def make_experiment(exp_id=1,
path="./Results/Temp/{domain}/{agent}/{representation}/",
boyan_N0=2120,
initial_learn_rate=.26,
lambda_=0.9,
resolution=8,
num_rbfs=4958):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 30000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
domain = FiniteCartPoleBalanceOriginal(good_reward=0.)
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}/",
boyan_N0=119,
initial_learn_rate=.06,
discretization=34):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 30000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 1
domain = FiniteCartPoleBalanceOriginal(good_reward=0.)
opt["domain"] = domain
representation = IncrementalTabular(domain, discretization=discretization)
policy = eGreedy(representation, epsilon=0.1)
opt["agent"] = Q_Learning(policy,
representation,
discount_factor=domain.discount_factor,
lambda_=0.9,
initial_learn_rate=initial_learn_rate,
learn_rate_decay_mode="boyan",
boyan_N0=boyan_N0)
experiment = Experiment(**opt)
return experiment
def makeComponents(self):
#noise = float(self.spNoise.value())
position1 = float(self.spPosition1.value())
position2 = float(self.spPosition2.value())
velocity1 = float(self.spVelocity1.value())
velocity2 = float(self.spVelocity2.value())
availForce1 = float(self.spAvailForce1.value())
availForce2 = float(self.spAvailForce2.value())
domain = FiniteCartPoleBalanceOriginal(good_reward=0.)
domain.POSITION_LIMITS = [position1, position2]
#: m/s - Default limits on cart velocity [per RL Community CartPole]
domain.VELOCITY_LIMITS = [velocity1, velocity2]
#: Newtons, N - Force values available as actions
domain.AVAIL_FORCE = np.array([availForce1, availForce2])
#: kilograms, kg - Mass of the pendulum arm
domain.MASS_PEND = float(self.spMassPend.value())
#: kilograms, kg - Mass of cart
domain.MASS_CART = float(self.spMassCart.value())
#: meters, m - Physical length of the pendulum, meters (note the moment-arm lies at half this distance)
domain.LENGTH = float(self.spLength.value())
# m - Length of moment-arm to center of mass (= half the pendulum length)
domain.MOMENT_ARM = domain.LENGTH / 2.
# 1/kg - Used in dynamics computations, equal to 1 / (MASS_PEND +
# MASS_CART)
domain._ALPHA_MASS = 1.0 / (domain.MASS_CART + domain.MASS_PEND)
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 make_experiment(exp_id=1,
path="./Results/Temp/{domain}/{agent}/{representation}/",
discover_threshold=.21,
boyan_N0=37.,
lambda_=.9,
initial_learn_rate=.07,
kernel_resolution=13.14):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 30000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
active_threshold = 0.01
max_base_feat_sim = 0.5
sparsify = 1
domain = FiniteCartPoleBalanceOriginal(good_reward=0.)
opt["domain"] = domain
# domain = FiniteCartPoleBalanceModern()
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=initial_learn_rate,
# lambda_=.0, learn_rate_decay_mode="boyan", boyan_N0=boyan_N0)
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=77.,
boyan_N0=11,
lambda_=0.9,
initial_learn_rate=.05,
discretization=47):
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 30000
opt["num_policy_checks"] = 20
opt["checks_per_policy"] = 10
sparsify = 1
domain = FiniteCartPoleBalanceOriginal(good_reward=0.)
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=initial_learn_rate,
# lambda_=.0, learn_rate_decay_mode="boyan", boyan_N0=boyan_N0)
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