Transcript INFORMS05

Forward and Backward
Deviation Measures and Robust
Optimization
Peng Sun (Duke)
with Xin Chen (UIUC) and Melvyn Sim (NUS)
Agenda




Overview of Robust Optimization
Forward and Backward Deviation Measures
Uncertainty Sets and Probability Bounds
Conclusions
Uncertain linear constraints

Uncertain linear constraint

Chance constraint

Hard to solve when  is small
Linear constraints with uncertainty

Affine Uncertainty
: zero mean, independent but not necessarily
identically distributed
Overview of Robust Optimization

Worst case



Relies only on the distribution support
Easy to solve (Soyster 1973)
Extremely conservative
Overview of Robust Optimization

Goal of Robust Optimization:
 Easy to obtain feasible solutions that satisfy
chance constraint
 Not as conservative as worst case
Overview of Robust Optimization

Design of Uncertainty Set


Subset of worst case set, W
: Budget of uncertainty


Control radius of uncertainty set, S
max: Worse case budget

For reasonable probability bound,  << max
Overview of Robust Optimization

Robust Counterpart

Semi-infinite constraint


Tractability
Polyhedral Uncertainty Set


Linear Programming (LP)
Conic Quadratic Uncertainty Set

Second Order Cone Programming (SOCP)
Overview of Robust Optimization

Ellipsoidal Uncertainty Set



Ben-Tal and Nemirovski (1997), El-Ghaoui et al.
(1996)
Robust Counterpart is a SOCP
Probability bound for symmetrically bounded
uncertainties:
Overview of Robust Optimization

Polyhedral Uncertainty Set
 Bertsimas and Sim (2000)
 Robust Counterpart is a LP
 Probability bound for symmetrically bounded uncertainties:

More conservative compared to Ellipsoidal Uncertainty Set
Overview of Robust Optimization

Norm Uncertainty Set


Robust Counterpart depends on the choice of norm
Probability bound for symmetrically bounded
uncertainties:
Overview of Robust Optimization

Modeling limitations of Classical
Uncertainty Sets


Uncertainty set is symmetrical, but
distributions are generally asymmetrical
Can we use more information on
distributions to obtain less conservative
solution in achieveing the same probability
bound?
--- Forward and Backward Deviation Measures
Forward and Backward Deviation
Measures

Set of forward deviation measures

Set of backward deviation measures
Alternatively,

Forward deviation

Backward deviation
Forward and Backward Deviation
Measures

Key idea
Important property

Subadditivity
Least conservative forward and backward
deviation measures

Suppose distribution is known

p* = q* if distribution is symmetrical
p*, q* ¸  (standard deviation)
p* = q* =  if distribution is Normal


Least conservative forward and backward
deviation measures

p* , q* can be obtained numerically


E.g: p* = q* = 0.58 for uniform distribution in [-1,1]
E.g:
Forward and backward deviation measures
– two point distribution example
Distribution not known

Suppose distribution is bounded in [-1,1]
(but not necessarily symmetrical)
Only know mean and support
Forward and Backward Deviation
Measures: Function g()
General Deviation Measures
General Deviation Measure
Agenda




Overview of Robust Optimization
Forward and Backward Deviation Measures
Uncertainty Sets and Probability Bounds
Conclusions
Uncertainty Sets and Probability Bounds

Model of Data Uncertainty

Recall: Affine Uncertainty
: zero mean, independent but not necessarily
identically distributed
Uncertainty Sets and Probability Bounds

Model of Data Uncertainty
Uncertainty Sets and Probability Bounds

New Uncertainty Set (Asymmetrical)
Uncertainty Sets and Probability Bounds

Recall: Norm Uncertainty Set (Symmetrical)
Uncertainty Sets and Probability Bounds

Generalizes Symmetrical Uncertainty Set
Uncertainty Sets and Probability Bounds
Uncertainty Sets and Probability Bounds
Uncertainty Sets and Probability Bounds

Robust Counterpart
Uncertainty Sets and Probability Bounds


Probability Bound
More information achieve less conservative
solution while preserving the bound. E.g


pj = qj = 1 for symmetric distribution in [-1,1]
pj = qj = 0.58 for uniform distribution in [-1,1]
Conclusions

RO framework



Forward and backward Deviation measures





Affine uncertainty constraints
Independent r.v.’s with asymmetric distributions
Defined from moment generating functions – implying
probability bound
Sub-additivity – linear combinations
Easy to calculate and approximate from support
information
Advantage -- less conservative solution
Next

Stochastic programming with chance constraints