Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M....

27
Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa [email protected]

Transcript of Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M....

Page 1: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems

George M. WhiteSITE

University of [email protected]

Page 2: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Examples of very difficult problems

medical personnel in hospitals contact centre personnel judicial staff assignments examination scheduling portfolio management

Page 3: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Page 4: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Page 5: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Examples

These are all examples of NP-hard assignment/scheduling problems. They are characterized by having a series of non-linear constraints

We wish to find solutions such that all constraints are satisfied

If this is not possible, we wish to find solutions such that a maximum number of constraints are satisfied.

Page 6: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Page 7: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Optimization

There is often more than one possible solution. In this case we want the one that is best (i.e. we want to optimize some property of the schedule) total wages paid overall satisfaction personnel coverage separation

Page 8: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Optimization

This implies that we must optimize some cost function(to the best value permitted by the constraints and the time available). unidimensional optimization multidimensional optimization

Page 9: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Optimization

This also means that we will have to use an approximation algorithm to find good solutions. Exact solutions require too much time for real life problems. tabu search particle swarm optimization simulated annealing great deluge partialcol IDWalk etc

Page 10: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

yor-s-83

Page 11: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

yor-s-83

Page 12: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

The shape of the curve

at some time in the future it seems reasonable to assume that the best penalty values will reach a limit, i.e.

Page 13: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

the form of dP/dt is unknown but it is reasonable to assume that it is some function of the current penalty

Page 14: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

expanding this as a Maclauren series yields

...!

)0(...

!2

)0(")0(')0()(

)(2 n

n

Pn

fP

fPffPf

or

Page 15: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

we want to simplify this equation as much as possible (but no further) so we try

dP/dt = a0

this doesn’t work

Page 16: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

try

dP/dt = ao + a1P

this doesn’t work

Page 17: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

the next simplest form is

it turns out that this is a plausibleform

Page 18: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

at the limiting value

2lim2lim1

221 0 PaPaPaPa

dt

dP

2

1lim a

aP

Page 19: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

this often appears in the literature with symbol substitution

21 aa

and the equation is written

2PPdt

dP

Page 20: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

The solution for this equation is

teP

tP

0

)(

where P0 = P(0)

Page 21: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

The limiting value of P(t) is

limP

Page 22: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

To estimate the limiting penalty of a data set

1. Collect the data representing the “current champion” over time.

2. Fit a curve to this data.3. Calculate the limiting value of this

curve.

Page 23: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Problems

Lack of data: The largest number of points for any of the data sets is 6.

Number of parameters: 3 parameters

Uncertain and irregular spacing in data:

Curious data points: The first (1996) data points:

Page 24: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Problems

Therefore, the numerical results must be regarded as preliminary estimates, subject to review as more data becomes available.

Page 25: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

yor-s-83

Page 26: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Similar behaviour has been

observed for other data sets of the same type.

Work continues on other sets of data from other real-world problems.

Page 27: Attempts to find an optimum solution penalty value for certain classes of NP-Hard problems George M. White SITE University of Ottawa white@site.uottawa.ca.

CORS - Ottawa

Thank you

George [email protected]