mserrors[:, i] = val - np.mean(val)
print noise, lambdas[np.argmin(val)]
#mserrors -= mserrors.min(axis=1)[:,None]
#mserrors /= mserrors.max(axis=1)[:,None]
return mserrors
if __name__ == "__main__":
plt.ion()
fn = "data/noise_lambda.npz"
if os.path.exists(fn):
d = np.load(fn)
globals().update(d)
else:
n_feat = 20
phi = features.lin_random(n_feat, n, constant=True)
mserrors2 = lambda_errors(phi, lambdas, noises)
phi = features.eye(n)
mserrors1 = lambda_errors(phi, lambdas, noises)
np.savez(fn,
mserrors1=mserrors1,
mserrors2=mserrors2,
noises=noises,
lambdas=lambdas)
mserrors2 -= mserrors2.min(axis=0)
mserrors2 /= mserrors2.max(axis=0)
mserrors1 -= mserrors1.min(axis=0)
#mserrors1 /= mserrors1.max(axis=0)
plt.figure()
mymap = mpl.colors.LinearSegmentedColormap.from_list(
Uniformly sampled random MDP with discrete states, on-policy case
"""
__author__ = "Christoph Dann <*****@*****.**>"
import td
import examples
from task import LinearDiscreteValuePredictionTask
import numpy as np
import features
import policies
import regtd
n = 400
n_a = 10
n_feat = 200
mdp = examples.RandomMDP(n, n_a)
phi = features.lin_random(n_feat, n, constant=True)
gamma = .95
np.random.seed(3)
beh_pol = policies.Discrete(np.random.rand(n, n_a))
tar_pol = beh_pol
task = LinearDiscreteValuePredictionTask(mdp, gamma, phi, np.zeros(phi.dim),
policy=beh_pol)
methods = []
alpha = 0.007
mu = .0001
gtd = td.GTD(alpha=alpha, beta=mu * alpha, phi=phi)
gtd.name = r"GTD $\alpha$={} $\mu$={}".format(alpha, mu)
gtd.color = "r"
methods.append(gtd)