Signomial programming finds a local solution to a problem of the form:
where each \(f\) is monomial while each \(g\) and \(h\) is a posynomial.
This requires multiple solutions of geometric programs, and so will take longer to solve than an equivalent geometric programming formulation.
The specification of a signomial problem can affect its solve time in a nuanced way:
gpkit.Model(x, [x >= 0.1, x+y >= 1, y <= 0.1]).localsolve() takes about twice as long to solve with cvxopt as
gpkit.Model(x, [x >= 1-y, y <= 0.1]).localsolve(), despite the two formulations being arithmetically equivalent and taking the same number of iterations.
In general, when given the choice of which variables to include in the positive-posynomial / \(g\) side of the constraint, the modeler should:
- maximize the number of variables in \(g\),
- prioritize variables that are in the objective,
- then prioritize variables that are present in other constraints.
.localsolve syntax was chosen to emphasize that signomial programming returns a local optimum. For the same reason, calling
.solve on an SP will raise an error.
By default, signomial programs are first solved conservatively (by assuming each \(h\) is equal only to its constant portion) and then become less conservative on each iteration.
"""Adapted from t_SP in tests/t_geometric_program.py""" import gpkit # Decision variables x = gpkit.Variable('x') y = gpkit.Variable('y') # must enable signomials for subtraction with gpkit.SignomialsEnabled(): constraints = [x >= 1-y, y <= 0.1] # create and solve the SP m = gpkit.Model(x, constraints) sol = m.localsolve(verbosity=0) print sol.table() assert abs(sol(x) - 0.9) < 1e-6
When using the
localsolve method, the
reltol argument specifies the relative tolerance of the solver: that is, by what percent does the solution have to improve between iterations? If any iteration improves less than that amount, the solver stops and returns its value.
If you wish to start the local optimization at a particular point \(x_k\), however, you may do so by putting that position (a dictionary formatted as you would a substitution) as the
Sequential Geometric Programs¶
The method of solving local GP approximations of a non-GP compatible model can be generalized, at the cost of the general smoothness and lack of a need for trust regions that SPs guarantee.
For some applications, it is useful to call external codes which may not be GP compatible. Imagine we wished to solve the following optimization problem:
This problem is not GP compatible due to the sin(x) constraint. One approach might be to take the first term of the Taylor expansion of sin(x) and attempt to solve:
"Can be found in gpkit/docs/source/examples/sin_approx_example.py" import numpy as np from gpkit import Variable, Model x = Variable("x") y = Variable("y") objective = y constraints = [y >= x, x <= np.pi/2., x >= np.pi/4., ] m = Model(objective, constraints) sol = m.solve(verbosity=0) print sol.table()
Cost ---- 0.7854 Free Variables -------------- x : 0.7854 y : 0.7854
We can do better, however, by utilizing some built in functionality of GPkit. Assume we have some external code which is capable of evaluating our incompatible function:
"""External function for GPkit to call. Can be found in gpkit/docs/source/examples/external_function.py""" import numpy as np def external_code(x): "Returns sin(x)" return np.sin(x)
Now, we can create a ConstraintSet that allows GPkit to treat the incompatible constraint as though it were a signomial programming constraint:
"Can be found in gpkit/docs/source/examples/external_constraint.py" from gpkit import ConstraintSet from gpkit.exceptions import InvalidGPConstraint from external_function import external_code class ExternalConstraint(ConstraintSet): "Class for external calling" # Overloading the __init__ function here permits the constraint class to be # called more cleanly at the top level GP. def __init__(self, x, y, **kwargs): # Calls the ConstriantSet __init__ function super(ExternalConstraint, self).__init__(, **kwargs) # We need a GPkit variable defined to return in our constraint. The # easiest way to do this is to read in the parameters of interest in # the initiation of the class and store them here. self.x = x self.y = y # Prevents the ExternalConstraint class from solving in a GP, thus forcing # iteration def as_posyslt1(self, substitutions=None): raise InvalidGPConstraint("ExternalConstraint cannot solve as a GP.") # Returns the ExternalConstraint class as a GP compatible constraint when # requested by the GPkit solver def as_gpconstr(self, x0): # Unpacking the GPkit variables x = self.x y = self.y # Creating a default constraint for the first solve if not x0: return (y >= x) # Returns constraint updated with new call to the external code else: # Unpack Design Variables at the current point x_star = x0["x"] # Call external code res = external_code(x_star) # Return linearized constraint return (y >= res*x/x_star)
and replace the incompatible constraint in our GP:
"Can be found in gpkit/docs/source/examples/external_sp.py" import numpy as np from gpkit import Variable, Model from external_constraint import ExternalConstraint x = Variable("x") y = Variable("y") objective = y constraints = [ExternalConstraint(x, y), x <= np.pi/2., x >= np.pi/4., ] m = Model(objective, constraints) sol = m.localsolve(verbosity=0) print sol.table()
Cost ---- 0.7071 Free Variables -------------- x : 0.7854 y : 0.7071
which is the expected result. This method has been generalized to larger problems, such as calling XFOIL and AVL.
If you wish to start the local optimization at a particular point \(x_0\), however, you may do so by putting that position (a dictionary formatted as you would a substitution) as the