def feature_extractor_2(state, action, normalize=True):
# features are
# - bias
# - the elements of the state vector
# - all second order terms (mixed and pure)
# - sines of every first and second order term
# - cosines of every first and second order term
# all of that times two (i.e. for each of the actions)
s_prime = np.r_[1, state] # dim(state) + 1
# dim(state) (dim(state) + 1) / 2
quadratic = np.outer(s_prime, s_prime)[np.tril_indices(s_prime.shape[0])].reshape(-1)
# (dim(state) (dim(state) + 1) / 2) - 1
sines = np.sin(quadratic[1:])
cosines = np.cos(quadratic[1:])
# dim(state)(dim(state) + 1) - 2 + dim(state)(dim(state) + 1) / 2
state_feats = np.r_[quadratic, sines, cosines]
# dim(state_feats) * 2
features = np.outer(state_feats, np.array([0, 1]) == action).T.reshape(-1)
if normalize:
# normalize everything but the bias.
norm = features / (np.linalg.norm(features))
norm[0] = int(action == 0) * 1.0
norm[state_feats.shape[0]] = int(action == 1) * 1.0
return norm
else:
return features
评论列表
文章目录
