reward_function = 'simple' # simple or lbph
seed = random.randint(0, 1000)
print "Using seed=%i" % seed
ce = CubeEnvironment(n=3, seed=seed, whiteplastic=False, reward_function=reward_function)
ce.randomize(1)
# Create actor-critic models
n_hidden = 50
activation_func = 'gaussian'
t = 10
gamma = 0.99
lambda_ = 0.3 # default: 0.7
N = 5000
M = 8 # Total movements
policy_model = PolicyModel(ce, LBPFeatureTransformer(), t=t, lambda_=lambda_, gamma=gamma)
value_model = ValueModel(ce, LBPFeatureTransformer(), n_hidden=n_hidden, activation_func=activation_func,
t=t, lambda_=lambda_, gamma=gamma)
max_movements_for_cur_iter, max_movements_for_random = 0, 0
total_rewards = np.empty(N)
total_iters = np.empty(N)
for n in range(N):
#eps = 1.0/np.sqrt(n+1)
eps = 1.0/(0.001*n+1)
#eps = 0.7
prev_max_movements_for_cur_iter, prev_max_movements_for_random = max_movements_for_cur_iter, max_movements_for_random
max_movements_for_cur_iter, max_movements_for_random = max_movements(n, N, M)
if prev_max_movements_for_cur_iter != max_movements_for_cur_iter:
print "Now playing for a max of %i movements..." % max_movements_for_cur_iter
def __init__(self, env):
self.env = env
self.models = {}
self.feature_transformer = LBPFeatureTransformer()
for a in env.actions_available:
self.models[a] = dict()
from rl3.environment.base import CubeEnvironment
import matplotlib.pyplot as plt
import numpy as np
if __name__ == '__main__':
reward_function = 'lbph' # simple or lbph
seed = 39 # np.random.randint(0, 100)
print "Using seed=%i" % seed
ce = CubeEnvironment(n=3,
seed=seed,
whiteplastic=False,
reward_function=reward_function)
ce.randomize(1)
#model = QubeTabularAgent(ce)
model = QubeRegAgent(ce, LBPFeatureTransformer())
# TODO: crear un mecanismo de attachments para el env (por ej, para monitorear algoritmos seguidos en cada iter)
gamma = 0.99
N = 3000
totalrewards = np.empty(N)
costs = np.empty(N)
for n in range(N):
eps = 1.0 / np.sqrt(n + 1)
totalreward = play_one(model, ce, eps, gamma, max_iters=100)
totalrewards[n] = totalreward
if n % 100 == 0:
print "episode:", n, "total reward:", totalreward, "eps:", eps, "avg reward (last 100):", \
totalrewards[max(0, n-100):(n+1)].mean()
# print "Algorithm followed: %s" % ce.actions_taken
sum_movements = sum(movements)
probabilities = [m / float(sum_movements) for m in movements]
return int(np.random.choice(movements, p=probabilities))
if __name__ == '__main__':
reward_function = 'simple' # simple or lbph
seed = random.randint(0, 1000)
print "Using seed=%i" % seed
ce = CubeEnvironment(n=3,
seed=seed,
whiteplastic=False,
reward_function=reward_function)
ce.randomize(1)
#model = QubeTabularAgent(ce)
model = QubeBoltzmannRegAgent(ce, LBPFeatureTransformer())
# TODO: crear un mecanismo de attachments para el env (por ej, para monitorear algoritmos seguidos en cada iter)
# TODO: experience replay, pero mayormente de los estados que llevaron a un reward positivo
gamma = 0.99
lambda_ = 0.3 # default: 0.7
N = 5000
M = 10 # Total movements
max_movements_for_cur_iter, max_movements_for_random = 0, 0
totalrewards = np.empty(N)
for n in range(N):
#eps = 1.0/np.sqrt(n+1)
eps = 1.0 / (0.001 * n + 1)
#eps = 0.7
prev_max_movements_for_cur_iter, prev_max_movements_for_random = max_movements_for_cur_iter, max_movements_for_random
#max_movs = int(2.4270509831248424 ** max_movs_for_random) # num_aureo * n_caras / n vertices
max_movs = min(max_movs, 100)
return max_movs, max_movs_for_random
if __name__ == '__main__':
reward_function = 'lbph' # simple or lbph
seed = random.randint(0, 1000)
print "Using seed=%i" % seed
ce = CubeEnvironment(n=3,
seed=seed,
whiteplastic=False,
reward_function=reward_function)
ce.randomize(1)
#model = QubeTabularAgent(ce)
model = QubeBloomPARAgent(ce, LBPFeatureTransformer())
#model = QubeBloomPARAgent(ce, LBPFeatureTransformer())
#model = QubeBloomAgent(ce, LBPFeatureTransformer())
# TODO: crear un mecanismo de attachments para el env (por ej, para monitorear algoritmos seguidos en cada iter)
gamma = 0.99
lambda_ = 0.7 # default: 0.7
N = 10000
M = 10 # Total movements
max_movements_for_cur_iter, max_movements_for_random = 0, 0
totalrewards = np.empty(N)
for n in range(N):
#eps = 1.0/np.sqrt(n+1)
eps = 1.0 / (0.001 * n + 1)
#eps = 0.7