def random_test():
discovery_threshold
TRIALS = 200
K = 5 # number of randomly activated features
domain = Domains.PST()
np.random.seed(999999999)
initialRep = IndependentDiscretizationCompactBinary(domain)
rep = iFDD(
domain,
discovery_threshold,
initialRep,
debug=0,
useCache=1,
iFDDPlus=0)
rep.theta = np.arange(rep.features_num * domain.actions_num) * 10
n = rep.features_num
for i in xrange(TRIALS):
phi = np.zeros(n, 'bool')
for j in xrange(K):
ind = np.random.randint(0, n - 1)
phi[ind] = 1
threshold = np.random.rand() * 2 - 1
print '%d: %0.2f >> %s' % (i + 1, threshold, str(phi.nonzero()[0]))
rep.discover(phi, threshold)
rep.show()
def make_experiment(exp_id=1,
path="./Results/Temp/{domain}/{agent}/{representation}/",
discover_threshold=1.0,
lambda_=0.,
boyan_N0=20.1,
initial_learn_rate=0.330):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 100000
opt["num_policy_checks"] = 10
opt["checks_per_policy"] = 1
sparsify = 1
ifddeps = 1e-7
domain = IntruderMonitoring()
opt["domain"] = domain
initial_rep = IndependentDiscretization(domain)
representation = iFDD(domain,
discover_threshold,
initial_rep,
sparsify=sparsify,
useCache=True,
iFDDPlus=1 - ifddeps)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = SARSA(policy,
representation,
discount_factor=domain.discount_factor,
lambda_=lambda_,
initial_learn_rate=initial_learn_rate,
learn_rate_decay_mode="boyan",
boyan_N0=boyan_N0)
experiment = Experiment(**opt)
return experiment
def make_experiment(
exp_id=1, path="./Results/Temp/{domain}/{agent}/{representation}/",
discover_threshold=1.0,
lambda_=0.,
boyan_N0=20.1,
initial_learn_rate=0.330):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 100000
opt["num_policy_checks"] = 10
opt["checks_per_policy"] = 1
sparsify = 1
ifddeps = 1e-7
domain = IntruderMonitoring()
opt["domain"] = domain
initial_rep = IndependentDiscretization(domain)
representation = iFDD(domain, discover_threshold, initial_rep,
sparsify=sparsify,
useCache=True,
iFDDPlus=1 - ifddeps)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = SARSA(
policy, representation, discount_factor=domain.discount_factor,
lambda_=lambda_, initial_learn_rate=initial_learn_rate,
learn_rate_decay_mode="boyan", boyan_N0=boyan_N0)
experiment = Experiment(**opt)
return experiment
def random_test():
discovery_threshold
TRIALS = 200
K = 5 # number of randomly activated features
domain = Domains.PST()
np.random.seed(999999999)
initialRep = IndependentDiscretizationCompactBinary(domain)
rep = iFDD(
domain,
discovery_threshold,
initialRep,
debug=0,
useCache=1,
iFDDPlus=0)
rep.theta = np.arange(rep.features_num * domain.actions_num) * 10
n = rep.features_num
for i in range(TRIALS):
phi = np.zeros(n, 'bool')
for j in range(K):
ind = np.random.randint(0, n - 1)
phi[ind] = 1
threshold = np.random.rand() * 2 - 1
print('%d: %0.2f >> %s' % (i + 1, threshold, str(phi.nonzero()[0])))
rep.discover(phi, threshold)
rep.show()
def deterministic_test():
discovery_threshold = 1
sparsify = True
domain = rlpy.Domains.SystemAdministrator()
initialRep = IndependentDiscretizationCompactBinary(domain)
rep = iFDD(domain, discovery_threshold, initialRep,
debug=0, useCache=1, sparsify=sparsify)
rep.theta = np.arange(rep.features_num * domain.actions_num) * 10
print 'Initial [0,1] => ',
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1]))
print ANSWER
assert np.array_equal(ANSWER, np.array([0, 1]))
# rep.show()
print rep.inspectPair(0, 1, discovery_threshold + 1)
# rep.show()
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1]))
print ANSWER
assert np.array_equal(ANSWER, np.array([21]))
print 'Initial [2,3] => ',
ANSWER = np.sort(rep.findFinalActiveFeatures([2, 3]))
print ANSWER
assert np.array_equal(ANSWER, np.array([2, 3]))
# rep.showCache()
# rep.showFeatures()
# rep.showCache()
print 'Initial [0,20] => ',
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 20]))
print ANSWER
assert np.array_equal(ANSWER, np.array([0, 20]))
print 'Initial [0,1,20] => ',
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1, 20]))
print ANSWER
assert np.array_equal(ANSWER, np.array([20, 21]))
rep.showCache()
# Change the weight for new feature 40
rep.theta[40] = -100
print 'Initial [0,20] => ',
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 20]))
print ANSWER
assert np.array_equal(ANSWER, np.array([0, 20]))
print 'discover 0,1,20'
rep.inspectPair(20, rep.features_num - 1, discovery_threshold + 1)
# rep.showFeatures()
# rep.showCache()
print 'Initial [0,1,20] => ',
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1, 20]))
print ANSWER
assert np.array_equal(ANSWER, np.array([22]))
rep.showCache()
print 'Initial [0,1,2,3,4,5,6,7,8,20] => ',
ANSWER = np.sort(
rep.findFinalActiveFeatures([0, 1, 2, 3, 4, 5, 6, 7, 8, 20]))
print ANSWER
assert np.array_equal(ANSWER, np.array([2, 3, 4, 5, 6, 7, 8, 22]))
def deterministic_test():
discovery_threshold = 1
sparsify = True
domain = rlpy.Domains.SystemAdministrator()
initialRep = IndependentDiscretizationCompactBinary(domain)
rep = iFDD(domain, discovery_threshold, initialRep,
debug=0, useCache=1, sparsify=sparsify)
rep.theta = np.arange(rep.features_num * domain.actions_num) * 10
print('Initial [0,1] => ', end=' ')
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([0, 1]))
# rep.show()
print(rep.inspectPair(0, 1, discovery_threshold + 1))
# rep.show()
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([21]))
print('Initial [2,3] => ', end=' ')
ANSWER = np.sort(rep.findFinalActiveFeatures([2, 3]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([2, 3]))
# rep.showCache()
# rep.showFeatures()
# rep.showCache()
print('Initial [0,20] => ', end=' ')
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 20]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([0, 20]))
print('Initial [0,1,20] => ', end=' ')
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1, 20]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([20, 21]))
rep.showCache()
# Change the weight for new feature 40
rep.theta[40] = -100
print('Initial [0,20] => ', end=' ')
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 20]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([0, 20]))
print('discover 0,1,20')
rep.inspectPair(20, rep.features_num - 1, discovery_threshold + 1)
# rep.showFeatures()
# rep.showCache()
print('Initial [0,1,20] => ', end=' ')
ANSWER = np.sort(rep.findFinalActiveFeatures([0, 1, 20]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([22]))
rep.showCache()
print('Initial [0,1,2,3,4,5,6,7,8,20] => ', end=' ')
ANSWER = np.sort(
rep.findFinalActiveFeatures([0, 1, 2, 3, 4, 5, 6, 7, 8, 20]))
print(ANSWER)
assert np.array_equal(ANSWER, np.array([2, 3, 4, 5, 6, 7, 8, 22]))
def select_agent(name: Optional[str], seed: int) -> Agent:
tabular = Tabular(DOMAIN, discretization=20)
if name is None or name == 'tabular-lspi':
policy = eGreedy(tabular, epsilon=0.1)
return LSPI(policy, tabular, DOMAIN.discount_factor, MAX_STEPS, 1000)
elif name == 'tabular-q':
policy = eGreedy(tabular, epsilon=0.1)
return Q_Learning(policy, tabular, DOMAIN.discount_factor, lambda_=0.3)
elif name == 'tabular-sarsa':
policy = eGreedy(tabular, epsilon=0.1)
return SARSA(policy, tabular, DOMAIN.discount_factor, lambda_=0.3)
elif name == 'ifdd-q':
lambda_, boyan_N0 = 0.42, 202
discretization = 18
initial_rep = IndependentDiscretization(DOMAIN, discretization=discretization)
ifdd = iFDD(
DOMAIN,
discovery_threshold=8.63917,
initial_representation=initial_rep,
useCache=True,
iFDDPlus=True,
)
return Q_Learning(
eGreedy(ifdd, epsilon=0.1),
ifdd,
discount_factor=DOMAIN.discount_factor,
lambda_=lambda_,
initial_learn_rate=0.7422,
learn_rate_decay_mode='boyan',
boyan_N0=boyan_N0,
)
elif name == 'kifdd-q':
lambda_, boyan_N0 = 0.52738, 389.56
kernel_resolution = 8.567677
kernel_width = (DOMAIN.statespace_limits[:, 1] - DOMAIN.statespace_limits[:, 0]) \
/ kernel_resolution
kifdd = KernelizediFDD(
DOMAIN,
sparsify=True,
kernel=gaussian_kernel,
kernel_args=[kernel_width],
active_threshold=0.01,
discover_threshold=0.0807,
normalization=True,
max_active_base_feat=10,
max_base_feat_sim=0.5
)
policy = eGreedy(kifdd, epsilon=0.1)
return Q_Learning(
policy,
kifdd,
discount_factor=DOMAIN.discount_factor,
lambda_=lambda_,
initial_learn_rate=0.4244,
learn_rate_decay_mode='boyan',
boyan_N0=boyan_N0,
)
elif name == 'rbfs-q':
rbf = RBF(
DOMAIN,
num_rbfs=96,
resolution_max=21.0,
resolution_min=21.0,
const_feature=False,
normalize=True,
seed=seed,
)
policy = eGreedy(rbf, epsilon=0.1)
return Q_Learning(
policy,
rbf,
discount_factor=DOMAIN.discount_factor,
lambda_=0.1953,
initial_learn_rate=0.6633,
learn_rate_decay_mode='boyan',
boyan_N0=13444.0,
)
else:
raise NotImplementedError()
