def __init__(self, g, H, **kwargs):
TruncatedCG.__init__(self, g, H, **kwargs)
self.name = 'Suff-CG'
self.qval = 0.0 # Initial value of quadratic objective.
self.best_decrease = 0
self.cg_reltol = kwargs.get('cg_reltol', 0.1)
self.detect_stalling = kwargs.get('detect_stalling', True)
def __init__(self, g, H, **kwargs):
TruncatedCG.__init__(self, g, H, **kwargs)
self.name = 'Suff-CG'
self.qval = 0.0 # Initial value of quadratic objective.
self.best_decrease = 0
self.cg_reltol = kwargs.get('cg_reltol', 0.1)
self.detect_stalling = kwargs.get('detect_stalling', True)
def __init__(self, g, H, **kwargs):
TrustRegionSolver.__init__(self, g, **kwargs)
self.cgSolver = TruncatedCG(g, H, **kwargs)
self.niter = 0
self.stepNorm = 0.0
self.step = None
self.m = None # Model value at candidate solution
class TrustRegionCG(TrustRegionSolver):
"""
Instantiate a trust-region subproblem solver based on the truncated
conjugate gradient method of Steihaug and Toint. See the :mod:`pcg` module
for more information.
The main difference between this class and the :class:`TrustRegionPCG`
class is that :class:`TrustRegionPCG` only accepts explicit
preconditioners (i.e. in matrix form). This class accepts an implicit
preconditioner, i.e., any callable object.
"""
def __init__(self, g, H, **kwargs):
TrustRegionSolver.__init__(self, g, **kwargs)
self.cgSolver = TruncatedCG(g, H, **kwargs)
self.niter = 0
self.stepNorm = 0.0
self.step = None
self.m = None # Model value at candidate solution
def Solve(self, **kwargs):
"""
Solve trust-region subproblem using the truncated conjugate-gradient
algorithm.
"""
self.cgSolver.Solve(**kwargs)
self.niter = self.cgSolver.niter
self.stepNorm = self.cgSolver.stepNorm
self.step = self.cgSolver.step
# Compute model reduction.
self.m = np.dot(self.g, self.step)
self.m += 0.5 * np.dot(self.step, self.cgSolver.H * self.step)
return
def __init__(self, J, c, radius=None, transposed=False, **kwargs):
"""
:parameters:
:J: coefficient matrix (may be rectangular)
:c: constant vector (numpy array)
:radius: positive real number or None (default: None)
:transpose: if set to True, replace J with its transpose.
`J` should be a linear operator.
Additional keyword arguments are passed directly to `TruncatedCG`.
Upon completion, the member `step` is set to the most recent solution
estimate. See the documentation of `TruncatedCG` for more
information.
"""
self.op = SquaredLinearOperator(J, transposed=transposed)
TruncatedCG.__init__(self, J.T * c, self.op, radius=radius, **kwargs)
return
def __init__(self, J, c, radius=None, transposed=False, **kwargs):
"""
:parameters:
:J: coefficient matrix (may be rectangular)
:c: constant vector (numpy array)
:radius: positive real number or None (default: None)
:transpose: if set to True, replace J with its transpose.
`J` should be a linear operator.
Additional keyword arguments are passed directly to `TruncatedCG`.
Upon completion, the member `step` is set to the most recent solution
estimate. See the documentation of `TruncatedCG` for more
information.
"""
self.op = SquaredLinearOperator(J, transposed=transposed)
TruncatedCG.__init__(self, J.T * c, self.op, radius=radius, **kwargs)
return
