def test_visual():
""" Test 2-D basis func visualization. """
kernel = gaussian_kernel
normalization=False
domain = InfiniteTrackCartPole.InfTrackCartPole() #2 continuous dims
discretization = 20 # not used
num = 1 # number of basis functions to use
resolution_min=1
resolution_max=5
rep = RandomLocalBases(domain, kernel, num, resolution_min,
resolution_max,
seed=1, normalization=normalization,
discretization=discretization)
rep.plot_2d_feature_centers()
def test_visual():
""" Test 2-D basis func visualization. """
kernel = gaussian_kernel
normalization=False
domain = InfiniteTrackCartPole.InfTrackCartPole() #2 continuous dims
discretization = 20 # not used
num = 1 # number of basis functions to use
resolution_min=1
resolution_max=5
rep = RandomLocalBases(domain, kernel, num, resolution_min,
resolution_max,
seed=1, normalization=normalization,
discretization=discretization)
rep.plot_2d_feature_centers()
def test_parametric_rep():
"""
For fixed representations: test successful kernel function use, using
varying number of features.
Ensure get expected result. Test normalization, ensure expected result.
"""
for normalization in [False, True]: # verify everything with/out norm
kernel = gaussian_kernel
domain = InfiniteTrackCartPole.InfTrackCartPole() #2 continuous dims
discretization = 20 # not used
num = 1 # number of basis functions to use IN EACH DIMENSION
resolution_min = 1
resolution_max = 5
rep = RandomLocalBases(domain,
kernel,
num,
resolution_min,
resolution_max,
seed=1,
normalization=normalization,
discretization=discretization)
assert rep.features_num == num # in reality, theres one in each dim.
# Center lies within statespace limits
assert np.all(domain.statespace_limits[:, 0] <= rep.centers[0])
assert np.all(rep.centers[0] <= domain.statespace_limits[:, 1])
# width lies within resolution bounds
statespace_range = domain.statespace_limits[:,
1] - domain.statespace_limits[:,
0]
assert np.all(statespace_range / resolution_max <=
rep.widths[0]) # widths[0] has `state_space_dims` cols
assert np.all(rep.widths[0] <= statespace_range / resolution_min)
phiVecOrigin = rep.phi(np.array([0, 0], dtype=np.float),
terminal=False)
assert np.all(phiVecOrigin >= 0) # nonnegative feat func values
# feature func only dependent on own dimension
phiVec2 = rep.phi(np.array([0, 1], dtype=np.float), terminal=False)
if normalization:
assert sum(phiVecOrigin) == 1
assert sum(phiVec2) == 1
def make_experiment(exp_id=1,
path="./Results/Temp",
initial_learn_rate=.40,
lambda_=0.,
resolution=25,
num_rbfs=300):
"""
Each file specifying an experimental setup should contain a
make_experiment function which returns an instance of the Experiment
class with everything set up.
@param id: number used to seed the random number generators
@param path: output directory where logs and results are stored
"""
# import sys
# import os
# cur_dir = os.path.expanduser("~/work/clipper/models/rl/")
# sys.path.append(cur_dir)
# from Domains import RCCarModified
# from Policies import RCCarGreedy
# Experiment variables
opt = {}
opt["path"] = path
opt["exp_id"] = exp_id
opt["max_steps"] = 200000
opt["num_policy_checks"] = 15
opt["checks_per_policy"] = 2
# Logging
domain = RCCarLeftTurn(noise=0.)
opt["domain"] = domain
# Representation
kernel = gaussian_kernel
representation = RandomLocalBases(domain,
gaussian_kernel,
num=int(num_rbfs),
normalization=True,
resolution_max=resolution,
seed=exp_id)
policy = eGreedy(representation, epsilon=0.15)
# if biasedaction > -1:
# print "No Random starts with biasing {}".format(i % 4)
# policy = BiasedGreedy(representation, epsilon=0.5, biasedaction=biasedaction)
# Agent
opt["agent"] = Q_Learning(policy,
representation,
domain.discount_factor,
initial_learn_rate=initial_learn_rate,
lambda_=lambda_,
learn_rate_decay_mode="const")
experiment = Experiment(**opt)
return experiment
def test_parametric_rep():
"""
For fixed representations: test successful kernel function use, using
varying number of features.
Ensure get expected result. Test normalization, ensure expected result.
"""
for normalization in [False, True]: # verify everything with/out norm
kernel = gaussian_kernel
domain = InfiniteTrackCartPole.InfTrackCartPole() #2 continuous dims
discretization = 20 # not used
num = 1 # number of basis functions to use IN EACH DIMENSION
resolution_min=1
resolution_max=5
rep = RandomLocalBases(domain, kernel, num, resolution_min,
resolution_max,
seed=1, normalization=normalization,
discretization=discretization)
assert rep.features_num == num # in reality, theres one in each dim.
# Center lies within statespace limits
assert np.all(domain.statespace_limits[:,0] <= rep.centers[0])
assert np.all(rep.centers[0] <= domain.statespace_limits[:,1])
# width lies within resolution bounds
statespace_range = domain.statespace_limits[:, 1] - domain.statespace_limits[:, 0]
assert np.all(statespace_range / resolution_max <= rep.widths[0]) # widths[0] has `state_space_dims` cols
assert np.all(rep.widths[0] <= statespace_range / resolution_min)
phiVecOrigin = rep.phi(np.array([0,0], dtype=np.float), terminal=False)
assert np.all(phiVecOrigin >= 0) # nonnegative feat func values
# feature func only dependent on own dimension
phiVec2 = rep.phi(np.array([0,1], dtype=np.float), terminal=False)
if normalization:
assert sum(phiVecOrigin) == 1
assert sum(phiVec2) == 1
