S ystems Analysis Laboratory Helsinki University of Practical dominance and process support in the Even Swaps method Jyri Mustajoki Raimo P. Hämäläinen Systems Analysis Laboratory Helsinki University of Technology www.sal.hut.fi
S ystemsAnalysis LaboratoryHelsinki University of Technology
Practical dominance and process support in the Even
Swaps method
Jyri MustajokiRaimo P. Hämäläinen
Systems Analysis LaboratoryHelsinki University of Technology
www.sal.hut.fi
S ystemsAnalysis LaboratoryHelsinki University of Technology
Presentation outline
• Introduction to the Even Swaps method• Hammond, Keeney and Raiffa (1998, 1999)
• Two new techniques to support the method• New concept based on the PAIRS method
• Salo and Hämäläinen (1992)
• Aim to provide support for tasks needing mechanical scanning
• Smart-Swaps software• The first software for supporting the method
S ystemsAnalysis LaboratoryHelsinki University of Technology
Even Swaps method
• Multicriteria method to find the best alternative
• Based on even swaps• Value trade-off, where the value change in
one attribute is compensated in some other attribute
• The alternative with these changed values is equally preferred to the initial one
It can be used instead
S ystemsAnalysis LaboratoryHelsinki University of Technology
Elimination process
• Aim to carry out even swaps that make• Alternatives dominated
• Some other alternative is equal or better than this one in every attribute, and better at least in one attribute
• Attributes irrelevant• Every alternative has the same value on this
attribute
These can be eliminated
• Process continues until one alternative (i.e. the best one) remains
S ystemsAnalysis LaboratoryHelsinki University of Technology
Practical dominance
• If alternative x is better than alternative y in several attributes, but slightly worse in one attribute x practically dominates y y can be eliminated
• Aim to reduce the size of the problem in obvious cases• No need to carry out an even swap task
S ystemsAnalysis LaboratoryHelsinki University of Technology
Example
• Office selection problem (Hammond et al. 1999)
78
25
Practicallydominated
byMontana
Dominatedby
Lombard
S ystemsAnalysis LaboratoryHelsinki University of Technology
Two new techniques
• Modeling of the practical dominance• Support for looking for efficient even swaps
• New concept based on the PAIRS method• Aim to provide support for tasks needing
mechanical scanning• Computer support to help in these tasks
• For supporting the process – not for automating it
S ystemsAnalysis LaboratoryHelsinki University of Technology
PAIRS Method• Additive value function• Imprecise statements by intervals on
• Attribute weight ratios (e.g. 1 w1 / w2 5) Feasible region of the weights• Ratings of the alternatives (e.g. 0.6 v1(x1) 0.8)
Intervals for overall values• Lower bound for the overall value of x:
• Upper bound correspondingly
n
iiii xvwxv
1
)(min)(
S ystemsAnalysis LaboratoryHelsinki University of Technology
Pairwise dominance
• x dominates y in a pairwise sense if
i.e. if the overall value of x is greater than the one of y with any feasible weights of attributes and ratings of alternatives
0])()([min1
n
iiiiii
wyvxvw
S ystemsAnalysis LaboratoryHelsinki University of Technology
Modeling practical dominance
• General constraints for the weight ratios and value functions• These should cover all the plausible weights
and values
• If x dominates y in a pairwise sense with these general constraints y can be seen as practically dominated
S ystemsAnalysis LaboratoryHelsinki University of Technology
General constraints
• On weight ratios
• On value functions• E.g. exponential value
function constraints• Any value function within
the constraints allowed• Additional constraints, e.g.
for the slope
jirw
w
j
i ,,
1
0 xi
vi(xi)
S ystemsAnalysis LaboratoryHelsinki University of Technology
Use of even swaps information
• With each even swap the user reveals information about his/her preferences
• This information can be utilized in the process
Tighter weight ratio constraints elicited from the given even swaps
Better estimates for practical dominances
S ystemsAnalysis LaboratoryHelsinki University of Technology
Support for looking for efficient even swaps
• Aim to carry out as few swaps as possible to eliminate alternatives or attributes
Scanning through the consequences table
• There may also be other objectives• E.g. easiness of the swaps
Different types of suggestions of even swaps for the decision maker
S ystemsAnalysis LaboratoryHelsinki University of Technology
Irrelevant attributes
• Look for an attribute in which the most alternatives have the same value
Carry out such even swaps that make the values of all the alternatives the same in this attribute
• Compensation in attribute with which new dominances could also be obtained• Possible reduction also in the number of the
alternatives
S ystemsAnalysis LaboratoryHelsinki University of Technology
Dominated alternatives
• Look for such pair of alternatives, where dominance between these could be obtained with fewest swaps• E.g., if x outranks y only in one attribute, carry
out an even swap that makes the values of these alternatives the same in this attribute
• However, the ranking of the alternatives can change in compensating attribute
We cannot be sure that the other alternative is dominated after the swap
S ystemsAnalysis LaboratoryHelsinki University of Technology
Dominated alternatives
• An estimate for each swap, how far we relatively are from dominance• The ratio between
• The allowed value change in compensating attribute, and
• The maximum estimated value change in this• Estimated from general constraints
• d(y, x) = 'likelihood' of y dominating x after this even swap
S ystemsAnalysis LaboratoryHelsinki University of Technology
Example
36 different options to carry out an even swap which may lead to dominanceE.g. change in Monthly Costs of Montana from 1900 to 1500:Compensation in Client Access:
d(Mon, Bar) = ((85-78)/(85-50)) / ((1900-1500)/(1900-1500)) = 0.20d(Mon, Lom) = ((85-80)/(85-50)) / ((1900-1500)/(1900-1500)) = 0.14
Compensation in Office Size:d(Mon, Bar) = ((950-500)/(950-500)) / ((1900-1500)/(1900-1500)) = 1.00d(Mon, Lom) = ((950-700)/(950-500)) / ((1900-1500)/(1900-1500)) = 0.56(Assumptions: linear estimates for value functions; weight ratios = 1)
Initial Range:
85 - 50
A - C
950 - 500
1500 -1900
S ystemsAnalysis LaboratoryHelsinki University of Technology
Use in practice
• The proposed techniques assume an additive value function• Not explicitly assumed in the Even Swaps
method• Can still be used approximatively Suggestions should be confirmed by the
decision maker
S ystemsAnalysis LaboratoryHelsinki University of Technology
Smart-Swaps softwarewww.smart-swaps.hut.fi
• Support for the proposed approaches• Identification of practical dominances• Suggestions for even swaps
• Additional support• Information about what may happen with
each swap• Notification of dominances• Rank colors• Process history
S ystemsAnalysis LaboratoryHelsinki University of Technology
Smart-Swaps software
S ystemsAnalysis LaboratoryHelsinki University of Technology
www.Decisionarium.hut.fi
Software for different types of problems:• Smart-Swaps (www.smart-swaps.hut.fi)• Opinions-Online (www.opinions.hut.fi)
• Global participation, voting, surveys & group decisions
• Web-HIPRE (www.hipre.hut.fi)• Value tree based decision analysis and support
• Joint Gains (www.jointgains.hut.fi)• Multi-party negotiation support
• RICH Decisions (www.rich.hut.fi)• Rank inclusion in criteria hierarchies
S ystemsAnalysis LaboratoryHelsinki University of Technology
Conclusions• Techniques to support the even swaps
process presented• Modeling the practical dominance• Support for looking for efficient even swaps• New concept based on the PAIRS method
• Support for tasks needing mechanical scanning• Especially useful in large problems
• Computer support needed in practice• Smart-Swaps software introduced
S ystemsAnalysis LaboratoryHelsinki University of Technology
ReferencesHammond, J.S., Keeney, R.L., Raiffa, H., 1998. Even swaps: A rational
method for making trade-offs, Harvard Business Review, 76(2), 137-149.Hammond, J.S., Keeney, R.L., Raiffa, H., 1999. Smart choices. A practical
guide to making better decisions, Harvard Business School Press, Boston, MA.
Mustajoki, J., Hämäläinen, R.P., 2003. Practical dominance and process support in the Even Swaps method. Manuscript. Downloadable soon at www.sal.hut.fi/Publications/
Salo, A., Hämäläinen, R.P., 1992. Preference assessment by imprecise ratio statements, Operations Research, 40(6), 1053-1061.
Applications of Even Swaps:Gregory, R., Wellman, K., 2001. Bringing stakeholder values into
environmental policy choices: a community-based estuary case study, Ecological Economics, 39, 37-52.
Kajanus, M., Ahola, J., Kurttila, M., Pesonen, M., 2001. Application of even swaps for strategy selection in a rural enterprise, Management Decision, 39(5), 394-402.
S ystemsAnalysis LaboratoryHelsinki University of Technology
S ystemsAnalysis LaboratoryHelsinki University of Technology
xi
x'i
v(xi)-
v(xi)-
v(x'i)-
v(x'i)-
mini
maxi
0
1
Value function constraints
• Exponential value function constraint
where a (0, 1)
xN = (xi – mini) / (maxi – mini)
vi(maxi)=0, vi(maxi)=1
(here a=0.15)
1
1)(
a
axv
Nx
ii
Appendix
S ystemsAnalysis LaboratoryHelsinki University of Technology
xi
x'i
min(v(x'i)-v(x
i))
max(v(x'i)-v(x
i))
xD
mini
maxi
0
1
Value function constraints
• Slope constraints
wheres (0, 1)
Dx = (x'i – xi) / (maxi – mini)
vi(maxi)=0, vi(maxi)=1
(here s=0.5)
sx
xvxvs iiii 1)()'(
D
Appendix
S ystemsAnalysis LaboratoryHelsinki University of Technology
New constraints from the given trade-offs
• E.g. change xi x'i is compensated with the change xj x'j• Assume an additive value function:
wi v(xi) + wj v(xj) = wi v(x'i) + wj v(x'j)• General constraints for value functions New weight ratio constraint:
))()'(
)'()(max(
iiii
jjjj
j
i
xvxv
xvxv
w
w
Appendix
S ystemsAnalysis LaboratoryHelsinki University of Technology
The use of practical dominance in practice
• Suggestions - not automatization• The user should confirm the dominances
• Strict gereral constraints Smaller feasible region Alternatives may become incorrectly
identified as dominated ones
• Loose general costraints Larger feasible region Not as many dominances, but all these
should be real ones
Appendix
S ystemsAnalysis LaboratoryHelsinki University of Technology
Estimate how far we are from dominance• Assume, e.g. that
• The change xi yi (vi(xi) > vi(yi)) is compensated with the change xj x'j (vj(x'j) > vj(xj))
• x'j should remain under yj to make y dominate x
• The allowed value change in j:
The maximum plausible value change in j:
• Derived from general constraints in PAIRS
)()( jjjj xvyv
)))()((max())()'(max( iiiij
ijjjj yvxv
w
wxvxv
Appendix
S ystemsAnalysis LaboratoryHelsinki University of Technology
Estimate how far we are from dominance
• An estimate how close we are relatively to make y dominate x• The ratio between the allowed compensation
and the maximum plausible value change
• The bigger the ratio is, the better the dominance would be obtained
• Strict constraints can also be used instead of intervals
)))()((
))()((min(),(
iii
jjj
yvxvw
xvyvwxyd
Appendix