Exponential transformation: Difference between revisions

From Cornell University Computational Optimization Open Textbook - Optimization Wiki
Jump to navigation Jump to search
No edit summary
m (Update section link)
 
(34 intermediate revisions by 2 users not shown)
Line 1: Line 1:
Author: Daphne Duvivier, Daniela Gil, Jacqueline Jackson, Sinclaire Mills, Vanessa Nobre (SYSEN 5800, Fall 2021)
Author: Daphne Duvivier, Daniela Gil, Jacqueline Jackson, Sinclaire Mills, Vanessa Nobre (SYSEN 5800, Fall 2021)


== Introduction ==
== Introduction ==


Exponential transformations are simple algebraic transformation of monomial functions through a variable substitution with an exponential variable.  
An '''exponential transformation''' is a simple algebraic transformation of a monomial function through variable substitution with an exponential variable. In computational optimization, exponential transformations are used for convexification of geometric programming constraints in nonconvex optimization problems.
 
Information regarding the algebraic properties of exponential transformation
 
In computational optimization exponential transformations are used for convexification of geometric programming constraints (posynominal) nonconvex optimization problems. This transformation creates a convex function without changing the decision space of the problem. <ref> Li, D., Biswal, M.P. Exponential Transformation in Convexifying a Noninferior Frontier and Exponential Generating Method. Journal of Optimization Theory and Applications 99, 183–199 (1998). https://doi.org/10.1023/A:1021708412776 </ref> Through the use of exponential transformation the time to solve an Non-linear program (NLP) or a Mixed integer non-linear program (MINLP) is reduced by allowing the use of a global solve.
 
 


Exponential transformation creates a convex function without changing the decision space of the problem. <ref name =":0"> D. Li and M. P. Biswal, [https://doi.org/10.1023/A:1021708412776 "Exponential Transformation in Convexifying a Noninferior Frontier and Exponential Generating Method]," ''Journal of Optimization Theory and Applications'', vol. 99, pp. 183–199, 1998.</ref> This is done through a simple substitution of continuous variables with a natural exponent and simplification of binary variables through removal of the exponent. The transformation is verified to be convex if the Hessian, denoted by <math>H(x)</math>, is proven to be positive-definite.


By using exponential transformations, not only is the overall time to solve a Nonlinear Programming (NLP) or a Mixed Integer Nonlinear Programming (MINLP) problem reduced, but we can also simplify the solution space enough to utilize conventional NLP/MINLP solvers. Various real world optimization problems apply this transformation to simplify the solution space consisting of extensive quantities of constraints and variables.


== Theory & Methodology ==
== Theory, Methodology, and Algorithmic Discussions ==
=== Theory ===
Exponential transformations are most commonly applied to geometric programs. A '''geometric program''' is a mathematical optimization problem where the objective function is a posynomial being minimized. '''Posynomial''' functions are defined as positive polynomials. <ref name =":1"> S. Boyd, S. J. Kim, and L. Vandenberghe ''et al.'', [https://doi.org/10.1007/s11081-007-9001-7 "A tutorial on geometric programming]," ''Optimization and Engineering'', vol. 8, article 67, 2007.</ref> 


Exponential transformation can be applied to geometric programs
The standard form of a geometric program is represented by:


A geometric program is a mathematical optimization problem  
<math>\begin{align}
\min & \quad f_0(x) \\
s.t. & \quad f_i(x) \leq 1 \quad i = 1,....,m  \\
& \quad g_i(x) = 1  \quad i = 1,....,p
\end{align}</math>


Geometric Programs take the form of:
Where <math> f_0(x) </math> is a posynomial function, <math> f_i(x) </math>  is a posynomial function and <math> g_i(x) </math> is a monomial function.


<math> min f_0(x) </math>
To verify a geometric program is represented in its standard form, the following conditions must be true:
<math> s.t. f_i(x) \leq 1  i = 1,....,m </math>
# Objective function <math>f_0(x)</math> is a posynomial.
<math> s.t. g_i(x) = 1  i = 1,....,p </math>
# Inequality constraints <math>f_i(x)</math> must be posynomials less than or equal to 1.
# Equality constraints <math>g_i(x)</math> must be monomials equal to 1.


convexification of a non-inferior frontier - based on having a differentiable objective function  https://link.springer.com/content/pdf/10.1023/A:1021708412776.pdf
In this definition, monomials differ from the usual algebraic definition where the exponents must be nonnegative integers. For this application, exponents can be any positive number inclusive of fractions and negative exponents. <ref name =":1"></ref>


this would apply to any polynomial.
=== Methodology ===
A posynomial is defined as define polynomials:


Exponential transformation begins with a posynominal (Positive and Polynomial) noncovex function of the form <ref> Boyd, S., Kim, SJ., Vandenberghe, L. et al. A tutorial on geometric programming. Optim Eng 8, 67 (2007). https://doi.org/10.1007/s11081-007-9001-7 </ref> :
Exponential transformation begins with a posynomial noncovex function as depicted below. <ref name =":1"></ref> A posynomial begins with <math> x_1,...,x_n </math> where <math>x_n</math> are real non-negative variables.


<math> f(x) = \sum_{k=1}^N c_k{x_1}^{a_{1k}}{x_2}^{a_{2k}}....{x_n}^{a_{nk}} </math>
<math> f(x) = \sum_{k=1}^N c_k{x_1}^{a_{1k}}{x_2}^{a_{2k}}....{x_n}^{a_{nk}} </math>


where <math> c_k \geq 0 </math> and <math> x_n \geq0 </math>
Where <math> c_k \geq 0 </math> and <math> x_n \geq0 </math>
 
A transformation of <math> x_n = e^u_i </math> is applied  <ref> Grossmann, I.E. Review of Nonlinear Mixed-Integer and Disjunctive Programming Techniques. Optimization and Engineering 3, 227–252 (2002). https://doi.org/10.1023/A:1021039126272 </ref>


The transformed function is presented as:
A transformation is applied in which <math> x_n </math> is replaced with the natural logarithm base exponential <math> e^{u_n} </math>. <ref> I. E. Grossmann, [https://doi.org/10.1023/A:1021039126272 "Review of Nonlinear Mixed-Integer and Disjunctive Programming Techniques]," ''Optimization and Engineering'', vol. 3, pp. 227–252, 2002. </ref> The transformed function after substitution is presented as:


<math> f(u) = \sum_{k=1}^N c_k{{e}^{{u_1}{a_{1k}}}{e}^{{u_2}{a_{2k}}}....{e}^{{u_n}{a_{nk}}}} </math>
<math> f(u) = \sum_{k=1}^N c_k{{e}^{{u_1}{a_{1k}}}{e}^{{u_2}{a_{2k}}}....{e}^{{u_n}{a_{nk}}}} </math>


Properties of the exponent can be used to further simplify the transformation above, resulting in the sum of the exponents with a natural logarithm base.


===Exponential transformation in Computational Optimization===
<math> f(u) = \sum_{k=1}^N c_k{{e}^{{{u_1}{a_{1k}}}+{{u_2}{a_{2k}}}....+{{u_n}{a_{nk}}}}} </math>
Exponential transformation can be used for convexification of any Geometric MINLP that meet the criteria of equation 1. This is done by turning the problem into a nonlinear convex optimization problem through exponential transformation. Using the exponential substitution detailed above all continuous variables in the function are transformed while binary variables are not transformed.  


Additionally as presented in Theorem 1 and accompanying proof in "Global optimization of signomial geometric programming using linear relaxation" by P. Shen, K. Zhang, given that a function is being minimized it shows that after transformation all points on the transformed function are feasible in the original function and all objective values in the transformed function are the same or less than the original function. <ref> Boyd, S., Kim, SJ., Vandenberghe, L. et al. A tutorial on geometric programming. Optim Eng 8, 67 (2007). https://doi.org/10.1007/s11081-007-9001-7 </ref> Also presented by Li and Biswal the bounds of the problem are not altered through exponential transformation. <ref> Li, D., Biswal, M.P. Exponential Transformation in Convexifying a Noninferior Frontier and Exponential Generating Method. Journal of Optimization Theory and Applications 99, 183–199 (1998). https://doi.org/10.1023/A:1021708412776 </ref>
In order to prove the convexity of the final transformed function, the positive-definite test of the Hessian is used as defined in ''Optimization of Chemical Processes'' <ref> T. F. Edgar, D. M. Himmelblau, and L. S. Lasdon, ''Optimization of Chemical Processes'', McGraw-Hill, 2001.</ref>. The Hessian is defined as the following:


<math>  H(x) = H = \nabla^2f(x) =
\begin{bmatrix}
{\partial^2 f \over \partial {x_1^2}} & {\partial^2 f \over \partial x_1\partial x_2} \\
{\partial^2 f \over \partial x_2\partial x_1} & {\partial^2 f \over \partial {x_2^2}}
\end{bmatrix}
</math>


===Simple Numerical Example ===
to test that
<math> {\frac{x_1^3}{x_2^4}} + {x_1^2} + {\sqrt[3]{x_2^2}} \leq 4</math>


Reformulating to exponents
<math>  Q(x)\geq 0 </math> where <math>  Q(x) = x^THx </math> for all <math> x \neq 0</math>.


<math> {x_1^3}*{x_2^{-4}} + {x_1^2} + {x_2^{\frac{2}{3}}} \leq 4 </math>
===Exponential Transformation in Computational Optimization===


Substituting <math> x_1 = e^{u_1}, x_2 = e^{u_2} </math>
Exponential transformation can be used for convexification of any MINLP that meets the criteria for a geometric program. Using the exponential substitution detailed above, all continuous variables in the function are transformed while binary variables are not transformed. This transformation can also be applied to constraints to ensure convexification throughout the entire problem. 


<math>{{e^{u_1}}^3}*{{e^{u_2}}^{-4}} + {{e^{u_1}}^2} + {{e^{u_2}}^{\frac{2}{3}}} \leq 4 </math>
In a special case, binary exponential transformation can also be applied where binary variables are linearized. Binary exponential transformation can be done by using the following replacement: <math> {y^n} </math> is substituted by <math>y</math>. 


Simplifying by exponent properties
Additionally, all points on the transformed function are feasible in the original function, and all objective values in the transformed function are the same or less than the original function. <ref>P. Shen and K. Zhang, [https://doi.org/10.1016/S0096-3003(03)00200-5 "Global optimization of signomial geometric programming using linear relaxation]," ''Applied Mathematics and Computation'', vol. 150, issue 1, pp. 99-114, 2004. </ref> This creates a convex under estimator approach to the problem. Note that the bounds of the problem are not altered through exponential transformation. <ref name=":0"></ref>


<math> {e^{3{u_1} - 4{u_2}}} + {e^{{2}{u_1}}} + {{e^{{\frac{2}{3}}{u_2}}}} \leq 4 </math>
== Numerical Example ==
To provide an example, we begin with a simple nonconvex problem:


===Example of Convexification application in MINLP ===
<math> {\frac{x_1^3}{x_2^4}} + {x_1^2} + {\sqrt[3]{x_2^2}} \leq 4</math>


The following MINLP problem can take a Covexification approach using exponential transformation:


min <math> Z = 5{x_1^2}{x_2^8} + 2{x_1} + \frac{x_2^3} + 5{y_1} + 2 {y_2^2} </math>
'''Step 1:''' Convert the problem into standard form by reformulating radicals, fractions, etc. into exponents.


s.t.
<math> {x_1^3}*{x_2^{-4}} + {x_1^2} + {x_2^{\frac{2}{3}}} \leq 4 </math>


<math> {x_1} \leq 7{x_2^{0.2}} </math>
'''Step 2:''' Substitute all instances of <math>x_n</math> with <math>e^{u_n}</math>.


<math> 2{x_1^3} - y_1^2 \leq 1 </math>
<math>{e^{{3}{u_1}}}*{e^{{-4}{u_2}}} + {e^{{2}{u_1}}} + {{e^{{\frac{2}{3}}u_2}}} \leq 4 </math>


<math> x_1 \geq 0 </math>
'''Step 3:''' Simplify by applying exponent properties.


<math> x_2 \leq 4 </math>
<math> {e^{3{u_1} - 4{u_2}}} + {e^{{2}{u_1}}} + {{e^{{\frac{2}{3}}{u_2}}}} \leq 4 </math>
 
<math> y_1 = 0,1 </math> <math> y_2 = 0,1 </math>
 
Using the exponential transformation to continuous variables <math>x_1, x_2, x_3 </math>  by substituting <math> x_1 = e^{u_1} and x_2 = e^{u_2} </math> described the problem becomes the following:


min <math> Z = 5{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} + {\frac{e^{{u_2}{3}}}} + 5{y_1} + 2 {y_2^2} </math>
===Example of Convexification in MINLP ===


s.t
The following MINLP problem can take a convexification approach using exponential transformation.


<math> {e^{u_1}} \leq 7{e^{0.2{u_2}}} </math>
<math>\begin{align}
\min & \quad Z = 5{x_1^2}{x_2^8} + 2{x_1} + {x_2^3} + 5{y_1} + 2 {y_2^2} \\
s.t. & \quad {x_1} \leq 7{x_2^{0.2}} \\
& \quad 2{x_1^3} - y_1^2 \leq 1 \\
& \quad x_1 \geq 0 \\
& \quad x_2 \leq 4 \\
& \quad y_1 \isin \left \{ 0,1 \right \} \quad y_2 \isin \left \{ 0,1 \right \}
\end{align} </math>  


<math> 2{e^{3{u_1}}} - y_1^2 \leq 1 </math>


<math> e^{u_1} \geq 0 </math>
'''Step 1:''' Apply the exponential transformation to continuous variables <math>x_1</math> and <math>x_2</math> by substituting <math>x_1 = e^{u_1}</math> and <math>x_2 = e^{u_2} </math>.


<math> e^{u_2} \leq 4 </math>
<math>\begin{align}
\min & \quad Z = 5{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} + {e^{{3}{u_2}}} + 5{y_1} + 2 {y_2^2} \\
s.t. & \quad {e^{u_1}} \leq 7{e^{0.2{u_2}}} \\
& \quad 2{e^{3{u_1}}} - y_1^2 \leq 1 \\
& \quad e^{u_1} \geq 0 \\
& \quad e^{u_2} \leq 4 \\
& \quad y_1 \isin \left \{ 0,1 \right \} \quad y_2 \isin \left \{ 0,1 \right \}
\end{align}</math>


<math> y_1 = 0,1 </math> <math> y_2 = 0,1 </math>
'''Step 2:''' Simplify using the properties of exponents. (i.e. Combining the products of exponential terms as the sum of exponents with the same base)


With additional logarithmic simplification through properties of natural logarithm:
<math>\begin{align}
 
\min & \quad Z = 5{e^{2{u_1}+{8{u_2}}}} + 2{e^{u_1}}+{e^{{2u_2}+{3}{u_2}}} + 5{y_1} + 2 {y_2^2} \\
min <math> Z = 5{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} + {\frac{e^{{u_2}{3}}}} + 5{y_1} + 2 {y_2^2} </math>
s.t. & \quad u_1 \leq \ln 7 + 0.2{u_2} \\
 
& \quad 2{e^{3{u_1}}} - y_1^2 \leq 1 \\
s.t
& \quad {u_2} \leq \ln 4 \\
 
& \quad y_1 \isin \left \{ 0,1 \right \} \quad y_2 \isin \left \{ 0,1 \right \}
<math> u_1 \leq \ln 7 + 0.2{u_2} </math>
\end{align}</math>
 
<math> 2{e^{3{u_1}}} - y_1^2 \leq 1 </math>
 
<math> {u_2} \leq \ln 4 </math>
 
<math> y_1 = 0,1 </math> <math> y_2 = 0,1 </math>  


Where <math> u_1 </math> is unbounded due to logarithmic of 0 being indefinite.
Where <math> u_1 </math> is unbounded due to logarithmic of 0 being indefinite.


The transformed objective function can be show to be convex through the positive-definite test of the Hessian, for the example above the Hessian is as follows <ref> Chiang, Mung. (2005). Geometric Programming for Communication Systems. 10.1561/9781933019574; https://www.princeton.edu/~chiangm/gp.pdf  </ref>:
'''Step 3:''' Simplify binary variables by substituting <math>{y_1}^2</math> with <math>{y_1}</math> and <math>{y_2}^2</math> with <math>{y_2} </math>.


In order to prove the convexity of the transformed functions the positive definite test of Hessian is used as defined in "Optimization of Chemical Processes" <ref> T.F. Edgar, D.M. Himmelblau, L.S. Lasdon, Optimization of Chemical Processes.  McGraw-Hill, 2001.</ref> can be used. This tests the Hessian defined as:
<math>\begin{align}
\min & \quad Z = 5{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} + {e^{{u_2}{3}}} + 5{y_1} + 2 {y_2}  \\
s.t. & \quad u_1 \leq \ln 7 + 0.2{u_2} \\
& \quad 2{e^{3{u_1}}} - y_1 \leq 1 \\
& \quad {u_2} \leq \ln 4 \\
& \quad y_1 \isin \left \{ 0,1 \right \} \quad y_2 \isin \left \{ 0,1 \right \}
\end{align}</math>


<math> H(x) = H = \nabla^2f(x)</math>
==== Convexity Check ====
For the example above, the Hessian of the transformed matrix is as follows: <ref> M. Chiang, [https://www.princeton.edu/~chiangm/gp.pdf "Geometric Programming for Communication Systems]," 2005. </ref>


to test that
<math>\begin{bmatrix}
 
{\partial^2 Z(u) \over \partial u_1^2}  &  {\partial^2 Z(u) \over \partial u_1\partial u_2} \\
<math>  Q(x)\geq 0 </math>
{\partial^2 Z(u) \over \partial u_2\partial u_1} & {\partial^2 Z(u) \over \partial u_2^2} \\
 
\end{bmatrix}</math>
where


<math>  Q(x) = x^THx </math>


for all <math> x \neq 0 </math>
'''Step 1:''' Solve for each partial derivative.


for functions the Hessian is defined as:
<math>\begin{align}
{\partial Z(u) \over \partial u_1} = 10{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} \qquad & {\partial Z(u) \over \partial u_2} = 40{e^{2{u_1}}}{e^{8{u_2}}} + 4{e^{u_1}}{e^{2u_2}} + 3{e^{3u_2}} \\
{\partial^2 Z(u) \over \partial u_1^2} = 20{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} \qquad & {\partial^2 Z(u) \over \partial u_2^2} = 320{e^{2{u_1}}}{e^{8{u_2}}} + 8{e^{u_1}}{e^{2u_2}} + 9{e^{3u_2}} \\
{\partial^2 Z(u) \over \partial u_1\partial u_2} = 80{e^{2{u_1}}}{e^{8{u_2}}} + 4{e^{u_1}}{e^{2u_2}} \qquad & {\partial^2 Z(u) \over \partial u_2\partial u_1} = 80{e^{2{u_1}}}{e^{8{u_2}}} + 4{e^{u_1}}{e^{2u_2}}
\end{align}</math>


<math>
\begin{bmatrix}
X & X \\
X & Y
\end{bmatrix}
</math>


'''Step 2:''' Construct Hessian matrix, <math>H(x)</math>, from second derivatives.


In the example above the hessian is defined as:
<math>H(x) =
\begin{bmatrix}
20{e^{2{u_1}}}{e^{8{u_2}}} + 2{e^{u_1}}{e^{2u_2}} \qquad & 80{e^{2{u_1}}}{e^{8{u_2}}} + 4{e^{u_1}}{e^{2u_2}} \\
80{e^{2{u_1}}}{e^{8{u_2}}} + 4{e^{u_1}}{e^{2u_2}} \qquad & 320{e^{2{u_1}}}{e^{8{u_2}}} + 8{e^{u_1}}{e^{2u_2}} + 9{e^{3u_2}}
\end{bmatrix}</math>


<math>
\begin{bmatrix}
X & X \\
X & Y
\end{bmatrix}
</math>


Therefore H(x) is positive-definite and strictly convex.  
Because the second derivatives consist only of exponential equations and the exponential expression <math>e^x</math> is convex everywhere, <math>H(x)</math> is proven to be positive-definite and strictly convex.  


==Applications==
==Applications==


Currently various applications of exponential transformation can be seen in published journal articles and industry practices, due to the closeness with logarithmic transformation usually a combination of the approaches are used in practical solutions.   
Currently, various applications of exponential transformation can be seen in several published journal articles and industry practices. Many of these applications use exponential transformation to convexify their problem space. Due to the similarities between exponential and logarithmic transformations, a combination of both approaches is typically used in practical solutions.   


Global optimization of heat exchanger network synthesis problems with and without the isothermal mixing assumption:
=== Mechanical Engineering Applications ===
*As seen in eq(34) and (35) of the work by Björk and Westerlund they employ an exponential transformation to convexify their optimization problem to employ a global optimization approach. <ref>{{cite journal |title=Global optimization of heat exchanger network synthesis problems with and without the isothermal mixing assumption |journal=Computers & Chemical Engineering |year=2002 |last=Björk |first=Kaj-Mikael |last2=Westerlund |first2=Tapio |volume=26 |issue=11 |pages=1581-1593 |issn=ISSN 0098-1354 |doi=10.1016/S0098-1354(02)00129-1 |url=https://www.sciencedirect.com/science/article/pii/S0098135402001291?casa_token=G7OVOrBKagoAAAAA:1hUCHkGascVlawR3OfBpolNXlFqPSBUhWL6MkVAhn-ofKVfF-CbhVK6ZfSCKQ7i6mRQ9MTaqf9Q |accessdate=2021-11-27 }}</ref>
A global optimization approach is explored for the synthesis of heat exchanger networks. As seen in equations (34) and (35) of the paper by Björk and Westerlund, they employ an exponential transformation to convexify their optimization problem. <ref>K. J. Björk and T. Westerlund, [https://doi.org/10.1016/S0098-1354(02)00129-1 "Global optimization of heat exchanger network synthesis problems with and without the isothermal mixing assumption]," ''Computers & Chemical Engineering'', vol. 26, issue 11, pp. 1581-1593, 2002.</ref>
=== Electrical Engineering Application ===
In the optimization of VLSI circuit performance, a special geometric program defined as a '''unary geometric program''' is presented. The unary geometric program is a posynomial as defined in the [[#Theory, Methodology, and Algorithmic Discussions|Theory, Methodology, and Algorithmic Discussions]] section. The unary geometric program is derived through a greedy algorithm which implements a logarithmic transformation within lemma 5. <ref>C. Chu and D. F. Wong, [http://home.eng.iastate.edu/~cnchu/pubs/j08.pdf "VLSI Circuit Performance Optimization by Geometric Programming]," ''Annals of Operations Research'', vol. 105, pp. 37-60, 2001.</ref> While this is not a specific exponential transformation example, logarithmic transformations are within the same family and can also be used to convexify geometric programs.
=== Machining Economics ===
Applications in economics can be seen through geometric programming approaches. Examples and applications include analyzing the life of cutting tools in machining. In this approach, exponential transformations are used to convexify the problem. <ref>T. R. Jefferson and C. H. Scott, [https://link.springer.com/article/10.1007%2FBF02591746 "Quadratic geometric programming with application to machining economics]," ''Mathematical Programming'', vol. 31, pp. 137-152, 1985.</ref>


 
Overall, exponential transformations can be applied anywhere a geometric programming approach is taken to optimize the solution space. Some applications may perform a logarithmic transformation instead of an exponential transformation.  
Electrical Engineering Application: http://home.eng.iastate.edu/~cnchu/pubs/j08.pdf
 
 
Quadratic Geometric Programming
Ecconomics: https://link.springer.com/content/pdf/10.1007/BF02591746.pdf


==Conclusion==
==Conclusion==
Exponential transformation is a useful method to convexify Geometric MINLP and obtain a global solution to the problem. Exponential transformation does not alter the bounds of the problem and allows for a convex objective function and constraints given that the conditions described under the Theory and methodology section are satisfied. (Add jump metric) Geometric Programming transformation can be further explored through logarithmic transformation to address convexification.
Exponential transformation is a useful method to convexify geometric MINLP and obtain a global solution to the problem. Exponential transformation does not alter the bounds of the problem and allows for a convex objective function and constraints given that the prerequisite conditions described are satisfied. Geometric programming transformation can be further explored through logarithmic transformation to address convexification.


==References==
==References==
<references />
<references />
{{reflist}}

Latest revision as of 23:06, 14 December 2021

Author: Daphne Duvivier, Daniela Gil, Jacqueline Jackson, Sinclaire Mills, Vanessa Nobre (SYSEN 5800, Fall 2021)

Introduction

An exponential transformation is a simple algebraic transformation of a monomial function through variable substitution with an exponential variable. In computational optimization, exponential transformations are used for convexification of geometric programming constraints in nonconvex optimization problems.

Exponential transformation creates a convex function without changing the decision space of the problem. [1] This is done through a simple substitution of continuous variables with a natural exponent and simplification of binary variables through removal of the exponent. The transformation is verified to be convex if the Hessian, denoted by , is proven to be positive-definite.

By using exponential transformations, not only is the overall time to solve a Nonlinear Programming (NLP) or a Mixed Integer Nonlinear Programming (MINLP) problem reduced, but we can also simplify the solution space enough to utilize conventional NLP/MINLP solvers. Various real world optimization problems apply this transformation to simplify the solution space consisting of extensive quantities of constraints and variables.

Theory, Methodology, and Algorithmic Discussions

Theory

Exponential transformations are most commonly applied to geometric programs. A geometric program is a mathematical optimization problem where the objective function is a posynomial being minimized. Posynomial functions are defined as positive polynomials. [2]

The standard form of a geometric program is represented by:

Where is a posynomial function, is a posynomial function and is a monomial function.

To verify a geometric program is represented in its standard form, the following conditions must be true:

  1. Objective function is a posynomial.
  2. Inequality constraints must be posynomials less than or equal to 1.
  3. Equality constraints must be monomials equal to 1.

In this definition, monomials differ from the usual algebraic definition where the exponents must be nonnegative integers. For this application, exponents can be any positive number inclusive of fractions and negative exponents. [2]

Methodology

Exponential transformation begins with a posynomial noncovex function as depicted below. [2] A posynomial begins with where  are real non-negative variables.

Where and

A transformation is applied in which is replaced with the natural logarithm base exponential . [3] The transformed function after substitution is presented as:

Properties of the exponent can be used to further simplify the transformation above, resulting in the sum of the exponents with a natural logarithm base.

In order to prove the convexity of the final transformed function, the positive-definite test of the Hessian is used as defined in Optimization of Chemical Processes [4]. The Hessian is defined as the following:

to test that

where for all .

Exponential Transformation in Computational Optimization

Exponential transformation can be used for convexification of any MINLP that meets the criteria for a geometric program. Using the exponential substitution detailed above, all continuous variables in the function are transformed while binary variables are not transformed. This transformation can also be applied to constraints to ensure convexification throughout the entire problem.

In a special case, binary exponential transformation can also be applied where binary variables are linearized. Binary exponential transformation can be done by using the following replacement: is substituted by .

Additionally, all points on the transformed function are feasible in the original function, and all objective values in the transformed function are the same or less than the original function. [5] This creates a convex under estimator approach to the problem. Note that the bounds of the problem are not altered through exponential transformation. [1]

Numerical Example

To provide an example, we begin with a simple nonconvex problem:


Step 1: Convert the problem into standard form by reformulating radicals, fractions, etc. into exponents.

Step 2: Substitute all instances of with .

Step 3: Simplify by applying exponent properties.

Example of Convexification in MINLP

The following MINLP problem can take a convexification approach using exponential transformation.


Step 1: Apply the exponential transformation to continuous variables and by substituting and .

Step 2: Simplify using the properties of exponents. (i.e. Combining the products of exponential terms as the sum of exponents with the same base)

Where is unbounded due to logarithmic of 0 being indefinite.

Step 3: Simplify binary variables by substituting with and with .

Convexity Check

For the example above, the Hessian of the transformed matrix is as follows: [6]


Step 1: Solve for each partial derivative.


Step 2: Construct Hessian matrix, , from second derivatives.


Because the second derivatives consist only of exponential equations and the exponential expression is convex everywhere, is proven to be positive-definite and strictly convex.

Applications

Currently, various applications of exponential transformation can be seen in several published journal articles and industry practices. Many of these applications use exponential transformation to convexify their problem space. Due to the similarities between exponential and logarithmic transformations, a combination of both approaches is typically used in practical solutions.

Mechanical Engineering Applications

A global optimization approach is explored for the synthesis of heat exchanger networks. As seen in equations (34) and (35) of the paper by Björk and Westerlund, they employ an exponential transformation to convexify their optimization problem. [7]

Electrical Engineering Application

In the optimization of VLSI circuit performance, a special geometric program defined as a unary geometric program is presented. The unary geometric program is a posynomial as defined in the Theory, Methodology, and Algorithmic Discussions section. The unary geometric program is derived through a greedy algorithm which implements a logarithmic transformation within lemma 5. [8] While this is not a specific exponential transformation example, logarithmic transformations are within the same family and can also be used to convexify geometric programs.

Machining Economics

Applications in economics can be seen through geometric programming approaches. Examples and applications include analyzing the life of cutting tools in machining. In this approach, exponential transformations are used to convexify the problem. [9]

Overall, exponential transformations can be applied anywhere a geometric programming approach is taken to optimize the solution space. Some applications may perform a logarithmic transformation instead of an exponential transformation.

Conclusion

Exponential transformation is a useful method to convexify geometric MINLP and obtain a global solution to the problem. Exponential transformation does not alter the bounds of the problem and allows for a convex objective function and constraints given that the prerequisite conditions described are satisfied. Geometric programming transformation can be further explored through logarithmic transformation to address convexification.

References

  1. 1.0 1.1 D. Li and M. P. Biswal, "Exponential Transformation in Convexifying a Noninferior Frontier and Exponential Generating Method," Journal of Optimization Theory and Applications, vol. 99, pp. 183–199, 1998.
  2. 2.0 2.1 2.2 S. Boyd, S. J. Kim, and L. Vandenberghe et al., "A tutorial on geometric programming," Optimization and Engineering, vol. 8, article 67, 2007.
  3. I. E. Grossmann, "Review of Nonlinear Mixed-Integer and Disjunctive Programming Techniques," Optimization and Engineering, vol. 3, pp. 227–252, 2002.
  4. T. F. Edgar, D. M. Himmelblau, and L. S. Lasdon, Optimization of Chemical Processes, McGraw-Hill, 2001.
  5. P. Shen and K. Zhang, "Global optimization of signomial geometric programming using linear relaxation," Applied Mathematics and Computation, vol. 150, issue 1, pp. 99-114, 2004.
  6. M. Chiang, "Geometric Programming for Communication Systems," 2005.
  7. K. J. Björk and T. Westerlund, "Global optimization of heat exchanger network synthesis problems with and without the isothermal mixing assumption," Computers & Chemical Engineering, vol. 26, issue 11, pp. 1581-1593, 2002.
  8. C. Chu and D. F. Wong, "VLSI Circuit Performance Optimization by Geometric Programming," Annals of Operations Research, vol. 105, pp. 37-60, 2001.
  9. T. R. Jefferson and C. H. Scott, "Quadratic geometric programming with application to machining economics," Mathematical Programming, vol. 31, pp. 137-152, 1985.