Overview: Applications Optimization

General

On the fol­lo­wing sub­pages, you will find infor­ma­ti­on about the types of pro­blems that can be sol­ved using mathe­ma­ti­cal optimi­zation. The methods used and the basic theo­ry are brief­ly out­lined; code examp­les detail prac­ti­cal applications.

Relevance

When­ever it comes to choo­sing bet­ween mul­ti­ple opti­ons, the ques­ti­on of the best opti­on natu­ral­ly ari­ses — an optimi­zation pro­blem. The spe­ci­fic appli­ca­ti­on may invol­ve choo­sing pro­duct para­me­ters, desig­ning sche­du­les, esti­mat­ing unknown quan­ti­ties, or making decis­i­ons in com­plex sys­tems. Hence, optimi­zation is now found ever­y­whe­re and is used in finan­ce, esti­ma­ti­on theo­ry, opti­mal con­trol, sche­du­ling, indus­tri­al pro­duct design, metro­lo­gy, tele­com­mu­ni­ca­ti­ons, and much more.

Mathe­ma­ti­cal optimi­zation deals with the mini­miza­ti­on or maxi­miza­ti­on of an objec­ti­ve func­tion \(f(x)\) under cons­traints for the varia­ble \(x\). The for­mal nota­ti­on for this task is
$$ \begin{align} \min_x ~~~&f(x) \\ \text{subject to} ~~~&x \in D \end{align}$$
This encodes the ins­truc­tions that an \(x\) is to be found which mini­mi­zes the func­tion value \(f(x)\) (\(\min _x f(x)\)). Here, \(x\) can only be cho­sen from a spe­ci­fic set \(D\) (\(x\in D\)), which for­ma­li­zes the requi­re­ments for the optimi­zation varia­ble \(x\).

Examples

The pro­blem for­mu­la­ted abo­ve includes a lar­ge set of tasks rele­vant to prac­ti­ce. Often, \(f(x)\) quan­ti­fies cos­ts, mate­ri­al expen­dit­u­re, dura­ti­ons, uncer­tain­ties, pro­ba­bi­li­ties of error, or other nega­ti­ve per­for­mance mea­su­res who­se reduc­tion is desi­red. The (often high-dimen­sio­nal) varia­ble \(x\) repres­ents decis­i­ons with direct impacts on the per­for­mance mea­su­res; such as invest­ment decis­i­ons, pro­duct pro­per­ties, sche­du­ling, expe­ri­men­tal designs, or safe­ty mar­gins. The­se are sub­ject to user-defi­ned demands that, for exam­p­le, for­ma­li­ze the adhe­rence to cer­tain stan­dards. The fol­lo­wing illus­tra­ti­on demons­tra­tes this, and the table lists exem­pla­ry interpretations.

Exam­p­le\(f(x)\)\(x_1,x_2, …\)\(D\)
Stock port­fo­lio- Expec­ted profitInvest­ment into stock 1, 2, …Acqui­si­ti­on constraints
Truss topo­lo­gy designMate­ri­al consumptionLength, thic­k­ness of barsSta­bi­li­ty certificate
Faci­li­ty layoutPro­duc­tion timesLoca­ti­on devices 1, 2, …Machi­ne neighborhoods
Phar­maceu­ti­cal testsUncer­tain­ty, signi­fi­can­ce levelTests of class 1,2, …Depen­ci­es of test classes
Trans­por­ta­ti­on planningTime effortSequence of locationsTrans­por­ta­ti­on capacity
Job sche­du­lingWorkloadStart­ing times of tasksJob sequen­ces
Table with examp­les of appli­ca­ti­ons of mathe­ma­ti­cal optimi­zation. The term \(f(x)\) is the objec­ti­ve func­tion to be mini­mi­zed, \(x_1, x_2, …\), are the optimi­zation varia­bles, and \(D\) repres­ents the cons­traints on a solution.

Types of tasks

Many prac­ti­cal tasks from disci­pli­nes such as finan­ce, busi­ness admi­nis­tra­ti­on, logi­stics, phy­sics, engi­nee­ring, tele­com­mu­ni­ca­ti­ons, game theo­ry, data ana­ly­sis, … can be for­mu­la­ted as the optimi­zation pro­blem \(\min _x f(x), x\in D\). Regard­less of the spe­ci­fic appli­ca­ti­on, we iden­ti­fy three clas­ses of tasks into which tasks from the abo­ve disci­pli­nes can be categorized:

Optimal design

The optimi­zation varia­bles \(x\) repre­sent sta­tic design para­me­ters of a pro­duct. They are to be cho­sen, taking into account a model based on preli­mi­na­ry con­side­ra­ti­ons, such that the pro­duct maxi­mi­zes a cer­tain per­for­mance mea­su­re \(-f(x)\) while ensu­ring cer­tain pro­per­ties are maintained.

This class of tasks includes, for exam­p­le, expe­ri­men­tal design, port­fo­lio optimi­zation, trans­port, rou­ting, and logi­stics pro­blems, as well as the optimi­zation of machi­nes and struc­tures and their arrangement.

Optimal estimation

The optimi­zation varia­ble \(x\) repres­ents para­me­ters of a pre­dic­ti­ve model \(g(x,z)\) with inde­pen­dent varia­ble \(z\). The model aims to make pre­dic­tions about which out­co­me \(l\) occurs, depen­ding on a pre­vious­ly made obser­va­ti­on \(z\).

The rela­ti­onship bet­ween \(z\) and \(l\) is to be mode­led, whe­re data­sets \(\{z_k,l_k\}_{k=1}^n\) have been obser­ved, and the model \(g(x,z)\) needs to be adjus­ted so that it con­tra­dicts the data foun­da­ti­on as litt­le as pos­si­ble. This class of tasks includes clas­sic regres­si­on pro­blems of para­me­ter esti­ma­ti­on and rever­se engi­nee­ring, but also spa­ti­al sta­tis­tics, pre­dic­tion, lear­ning of cor­re­la­ti­on struc­tures, sta­tis­ti­cal signal sepa­ra­ti­on, and recon­s­truc­tion. The­se tasks occur, for exam­p­le, in com­plex mea­su­re­ment problems.

Optimal control

The optimi­zation varia­ble \(x\) repres­ents a sequence of con­trol inputs in a sys­tem that beha­ves dyna­mi­cal­ly and pos­si­bly ran­dom­ly. This input must be cho­sen so that the dis­crepan­cy \(f(x)\) bet­ween the deve­lo­p­ment car­ri­ed out by the sys­tem and a pre­fer­red tar­get deve­lo­p­ment is minimized.

In this pro­cess, the con­trol varia­bles \(x\) may be limi­t­ed to a spe­ci­fic area \(D\). Pro­blems of this task class include sequen­ti­al sta­tis­ti­cal decis­i­on pro­blems such as traf­fic flow regu­la­ti­on, sup­p­ly chain manage­ment, pri­cing for finan­cial deri­va­ti­ves, adap­ti­ve medi­cal tre­at­ment, as well as tasks of clas­si­cal con­trol engineering.

Outlook

Tasks from all the afo­re­men­tio­ned are­as have alre­a­dy been suc­cessful­ly for­mu­la­ted and sol­ved as mathe­ma­ti­cal optimi­zation pro­blems. The fol­lo­wing sel­ec­tion includes sim­pli­fied exam­p­le pro­blems that are non-tri­vi­al to sol­ve wit­hout the appro­pria­te tools. Illus­tra­ti­ons and brief descrip­ti­ons show­ca­se the pro­blem, and a com­men­ted for­mu­la­ti­on as a mathe­ma­ti­cal optimi­zation pro­blem cla­ri­fies the con­nec­tions bet­ween real-world issues and the mathe­ma­ti­cal mode­ling level. We use spe­cial for­mal arran­ge­ments of objec­ti­ve func­tions and equa­tions and ine­qua­li­ties that can be relia­bly sol­ved through line­ar pro­gramming, qua­dra­tic pro­gramming, semi­de­fi­ni­te pro­gramming, and dyna­mic programming.