs[:, 1]
# Compute auto
auto_corr_vectors = np.flip(np.fft.ifft(
np.fft.fft(np.flip(state_pm1, axis=1)) * np.fft.fft(state_pm1)),
axis=1)
mean_sqr_side_peak_auto = np.mean(
np.square(np.abs(auto_corr_vectors[:, 1:])))
var_sqr_side_peak_auto = np.var(
np.mean(np.square(np.abs(auto_corr_vectors[:, 1:])), axis=1))
var_sqr_side_peak_auto_norm = np.var(
np.mean(np.square(np.abs(auto_corr_vectors[:, 1:])), axis=1) /
(env.n * env.n))
print('Auto')
print(np.real(auto_corr_vectors))
print()
# Compute average balance
bal = np.mean(np.abs(np.sum(state_pm1, axis=1)))
# Compute cross
cross_corr_vectors = np.flip(np.fft.ifft(
np.fft.fft(np.flip(state_pm1[first_i, :], axis=1)) *
np.fft.fft(state_pm1[second_i, :])),
axis=1)
mean_sqr_side_peak_cross = np.mean(np.square(np.abs(cross_corr_vectors)))
var_sqr_side_peak_cross = np.var(
np.mean(np.square(np.abs(cross_corr_vectors)), axis=1))
var_sqr_side_peak_cross_norm = np.var(
np.mean(np.square(np.abs(cross_corr_vectors)), axis=1) /
(env.n * env.n))
print('Cross')
print(np.real(cross_corr_vectors))
print()
mean_sqr_side_peak = 0.5 * mean_sqr_side_peak_auto + 0.5 * mean_sqr_side_peak_cross
print('Mean sqr (auto):', mean_sqr_side_peak_auto)
print('Mean sqr (cross):', mean_sqr_side_peak_cross)
print('Mean sqr:', mean_sqr_side_peak)
print('Var sqr (auto):', var_sqr_side_peak_auto)
print('Var sqr (cross):', var_sqr_side_peak_cross)
print('Mean bal:', bal)
print()
print('----------Normalized----------')
print('Mean sqr (auto):', mean_sqr_side_peak_auto / (env.n * env.n))
print('Mean sqr (cross):', mean_sqr_side_peak_cross / (env.n * env.n))
print('Mean sqr:', mean_sqr_side_peak / (env.n * env.n))
print('Var sqr (auto):', var_sqr_side_peak_auto_norm)
print('Var sqr (cross):', var_sqr_side_peak_cross_norm)
print()
def main():
env = gym.make(ENV)
act = deepq.load_act(MODEL)
steps = 0
outfile = open(FILE, 'w')
bcfile = open(BC_FILE, 'w')
total_reward = 0
episodes = 0
while steps < 50000:
obs, done = env.reset(), False
episode_rew = 0
while not done:
#env.render()
state_1 = obs
if np.random.uniform(0, 1) <= RANDOM:
action = act(obs[None])[0]
else:
action = env.action_space.sample()
obs, rew, done, _ = env.step(action)
state_2 = obs
if RANDOM == DEFAULT:
# write to AON file
to_write = '['
for w in state_1:
to_write += str(w) + ','
to_write = to_write[:-1]
to_write += ']'
outfile.write(to_write)
outfile.write(" ")
to_write = '['
for w in state_2:
to_write += str(w) + ','
to_write = to_write[:-1]
to_write += ']'
outfile.write(to_write)
outfile.write("\n")
# write to BC file
to_write = '['
for w in state_1:
to_write += str(w) + ','
to_write = to_write[:-1]
to_write += ']'
bcfile.write(to_write)
bcfile.write(" ")
bcfile.write("[" + str(action) + "]")
bcfile.write(" ")
to_write = '['
for w in state_2:
to_write += str(w) + ','
to_write = to_write[:-1]
to_write += ']'
bcfile.write(to_write)
bcfile.write("\n")
episode_rew += rew
steps += 1
print(steps)
print("Episode reward", episode_rew)
total_reward += episode_rew
episodes += 1.
print("Average reward", total_reward / episodes)
outfile.close()
bcfile.close()
def __init__(self):
expert_path = data_root_path / "experts/mountaincar_deepq_custom.pickle"
from baselines import deepq
self.expert = deepq.load_act(expert_path)