26
MILP Formulations for Stochastic Dominance

Stochastic dominance

Embed Size (px)

Citation preview

MILP Formulationsfor Stochastic Dominance

JamesLuedtkeNew Formulations for Optimization Under Stochastic Dominance ConstraintsSIAM Journal on Optimization, 2008

Stochastic dominance

Recall:

So, for the case

The problems

No assumptions neitheron X nor g

A simplification

We can focus on

with

and assume:

Second-order stochastic dominance

Image: http://www.flickr.com/photos/admiriam/

A characterization

SDLP

constraintsvariables

Another characterization

and Y with finite means.

(Strassen)

cSSD1

variables

constraints

(Luedtke)

cSSD2

We can actually replace with

Better performance in practice(using CPLEX dual simplex)

First-order stochastic dominance

Image: http://www.flickr.com/photos/kome8/

A characterization

FDMIP

variables

constraintsPoor LP

relaxation bounds

cFSD

variables

constraints

(Luedtke)

FSD

The LP relaxationyields a formulationfor SSD

Solving the MILP

If: ● X is a polyhedron and ● g(x, \xi^iare affine in x

is a MILP

Branching

Select levelyi should begreater than

At most one of thesecan be positive

SOS1

Branching

Order preserving heuristic

Sort

Fix so it satisfies:

Solve

Computational results

Image: http://www.flickr.com/photos/piper/

Portfolio optimization

Portfolio optimization

● 435 stocks in S&P 500● N daily returns in years 2005, 2006, 2007

● CPLEX 9.0● 2.4 GHz, 2GB memory

Second-order dominance

Time limit of100000 secs

First-order dominance (root LP)

FDMIP: Before addingCPLEX cuts FDMIP.C: After addingCPLEX cuts

FDMIP

FDMIP.C

cFSD

First-order stochastic dominance

H: Heuristics B: Specialized branching