def make_experiment(
exp_id=1, path="./Results/Temp/{domain}/{agent}/{representation}/",
discover_threshold=0.03613232738,
lambda_=0.,
boyan_N0=12335.665,
initial_learn_rate=0.037282,
discretization=6.):
opt = {}
opt["exp_id"] = exp_id
opt["max_steps"] = 50000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
sparsify = True
kappa = 1e-7
domain = InfCartPoleBalance()
opt["domain"] = domain
initial_rep = IndependentDiscretization(
domain,
discretization=discretization)
representation = iFDD(domain, discover_threshold, initial_rep,
sparsify=sparsify,
discretization=discretization,
useCache=True,
iFDDPlus=1 - kappa)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = SARSA(policy, representation,
lambda_=lambda_,
discount_factor=domain.discount_factor,
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}/",
boyan_N0=753,
initial_learn_rate=.7,
resolution=25.,
num_rbfs=206.,
lambda_=0.75):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 10000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
domain = InfCartPoleBalance()
opt["domain"] = domain
representation = RBF(domain, num_rbfs=int(num_rbfs),
resolution_max=resolution, resolution_min=resolution,
const_feature=False, normalize=True, seed=exp_id)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = Q_Learning(
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}/",
lambda_=0.,
boyan_N0=116.7025,
initial_learn_rate=0.01402,
discretization=6.):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 50000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
domain = InfCartPoleBalance()
opt["domain"] = domain
representation = IndependentDiscretization(domain,
discretization=discretization)
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}/",
boyan_N0=753,
initial_learn_rate=.7,
discretization=20.,
lambda_=0.75):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 5000
opt["num_policy_checks"] = 10
opt["checks_per_policy"] = 10
domain = InfCartPoleBalance(episodeCap=1000)
opt["domain"] = domain
representation = Tabular(domain, discretization=discretization)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = Q_Learning(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 makeComponents(self):
domain = InfCartPoleBalance()
representation = RepresentationFactory.get(config=self.representationConfig,
name=str(self.lstRepresentation.currentItem().text()),
domain=domain)
policy = PolicyFactory.get(config=self.policyConfig,
name=str(self.lstPolicy.currentItem().text()),
representation=representation)
agent = AgentFactory.get(config=self.agentConfig,
name=str(self.lstAgent.currentItem().text()),
representation=representation,
policy=policy)
return domain, agent
def initConfig(self):
domain=InfCartPoleBalance()
kernel_width = (domain.statespace_limits[:, 1] - domain.statespace_limits[:, 0]) \
/ 45.016
self.agentConfig['QLearning'] = {'lambda':0.6596, 'gamma':0.9, 'alpha':0.993, 'alpha_decay_mode':'boyan', 'boyan_N0':235}
self.agentConfig['Sarsa'] = {'lambda':0.6596, 'gamma':0.9, 'alpha':0.993, 'alpha_decay_mode':'boyan', 'boyan_N0':235}
self.policyConfig['eGreedy'] = {'epsilon':0.1}
self.representationConfig['Tabular'] = {'discretization':6}
self.representationConfig['IncrementalTabular'] = {'discretization':6}
self.representationConfig['RBF'] = {'num_rbfs':206, 'resolution_max':25, 'resolution_min':25}
self.representationConfig['iFDD'] = {'discretization':6, 'discover_threshold':0.037282}
self.representationConfig['KernelizediFDD']={'sparsify':1,'kernel':gaussian_kernel,
'kernel_args':[kernel_width],
'active_threshold':0.01,
'discover_threshold':0.01356,
'max_active_base_feat':10,
'max_base_feat_sim':0.5,
'kernel_resolution':45.016}
def make_experiment(exp_id=1,
path="./Results/Temp/{domain}/{agent}/{representation}/",
discover_threshold=.01356,
boyan_N0=235.,
lambda_=0.6596,
initial_learn_rate=.993,
kernel_resolution=45.016):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 10000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
active_threshold = 0.01
max_base_feat_sim = 0.5
sparsify = 1
domain = InfCartPoleBalance()
opt["domain"] = domain
kernel_width = old_div(
(domain.statespace_limits[:, 1] - domain.statespace_limits[:, 0]),
kernel_resolution)
representation = KernelizediFDD(domain,
sparsify=sparsify,
kernel=gaussian_kernel,
kernel_args=[kernel_width],
active_threshold=active_threshold,
discover_threshold=discover_threshold,
normalization=True,
max_active_base_feat=10,
max_base_feat_sim=max_base_feat_sim)
policy = eGreedy(representation, epsilon=0.1)
# agent = SARSA(representation,policy,domain,initial_learn_rate=1.,
# lambda_=0., learn_rate_decay_mode="boyan", boyan_N0=100)
opt["agent"] = Q_LEARNING(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=.01256,
lambda_=0.81,
boyan_N0=9811.2337,
initial_learn_rate=.15,
discretization=22.):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 10000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
sparsify = True
domain = InfCartPoleBalance()
opt["domain"] = domain
initial_rep = IndependentDiscretization(domain,
discretization=discretization)
representation = iFDD(domain,
discover_threshold,
initial_rep,
sparsify=sparsify,
discretization=discretization,
useCache=True,
iFDDPlus=True)
policy = eGreedy(representation, epsilon=0.1)
# agent = SARSA(representation,policy,domain,initial_learn_rate=1.,
# lambda_=0., learn_rate_decay_mode="boyan", boyan_N0=100)
opt["agent"] = Q_LEARNING(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=0.013461679,
lambda_=0.,
boyan_N0=484.78006,
initial_learn_rate=0.5651405,
discretization=23.):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 50000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
sparsify = True
kappa = 1e-7
domain = InfCartPoleBalance()
opt["domain"] = domain
initial_rep = IndependentDiscretization(domain,
discretization=discretization)
representation = iFDD(domain,
discover_threshold,
initial_rep,
sparsify=sparsify,
discretization=discretization,
useCache=True,
iFDDPlus=1. - kappa)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = Greedy_GQ(policy,
representation,
discount_factor=domain.discount_factor,
lambda_=lambda_,
BetaCoef=1e-6,
initial_learn_rate=initial_learn_rate,
learn_rate_decay_mode="boyan",
boyan_N0=boyan_N0)
experiment = Experiment(**opt)
return experiment