Optimization¶
Let \(X\) be a nonempty topological space and \(A\) be a unital sub-algebra of continuous functions over \(X\) which separates points of \(X\). We consider the following optimization problem:
Denote the feasibility set of the above program by \(K\) (i.e., \(K=\{x\in X:g_i(x)\ge 0,~ i=1,\dots,m\}\)). Let \(\rho\) be the optimum value of the above program and \(\mathcal{M}_1^+(K)\) be the space of all probability Borel measures supported on \(K\). One can show that:
This associates a \(K\)-positive linear functional \(L_{\mu}\) to every measure \(\mu\in\mathcal{M}_1^+(K)\). Let us denote the set of all elements of \(A\) nonnegative on \(K\) by \(Psd_A(K)\). If \(\exists p\in Psd_A(K)\) such that \(p^{-1}([0, n])\) is compact for each \(n\in\mathbb{N}\), then one can show that every \(K\)-positive linear functional admits an integral representation via a Borel measure on \(K\) (Marshall’s generalization of Haviland’s theorem). Let \(Q_{\bf g}\) be the quadratic module generated by \(g_1,\dots,g_m\), i.e, the set of all elements in \(A\) of the form
where \(\sigma_0,\dots,\sigma_m\in\sum A^2\) are sums of squares of elements of \(A\). A quadratic module \(Q\) is said to be Archimedean if for every \(h\in A\) there exists \(M>0\) such that \(M\pm h\in Q\). By Jacobi’s representation theorem, if \(Q\) is Archimedean and \(h>0\) on \(K\), where \(K=\{x\in X:g(x)\ge0~\forall g\in Q\}\), then \(h\in Q\). Since \(Q\) is Archimedean, \(K\) is compact and this implies that if a linear functional on \(A\) is nonnegative on \(Q\), then it is \(K\)-positive and hence admits an integral representation. Therefore:
Let \(Q=Q_{\bf g}\) and \(L(Q)\subseteq[0,\infty)\). Then clearly \(L(\sum A^2)\subseteq[0,\infty)\) which means \(L\) is positive semidefinite. Moreover, for each \(i=1,\dots,m\), \(L(g_i\sum A^2)\subseteq[0,\infty)\) which means the maps
are positive semidefinite. So the optimum value of the following program is still equal to \(\rho\):
This, still is not a semidefinite program as each constraint is infinite dimensional. One plausible idea is to consider functionals on finite dimensional subspaces of \(A\) containing \(f, g_1,\dots,g_m\). This was done by Lasserre for polynomials [JBL].
Let \(B\subseteq A\) be a linear subspace. If \(L:A\longrightarrow\mathbb{R}\) is \(K\)-positive, so is its restriction on \(B\). But generally, \(K\)-positive maps on \(B\) do not extend to \(K\)-positive one on \(A\) and hence existence of integral representations are not guaranteed. Under a rather mild condition, this issue can be resolved:
Theorem. [GIKM] Let \(K\subseteq X\) be compact, \(B\subseteq A\) a linear subspace such that there exists \(p\in B\) strictly positive on \(K\). Then every linear functional \(L:B\longrightarrow\mathbb{R}\) satisfying \(L(Psd_B(K))\subseteq[0,\infty)\) admits an integral representation via a Borel measure supported on \(K\).
Now taking \(B\) to be a finite dimensional linear space containing \(f, g_1,\dots,g_m\) and satisfying the assumptions of the above theorem, turns (2) into a semidefinite program. Note that this does not imply that the optimum value of the resulting SDP is equal to \(\rho\) since
- \(Q_{\bf g}\cap B\neq Psd_{B}(K)\) and,
- there may not exist a decomposition of \(f-\rho\) as in (1) inside \(B\) (i.e., the summands may not belong to \(B\)).
Thus, the optimum value just gives a lower bound for \(\rho\). But walking through a \(K\)-frame, as explained in [GIKM] constructs a net of lower bounds for \(\rho\) which approaches \(\rho\), eventually.
I practice, one only needs to find a sufficiently big finite dimensional linear space which contains \(f, g_1,\dots,g_m\) and a (1) decomposition of \(f-\rho\) can be found within that space. Therefore, the convergence happens in finitely many steps, subject to finding a suitable \(K\)-frame for the problem.
The significance of this method is that it converts any optimization problem into finitely many semidefinite programs whose optimum values approaches the optimum value of the original program and semidefinite programs can be solved in polynomial time. Although, this suggests that the NP-complete problem of optimization can be solved in P-time, but since the number of SDPs that is required to reach the optimum is unknown and such a bound does not exists when dealing with Archimedean modules.
Note
1. One behavior that distinguishes this method from others is that using SDP relaxations always provides a lower bound for theminimum value of the objective function over the feasibility set. While other methods usually involve evaluation of the objectiveand hence the result is always an upper bound for the minimum.
2. The SDP relaxation method relies on symbolic computations which could be quite costly and slow. Therefore, dealing with rather large problems -although Irene takes advantage from multiple cores- can be rather slow.
[GIKM] | (1, 2, 3) M. Ghasemi, M. Infusino, S. Kuhlmann and M. Marshall, Truncated Moment Problem for unital commutative real algebras, to appear. |
[JBL] | (1, 2) J-B. Lasserre, Global optimization with polynomials and the problem of moments, SIAM J. Optim. 11(3) 796-817 (2000). |
Polynomial Optimization¶
The SDP relaxation method was originally introduced by Lasserre [JBL] for polynomial optimization problem and excellent software packages such as GloptiPoly and ncpol2sdpa exist to handle constraint polynomial optimization problems.
Irene uses sympy for symbolic computations, so, it always need to be imported and the symbolic variables must be
introduced. Once these steps are done, the objective and constraints should be entered using SetObjective
and AddConstraint methods.
the method MomentsOrd
takes the relaxation degree upon user’s request, otherwise the minimum relaxation degree will be used.
The default SDP solver is CVXOPT
which can be modified via SetSDPSolver
method. Currently CVXOPT
, DSDP
, SDPA
and CSDP
are supported.
Next step is initialization of the SDP by InitSDP
and finally solving the SDP via Minimize
and the output will be stored in the Solution
variable as a python dictionary.
Example Solve the following polynomial optimization problem:
The following program uses relaxation of degree 3 and sdpa to solve the above problem:
from sympy import *
from Irene import *
# introduce variables
x = Symbol('x')
y = Symbol('y')
z = Symbol('z')
# initiate the Relaxation object
Rlx = SDPRelaxations([x, y, z])
# set the objective
Rlx.SetObjective(-2 * x + y - z)
# add support constraints
Rlx.AddConstraint(24 - 20 * x + 9 * y - 13 * z + 4 * x**2 -
4 * x * y + 4 * x * z + 2 * y**2 - 2 * y * z + 2 * z**2 >= 0)
Rlx.AddConstraint(x + y + z <= 4)
Rlx.AddConstraint(3 * y + z <= 6)
Rlx.AddConstraint(x >= 0)
Rlx.AddConstraint(x <= 2)
Rlx.AddConstraint(y >= 0)
Rlx.AddConstraint(z >= 0)
Rlx.AddConstraint(z <= 3)
# set the relaxation order
Rlx.MomentsOrd(3)
# set the solver
Rlx.SetSDPSolver('dsdp')
# initialize the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# output
print Rlx.Solution
The output looks like:
Solution of a Semidefinite Program:
Solver: DSDP
Status: Optimal
Initialization Time: 8.04711222649 seconds
Run Time: 1.056733 seconds
Primal Objective Value: -4.06848294478
Dual Objective Value: -4.06848289445
Feasible solution for moments of order 3
Moment Constraints¶
Initially the only constraints forced on the moments are those in (2). We can also force user defined constraints on the moments
by calling MomentConstraint
on a Mom
object. The following adds two constraints \(\int xy~d\mu\ge\frac{1}{2}\) and
\(\int yz~d\mu + \int z~d\mu\ge 1\) to the previous example:
from sympy import *
from Irene import *
# introduce variables
x = Symbol('x')
y = Symbol('y')
z = Symbol('z')
# initiate the Relaxation object
Rlx = SDPRelaxations([x, y, z])
# set the objective
Rlx.SetObjective(-2 * x + y - z)
# add support constraints
Rlx.AddConstraint(24 - 20 * x + 9 * y - 13 * z + 4 * x**2 -
4 * x * y + 4 * x * z + 2 * y**2 - 2 * y * z + 2 * z**2 >= 0)
Rlx.AddConstraint(x + y + z <= 4)
Rlx.AddConstraint(3 * y + z <= 6)
Rlx.AddConstraint(x >= 0)
Rlx.AddConstraint(x <= 2)
Rlx.AddConstraint(y >= 0)
Rlx.AddConstraint(z >= 0)
Rlx.AddConstraint(z <= 3)
# add moment constraints
Rlx.MomentConstraint(Mom(x * y) >= .5)
Rlx.MomentConstraint(Mom(y * z) + Mom(z) >= 1)
# set the relaxation order
Rlx.MomentsOrd(3)
# set the solver
Rlx.SetSDPSolver('dsdp')
# initialize the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# output
print Rlx.Solution
print "Moment of x*y:", Rlx.Solution[x * y]
print "Moment of y*z + z:", Rlx.Solution[y * z] + Rlx.Solution[z]
Solution is:
Solution of a Semidefinite Program:
Solver: DSDP
Status: Optimal
Initialization Time: 7.91646790504 seconds
Run Time: 1.041935 seconds
Primal Objective Value: -4.03644346623
Dual Objective Value: -4.03644340796
Feasible solution for moments of order 3
Moment of x*y: 0.500000001712
Moment of y*z + z: 2.72623169152
Equality Constraints¶
Although it is possible to add equality constraints via AddConstraint
and MomentConstraint
, but
SDPRelaxation converts them to two inequalities and considers a certain margin of error.
For \(A=B\), it considers \(A\ge B - \varepsilon\) and \(A\leq B + \varepsilon\).
In this case the value of \(\varepsilon\) can be modified by setting SDPRelaxation.ErrorTolerance
which its default value is \(10^{-6}\).
Truncated Moment Problem¶
It must be clear that we can use SDPRelaxations.MomentConstraint
to introduce a typical truncated
moment problem over polynomials as described in [JNie].
Example Find the support of a measure \(\mu\) whose support is a subset of \([-1,1]^2\) and the followings hold:
The following code does the job:
from sympy import *
from Irene import *
# introduce variables
x = Symbol('x')
y = Symbol('y')
# initiate the Relaxation object
Rlx = SDPRelaxations([x, y])
# add support constraints
Rlx.AddConstraint(1 - x**2 >= 0)
Rlx.AddConstraint(1 - y**2 >= 0)
# add moment constraints
Rlx.MomentConstraint(Mom(x**2) == 1. / 3.)
Rlx.MomentConstraint(Mom(y**2) == 1. / 3.)
Rlx.MomentConstraint(Mom(x**2 * y) == 0.)
Rlx.MomentConstraint(Mom(x * y**2) == 0.)
Rlx.MomentConstraint(Mom(x**2 * y**2) == 1. / 9.)
Rlx.MomentConstraint(Mom(x**4 * y**2) == 1. / 15.)
Rlx.MomentConstraint(Mom(x**2 * y**4) == 1. / 15.)
# set the solver
Rlx.SetSDPSolver('dsdp')
# initialize the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# output
Rlx.Solution.ExtractSolution('lh', 2)
print Rlx.Solution
and the result is:
Solution of a Semidefinite Program:
Solver: DSDP
Status: Optimal
Initialization Time: 1.08686900139 seconds
Run Time: 0.122459 seconds
Primal Objective Value: 0.0
Dual Objective Value: -9.36054051771e-09
Support:
(0.40181215311129925, 0.54947643681480196)
(-0.40181215311127805, -0.54947643681498193)
Support solver: Lasserre--Henrion
Feasible solution for moments of order 3
Note that the solution is not necessarily unique.
[JNie] | J. Nie, The A-Truncated K-Moment Problem, Found. Comput. Math., Vol.14(6), 1243-1276 (2014). |
Optimization of Rational Functions¶
Given two polynomials \(p(X), q(X), g_1(X),\dots,g_m(X)\), the minimum of \(\frac{p(X)}{q(X)}\) over \(K=\{x:g_i(x)\ge0,~i=1,\dots,m\}\) is equal to
Note that in this case \(\mu\) is not taken to be a probability measure, but instead \(\int q(X)~d\mu = 1\).
We can use SDPRelaxations.Probability = False
to relax the probability condition on \(\mu\) and use moment
constraints to enforce \(\int q(X)~d\mu = 1\). The following example explains this.
Example: Find the minimum of \(\frac{x^2-2x}{x^2+2x+1}\):
from sympy import *
from Irene import *
# define the symbolic variable
x = Symbol('x')
# initiate the SDPRelaxations object
Rlx = SDPRelaxations([x])
# settings
Rlx.Probability = False
# set the objective
Rlx.SetObjective(x**2 - 2*x)
# moment constraint
Rlx.MomentConstraint(Mom(x**2+2*x+1) == 1)
# set the sdp solver
Rlx.SetSDPSolver('cvxopt')
# initiate the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
print Rlx.Solution
The result is:
Solution of a Semidefinite Program:
Solver: CVXOPT
Status: Optimal
Initialization Time: 0.167912006378 seconds
Run Time: 0.008987 seconds
Primal Objective Value: -0.333333666913
Dual Objective Value: -0.333333667469
Feasible solution for moments of order 1
Note
Beside SDPRelaxations.Probability
there is another attribute SDPRelaxations.PSDMoment
which by default is set to True
and makes sure that the sdp solver assumes positivity for
the moment matrix.
Optimization over Varieties¶
Now we employ the results of [GIKM] to solve more complex optimization problems. The main idea is to represent the given function space as a quotient of a suitable polynomial algebra.
Suppose that we want to optimize the function \(\sqrt[3]{(xy)^2}-x+y^2\) over the closed disk with radius 3.
In order to deal with the term \(\sqrt[3]{(xy)^2}\), we introduce an algebraic relation to SDPRelaxations
object and give a
monomial order for Groebner basis computations (default is lex for lexicographic order).
Clearly \(xy-\sqrt[3]{(xy)}^3=0\). Therefore by introducing an auxiliary variable or function symbol, say \(f(x,y)\) the problem
can be stated in the quotient of \(\frac{\mathbb{R}[x,y,f]}{\langle xy-f^3\rangle}\). To check the result of SDPRelaxations
we
employ scipy.optimize.minimize
with two solvers COBYLA
and COBYLA
as well as two solvers, Augmented Lagrangian Particle Swarm
Optimizer and Non Sorting Genetic Algorithm II from pyOpt:
from sympy import *
from Irene import *
# introduce variables
x = Symbol('x')
y = Symbol('y')
f = Function('f')(x, y)
# define algebraic relations
rel = [x * y - f**3]
# initiate the Relaxation object
Rlx = SDPRelaxations([x, y, f], rel)
# set the monomial order
Rlx.SetMonoOrd('lex')
# set the objective
Rlx.SetObjective(f**2 - x + y**2)
# add support constraints
Rlx.AddConstraint(9 - x**2 - y**2 >= 0)
# set the solver
Rlx.SetSDPSolver('cvxopt')
# Rlx.MomentsOrd(2)
# initialize the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# output
print Rlx.Solution
# using scipy
from numpy import power
from scipy.optimize import minimize
fun = lambda x: power(x[0]**2 * x[1]**2, 1. / 3.) - x[0] + x[1]**2
cons = (
{'type': 'ineq', 'fun': lambda x: 9 - x[0]**2 - x[1]**2})
sol1 = minimize(fun, (0, 0), method='COBYLA', constraints=cons)
sol2 = minimize(fun, (0, 0), method='SLSQP', constraints=cons)
print "solution according to 'COBYLA'"
print sol1
print "solution according to 'SLSQP'"
print sol2
# pyOpt
from pyOpt import *
def objfunc(x):
from numpy import power
f = power(x[0]**2 * x[1]**2, 1. / 3.) - x[0] + x[1]**2
g = [x[0]**2 + x[1]**2 - 9]
fail = 0
return f, g, fail
opt_prob = Optimization('A third root function', objfunc)
opt_prob.addVar('x1', 'c', lower=-3, upper=3, value=0.0)
opt_prob.addVar('x2', 'c', lower=-3, upper=3, value=0.0)
opt_prob.addObj('f')
opt_prob.addCon('g1', 'i')
# Augmented Lagrangian Particle Swarm Optimizer
alpso = ALPSO()
alpso(opt_prob)
print opt_prob.solution(0)
# Non Sorting Genetic Algorithm II
nsg2 = NSGA2()
nsg2(opt_prob)
print opt_prob.solution(1)
The output will be:
Solution of a Semidefinite Program:
Solver: CVXOPT
Status: Optimal
Initialization Time: 0.12473487854 seconds
Run Time: 0.004865 seconds
Primal Objective Value: -2.99999997394
Dual Objective Value: -2.9999999473
Feasible solution for moments of order 1
solution according to 'COBYLA'
fun: -0.99788411120450926
maxcv: 0.0
message: 'Optimization terminated successfully.'
nfev: 25
status: 1
success: True
x: array([ 9.99969494e-01, 9.52333693e-05])
solution according to 'SLSQP'
fun: -2.9999975825413681
jac: array([ -0.99999923, 689.00398242, 0. ])
message: 'Optimization terminated successfully.'
nfev: 64
nit: 13
njev: 13
status: 0
success: True
x: array([ 3.00000000e+00, -1.25290367e-09])
ALPSO Solution to A third root function
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.1174
Total Function Evaluations: 1720
Lambda: [ 0.00023458]
Seed: 1482111093.38230896
Objectives:
Name Value Optimum
f -2.99915 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c 3.000000 -3.00e+00 3.00e+00
x2 c 0.000008 -3.00e+00 3.00e+00
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 i -1.00e+21 <= 0.000000 <= 0.00e+00
--------------------------------------------------------------------------------
NSGA-II Solution to A third root function
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.3833
Total Function Evaluations:
Objectives:
Name Value Optimum
f -2.99898 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c 3.000000 -3.00e+00 3.00e+00
x2 c -0.000011 -3.00e+00 3.00e+00
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 i -1.00e+21 <= -0.000000 <= 0.00e+00
--------------------------------------------------------------------------------
Optimization over arbitrary functions¶
Any given algebra can be represented as a quotient of a suitable polynomial algebra (on possibly infinitely many variables). Since optimization problems usually involve finitely many functions and constraints, we can apply the technique introduced in the previous section, as soon as we figure out the quotient representation of the function space.
Let us walk through the procedure by solving some examples.
Example 1. Find the optimum value of the following program:
Let us introduce four symbols to represent trigonometric functions:
Then the quotient algebra \(\frac{\mathbb{R}[f,g,h,k]}{I}\) where \(I=\langle f^2+g^2-1, h^2+k^2-1\rangle\) is the right framework to solve
the optimization problem. We also compare the outcome of SDPRelaxations
with scipy
and pyswarm
:
from sympy import *
from Irene import *
# introduce variables
x = Symbol('x')
f = Function('f')(x)
g = Function('g')(x)
h = Function('h')(x)
k = Function('k')(x)
# define algebraic relations
rels = [f**2 + g**2 - 1, h**2 + k**2 - 1]
# initiate the Relaxation object
Rlx = SDPRelaxations([f, g, h, k], rels)
# set the monomial order
Rlx.SetMonoOrd('lex')
# set the objective
Rlx.SetObjective(-(f - 1)**3 - (f - k)**4 - (k - 3)**2)
# add support constraints
Rlx.AddConstraint(10 - (f - 1)**2 >= 0)
Rlx.AddConstraint(10 - (f - k)**2 >= 0)
Rlx.AddConstraint(10 - (k - 3)**2 >= 0)
# set the solver
Rlx.SetSDPSolver('csdp')
# initialize the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# output
print Rlx.Solution
# using scipy
from scipy.optimize import minimize
fun = lambda x: -(sin(x[0]) - 1)**3 - (sin(x[0]) -
cos(x[1]))**4 - (cos(x[1]) - 3)**2
cons = (
{'type': 'ineq', 'fun': lambda x: 10 - (sin(x[0]) - 1)**2},
{'type': 'ineq', 'fun': lambda x: 10 - (sin(x[0]) - cos(x[1]))**2},
{'type': 'ineq', 'fun': lambda x: 10 - (cos(x[1]) - 3)**2})
sol1 = minimize(fun, (0, 0), method='COBYLA', constraints=cons)
sol2 = minimize(fun, (0, 0), method='SLSQP', constraints=cons)
print "solution according to 'COBYLA':"
print sol1
print "solution according to 'SLSQP':"
print sol2
# pyOpt
from pyOpt import *
def objfunc(x):
from numpy import sin, cos
f = -(sin(x[0]) - 1)**3 - (sin(x[0]) - cos(x[1]))**4 - (cos(x[1]) - 3)**2
g = [
(sin(x[0]) - 1)**2 - 10,
(sin(x[0]) - cos(x[1]))**2 - 10,
(cos(x[1]) - 3)**2 - 10
]
fail = 0
return f, g, fail
opt_prob = Optimization('A trigonometric function', objfunc)
opt_prob.addVar('x1', 'c', lower=-10, upper=10, value=0.0)
opt_prob.addVar('x2', 'c', lower=-10, upper=10, value=0.0)
opt_prob.addObj('f')
opt_prob.addCon('g1', 'i')
opt_prob.addCon('g2', 'i')
opt_prob.addCon('g3', 'i')
# Augmented Lagrangian Particle Swarm Optimizer
alpso = ALPSO()
alpso(opt_prob)
print opt_prob.solution(0)
# Non Sorting Genetic Algorithm II
nsg2 = NSGA2()
nsg2(opt_prob)
print opt_prob.solution(1)
Solutions are:
Solution of a Semidefinite Program:
Solver: CSDP
Status: Optimal
Initialization Time: 3.22915506363 seconds
Run Time: 0.016662 seconds
Primal Objective Value: -12.0
Dual Objective Value: -12.0
Feasible solution for moments of order 2
solution according to 'COBYLA':
fun: -11.824901993777621
maxcv: 1.7763568394002505e-15
message: 'Optimization terminated successfully.'
nfev: 42
status: 1
success: True
x: array([ 1.57064986, 1.7337948 ])
solution according to 'SLSQP':
fun: -11.9999999999720
jac: array([ -2.94446945e-05, -1.78813934e-05, 0.00000000e+00])
message: 'Optimization terminated successfully.'
nfev: 23
nit: 5
njev: 5
status: 0
success: True
x: array([ -1.57079782e+00, -6.42618794e-07])
ALPSO Solution to A trigonometric function
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.3503
Total Function Evaluations: 3640
Lambda: [ 0. 0. 2.0077542]
Seed: 1482111691.32805490
Objectives:
Name Value Optimum
f -11.8237 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c 7.854321 -1.00e+01 1.00e+01
x2 c 4.549489 -1.00e+01 1.00e+01
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 i -1.00e+21 <= -10.000000 <= 0.00e+00
g2 i -1.00e+21 <= -8.649336 <= 0.00e+00
g3 i -1.00e+21 <= -0.000612 <= 0.00e+00
--------------------------------------------------------------------------------
NSGA-II Solution to A trigonometric function
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.7216
Total Function Evaluations:
Objectives:
Name Value Optimum
f -12 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c -7.854036 -1.00e+01 1.00e+01
x2 c 0.000004 -1.00e+01 1.00e+01
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 i -1.00e+21 <= -6.000000 <= 0.00e+00
g2 i -1.00e+21 <= -6.000000 <= 0.00e+00
g3 i -1.00e+21 <= -6.000000 <= 0.00e+00
--------------------------------------------------------------------------------
SOS Decomposition¶
Let \(f_*\) be the result of SDPRelaxations.Minimize()
, then \(f-f_*\in Q_{\bf g}\).
Therefore, there exist \(\sigma_0,\sigma_1,\dots,\sigma_m\in \sum A^2\) such that
\(f-f_*=\sigma_0+\sum_{i=1}^m\sigma_i g_i\). Once the Minimize()
is called, the method
SDPRelaxations.Decompose()
returns this a dictionary of elements of \(A\) of the form
{0:[a(0, 1), ..., a(0, k_0)], ..., m:[a(m, 1), ..., a(m, k_m)}
such that
where \(g_0=1\).
Usually there are extra coefficients that are very small in absolute value as a result of round off error that should be ignored.
The following example shows how to employ this functionality:
from sympy import *
from Irene import SDPRelaxations
# define the symbolic variables and functions
x = Symbol('x')
y = Symbol('y')
z = Symbol('z')
Rlx = SDPRelaxations([x, y, z])
Rlx.SetObjective(x**3 + x**2 * y**2 + z**2 * x * y - x * z)
Rlx.AddConstraint(9 - (x**2 + y**2 + z**2) >= 0)
# initiate the SDP
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
print Rlx.Solution
# extract decomposition
V = Rlx.Decompose()
# test the decomposition
sos = 0
for v in V:
# for g0 = 1
if v == 0:
sos = expand(Rlx.ReduceExp(sum([p**2 for p in V[v]])))
# for g1, the constraint
else:
sos = expand(Rlx.ReduceExp(
sos + Rlx.Constraints[v - 1] * sum([p**2 for p in V[v]])))
sos = sos.subs(Rlx.RevSymDict)
pln = Poly(sos).as_dict()
pln = {ex:round(pln[ex],5) for ex in pln}
print Poly(pln, (x,y,z)).as_expr()
The output looks like this:
Solution of a Semidefinite Program:
Solver: CVXOPT
Status: Optimal
Initialization Time: 0.875229120255 seconds
Run Time: 0.031426 seconds
Primal Objective Value: -27.4974076889
Dual Objective Value: -27.4974076213
Feasible solution for moments of order 2
1.0*x**3 + 1.0*x**2*y**2 + 1.0*x*y*z**2 - 1.0*x*z + 27.49741
The Resume
method¶
It happens from time to time that one needs to stop the process of SDPRelaxations
to look into
its progress and/or run the code later. This has been accommodated thanks to python’s support for
serialization and error handling.
Since the initialization of the final SDP is the most time consuming part of the process, if one
breaks this via Ctrl-c, the object will save all the computation that has been done so far in
a .rlx file named with the name of the object. So, if one wants to resume the process later, it
suffices to call the Resume
method after instantiation and leave the program out and continue
the initialization via calling InitSDP
method.
The SDRelaxSol
¶
This object is a container for the solution of SDPRelaxation
objects.
It contains the following informations:
- Primal: the value of the SDP in primal form,
- Dual: the value of the SDP in dual form,
- RunTime: the run time of the sdp solver,
- InitTime: the total time consumed for initialization of the sdp,
- Solver: the name of sdp solver,
- Status: final status of the sdp solver,
- RelaxationOrd: order of relaxation,
- TruncatedMmntSeq: a dictionary of resulted moments,
- MomentMatrix: the resulted moment matrix,
- ScipySolver: the scipy solver to extract solutions,
- err_tol: the minimum value which is considered to be nonzero,
- Support: the support of discrete measure resulted from
SDPRelaxation.Minimize()
,- Weights: corresponding weights for the Dirac measures.
The SDRelaxSol
after initiation is an iterable object. The moments can be retrieved by
passing the index to the iterable SDRelaxSol[idx]
.
Extracting solutions¶
By default, the support of the measure is not calculated, but it can be approximated by calling
the method SDRelaxSol.ExtractSolution()
.
There exists an exact theoretical method for extracting the support of the solution measure as explained
in [HL]. But because of the numerical error of sdp solvers, computing rank and hence the support is quite
difficult. So, SDRelaxSol.ExtractSolution()
estimates the rank numerically by assuming that eigenvalues
with absolute value less than err_tol
which by default is set to SDPRelaxation.ErrorTolerance
.
Two methods are implemented for extracting solutions:
- Lasserre-Henrion method as explained in [HL]. To employ this method simply call
SDRelaxSol.ExtractSolution('LH', card)
, wherecard
is the maximum cardinality of the support.- Moment Matching method which employs
scipy.optimize.root
to approximate the support. The defaultscipy
solver is set to lm, but other solvers can be selected usingSDRelaxSol.SetScipySolver(solver)
. It is not guaranteed that scipy solvers return a reliable answer, but modifying sdp solvers and other parameters likeSDPRelaxation.ErrorTolerance
may help to get better results. To use this method callSDRelaxSol.ExtractSolution('scipy', card)
wherecard
is as above.
Example 1. Solve and find minimizers of \(x^2+y^2+z^4\) where \(x+y+z=4\):
from sympy import *
from Irene import *
x, y, z = symbols('x,y,z')
Rlx = SDPRelaxations([x, y, z])
Rlx.SetSDPSolver('cvxopt')
Rlx.SetObjective(x**2 + y**2 + z**4)
Rlx.AddConstraint(Eq(x + y + z, 4))
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# extract support
Rlx.Solution.ExtractSolution('LH', 1)
print Rlx.Solution
# pyOpt
from pyOpt import *
def objfunc(x):
f = x[0]**2 + x[1]**2 + x[2]**4
g = [x[0] + x[1] + x[2] - 4]
fail = 0
return f, g, fail
opt_prob = Optimization('Testing solutions', objfunc)
opt_prob.addVar('x1', 'c', lower=-4, upper=4, value=0.0)
opt_prob.addVar('x2', 'c', lower=-4, upper=4, value=0.0)
opt_prob.addVar('x3', 'c', lower=-4, upper=4, value=0.0)
opt_prob.addObj('f')
opt_prob.addCon('g1', 'e')
# Augmented Lagrangian Particle Swarm Optimizer
alpso = ALPSO()
alpso(opt_prob)
print opt_prob.solution(0)
The output is:
Solution of a Semidefinite Program:
Solver: CVXOPT
Status: Optimal
Initialization Time: 1.59334087372 seconds
Run Time: 0.021102 seconds
Primal Objective Value: 5.45953579912
Dual Objective Value: 5.45953586121
Support:
(0.91685039306810523, 1.541574317520042, 1.5415743175200163)
Support solver: Lasserre--Henrion
Feasible solution for moments of order 2
ALPSO Solution to Testing solutions
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.1443
Total Function Evaluations: 1720
Lambda: [-3.09182651]
Seed: 1482274189.55335808
Objectives:
Name Value Optimum
f 5.46051 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c 1.542371 -4.00e+00 4.00e+00
x2 c 1.541094 -4.00e+00 4.00e+00
x3 c 0.916848 -4.00e+00 4.00e+00
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 e 0.000314 = 0.00e+00
--------------------------------------------------------------------------------
Example 2. Minimize \(-(x-1)^2-(x-y)^2-(y-3)^2\) where \(1-(x-1)^2\ge0\), \(1-(x-y)^2\ge0\) and \(1-(y-3)^2\ge0\). It has three minimizers \((2, 3), (1, 2)\), and \((2, 2)\):
from sympy import *
from Irene import *
x, y = symbols('x, y')
Rlx = SDPRelaxations([x, y])
Rlx.SetSDPSolver('csdp')
Rlx.SetObjective(-(x - 1)**2 - (x - y)**2 - (y - 3)**2)
Rlx.AddConstraint(1 - (x - 1)**2 >= 0)
Rlx.AddConstraint(1 - (x - y)**2 >= 0)
Rlx.AddConstraint(1 - (y - 3)**2 >= 0)
Rlx.MomentsOrd(2)
Rlx.InitSDP()
# solve the SDP
Rlx.Minimize()
# extract support
Rlx.Solution.ExtractSolution('LH')
print Rlx.Solution
# pyOpt
from pyOpt import *
def objfunc(x):
f = -(x[0] - 1)**2 - (x[0] - x[1])**2 - (x[1] - 3)**2
g = [
(x[0] - 1)**2 - 1,
(x[0] - x[1])**2 - 1,
(x[1] - 3)**2 - 1
]
fail = 0
return f, g, fail
opt_prob = Optimization("Lasserre's Example", objfunc)
opt_prob.addVar('x1', 'c', lower=-3, upper=3, value=0.0)
opt_prob.addVar('x2', 'c', lower=-3, upper=3, value=0.0)
opt_prob.addObj('f')
opt_prob.addCon('g1', 'i')
opt_prob.addCon('g2', 'i')
opt_prob.addCon('g3', 'i')
# Augmented Lagrangian Particle Swarm Optimizer
alpso = ALPSO()
alpso(opt_prob)
print opt_prob.solution(0)
# Non Sorting Genetic Algorithm II
nsg2 = NSGA2()
nsg2(opt_prob)
print opt_prob.solution(1)
which results in:
Solution of a Semidefinite Program:
Solver: CSDP
Status: Optimal
Initialization Time: 0.861004114151 seconds
Run Time: 0.00645 seconds
Primal Objective Value: -2.0
Dual Objective Value: -2.0
Support:
(2.000000006497352, 3.000000045123556)
(0.99999993829586131, 1.9999999487412694)
(1.9999999970209055, 1.9999999029899564)
Support solver: Lasserre--Henrion
Feasible solution for moments of order 2
ALPSO Solution to Lasserre's Example
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.1353
Total Function Evaluations: 1720
Lambda: [ 0.08278879 0.08220848 0. ]
Seed: 1482307696.27431393
Objectives:
Name Value Optimum
f -2 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c 1.999967 -3.00e+00 3.00e+00
x2 c 3.000000 -3.00e+00 3.00e+00
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 i -1.00e+21 <= -0.000065 <= 0.00e+00
g2 i -1.00e+21 <= 0.000065 <= 0.00e+00
g3 i -1.00e+21 <= -1.000000 <= 0.00e+00
--------------------------------------------------------------------------------
NSGA-II Solution to Lasserre's Example
================================================================================
Objective Function: objfunc
Solution:
--------------------------------------------------------------------------------
Total Time: 0.2406
Total Function Evaluations:
Objectives:
Name Value Optimum
f -1.99941 0
Variables (c - continuous, i - integer, d - discrete):
Name Type Value Lower Bound Upper Bound
x1 c 1.999947 -3.00e+00 3.00e+00
x2 c 2.000243 -3.00e+00 3.00e+00
Constraints (i - inequality, e - equality):
Name Type Bounds
g1 i -1.00e+21 <= -0.000106 <= 0.00e+00
g2 i -1.00e+21 <= -1.000000 <= 0.00e+00
g3 i -1.00e+21 <= -0.000486 <= 0.00e+00
--------------------------------------------------------------------------------
Irene detects all minimizers correctly, but each pyOpt solvers only detect one. Note that we did not specify number of solutions, but the solver extracted them all.
[HL] | (1, 2) D. Henrion and J-B. Lasserre, Detecting Global Optimality and Extracting Solutions in GloptiPoly, Positive Polynomials in Control, LNCIS 312, 293-310 (2005). |