The evolution of optimization technologies - INFORMSnymetro.chapter.informs.org/prac_cor_pubs/optim.pdf · The evolution of optimization technologies ... • Frito-Lay (Production

The evolution of optimization technologies

Dash Optimization, Alkis VazacopoulosINFORMS New York Metro ClubOctober 11, 2005, The Penn Club, New York

3

Agenda

• Optimization Applications– Companies that offer optimization – Companies that use optimization

• Development priorities past and now• Development priorities – Future• Case Studies

4

Companies that Offer Optimization

• Modeling and Optimization– Dash Optimization ****– AIMMS– Ilog ****– Maximal– GAMS– AMPL– **** develop State of the Art solvers

5

Companies – complete solutions

• SAP (Supply Chain) *** ERP • Oracle (Retail, Supply Chain) *** ERP• Combinenet (procurement)• Carmen Systems (airlines ) • i2 (supply chain )• Smartops (inventory Optimization)• Fair Isaac (Banks, Marketing

Optimization)

6

Customers that use Optimization • NFL (Scheduling) (Current Schedule)• Baseball League (Scheduling)• Frito-Lay (Production Planning )• American Airlines (Training Pilots)• Wachovia (Financial)• Siemens (process scheduling)• Toyota (discrete planning and

scheduling)

7

Verticals

• Process Industry• Airlines• Energy• Retail• Financial Applications• Pharma

8

Process industry problems

• Refinery Scheduling • Vehicle Routing Problems• Refinery Expansion/sizing

9

Airlines

• Crew rostering and scheduling• Pricing (revenue Management)• Capacity Planning (size of fleet)• Service maintenance• Parts inventory

10

Energy

• Unit commitment problems • Capacity Planning• Pricing based on Uncertain Demand• Project selection

11

Retail

• Markdown Optimization • Replenishment of Inventory (inventory

Optimization)• Warehouse Management

12

Financial Applications

• Security valuation and selection– Mortgage backed securities

• Bond Trading• Crossing• Index tracking

– Track sector indices at minimum cost• Portfolio tracking

– Track performance at minimum cost

13

Pharma

• Sales promotions• Assign sales reps to physicians

14

Edelman Award

• UPS• NBC• Procter & Gamble• Aspentech (last year)

15

Characteristics

• Size of Problems– 100000 variables

• CPU Time – A few seconds

16

Development DirectionsEase of use

Type of Problem

Performance

17

Development DirectionsEase of use

Type of Problem

•Speed

•Size of problem

•Reliability

Performance

18

Development DirectionsEase of use

Type of Problem

•Speed

•Size of problem

•Reliability

Performance

Out-of-the-box

19

Primal SimplexSolution Times of Six Models with Recent

Releases of Xpress-MP Primal

0

5000

10000

15000

20000

25000

20772 5637 4454 4194 3588 451

Rel 10.28 Rel 11.14 Rel 12.13 Rel 13.10 Rel 14.05 Rel 15.00

Watson_1PreleKen-18EnergyChineseArtur

20

Dual SimplexSolution Times of Six Models with Recent

Releases of Xpress-MP Dual

0

2000

4000

6000

8000

10000

12000

14000

16000

18000

15850 3741 190 158 151 151

Rel 10.28 Rel 11.14 Rel 12.13 Rel 13.10 Rel 14.05 Rel 15.00

Watson_1PreleKen-18EnergyChineseArtur

21

Barrier Method (Interior Point)Solution Times of Six Models with Recent

Releases of Xpress-MP Barrier

0

500

1000

1500

2000

2500

3000

2440 1093 779 539 167 160

Rel 10.28 Rel 11.14 Rel 12.13 Rel 13.10 Rel 14.05 Rel 15.00

Watson_1PreleKen-18EnergyChineseArtur

22

Optimizer Speed

• Last Ten Years– 30 times faster– 100 times hardware– 3,000 times overall– Year on year gains of 125%

• Expectation– 60% per annum from hardware– 30% per annum from software

23

MIP Good Solutions - Fast

24

MIP Performance

XPR 14.10 XPR 15.00neos1 >10000 4neos2 4412 6neos3 >10000 29neos13 >10000 1242neos10 261 126

XPR 14.10 XPR 15.00

25

Development DirectionsEase of use

Type of Problem

Performance

26

Development Directions•Building Applications

•DeploymentEase of use

Type of Problem

Performance

27

Modeling Environment

28

Development DirectionsEase of use

Type of Problem

Performance

29

Type of problem - Technologies

• LP• MIP• QP• MIQP• Constraint

Programming • Nonlinear• Heuristics

• Network design• Capacity Planning• Pricing• Strategic Sourcing• Scheduling

30

Supply ChainCapacity Planning,LP, MIP

Network Design LP, MIP

Scheduling,LP, MIP, CP,

HeuristicsSupply Chain

Pricing,LP,

Nonlinear Strategic Sourcing,

LP,MIP, MIQP

31

Development DirectionsEase of use

Type of Problem

Performance

32

64 bit & Parallel

33

Future

• Grid computing– Solve millions of small scale lps and mips or

nlps in seconds– Examples – Pricing for Airlines-Hotels-Cars

34

Large Scale Programs

• Solve large scale optimization problems• Decompose them• Use different processors to calculate

upper bounds (find feasible solutions and lower bounds – optimality)– Example – pricing for energy in a large

scale (pricing gas in a specific time)– Example (deploy resources after a

catastrophic event – terrorism etc…)

35

Development DirectionsEase of use

Type of Problem

Performance

36

Clever Modeling Languages

• Solve Models in Parallel • Profiling models• Wizard – model development• Development of Clever algorithms

– Expert systems – Artificial intelligence

37

Wizard – decision variables

38

Xpress-Application Developer

39

Screenshots of XAD apps

40

Development DirectionsEase of use

Type of Problem

Performance

41

Xpress-SLP

• Non-Linear• MINLP Option• Large scale• Applications:

– Process Industry– Pricing– Energy

42

• Optimization Technologies

– Mixed Integer Programming: MIP– Finite Domain Constraint Programming: CP

• Planning & Scheduling

– Long and mid-term planning: MIP– Short-term planning, scheduling: CP

Supply chain optimizationRequires MIP, CP, and their combination

CP, MIP and their combination

CP MIPCP+MIP

Time Horizon

43

Xpress-CP Architecture (cooperative solver approach)

Mosel Planning and Scheduling Model

Mosel/IVE ENVIRONMENT

Xpress-MP KALIS

44

Xpress-CP Adopters

• BASF– Chemicals

• Peugeot– Cars

• Procter and Gamble– Discrete products

45

Stochastic Programming

• Basis– Represent uncertainty– Build problem efficiently– Manage scenarios

• Applications– Planning– Supply chain– Interaction with forecasting

46

Stochastic Programming

47

Future

• Data mining • Robust Optimization

48

Classification Problems:example

49

Classification Problems:Application Examples

• Cancer Diagnosis(Mangasarian et al, 1995 –linear separating surfaces)

• Classification of Credit Card Applications, Bonds Rating

(Bugera et al, 2002, 2003 –quadratic separating surfaces)

50

Problems

There are two major biomedical data mining problems associated with microarray data:

• Classification: Determine classes of the test samples.

• Gene detection: For each of the classes, select a subset of genes responsible for creating the condition corresponding to the class.Usual noisiness of microarray data complicates solution of these problems.

51

Credit Rating

• Logical Analysis of Data• Credit scoring of Countries• Credit scoring of Companies

52

Medical Applications

• Different types of Cancer• Obesity

53

Robust Optimization

54

Other examples

• Robust Portfolio Optimization• Robust Inventory policies

55

MOSEL

LP

MIP

SLP

IVE

XAD

MIQP

QPLP

MIP

STO

MISLP

HEURISTICS

VERTICAL APP.

GUI / STUDIO

EXTENSIONS/NAT

SOLVERS

MODELING PLATFORM

56

Some Trends• Profitable-to-Promise• Very large problems• Time of the essence• Custom driven supply chains• Flexibility/lifecycle• Analytics & Forecasting /

Data Mining• Out-of-box performance• Non-expert users

57

Portfolio Management

• Wrap Portfolio– Tracks another index / actively managed

portfolio– Realigns periodically to match risks and

returns– Provides similar performance at lower

management costs

58

WRAP PortfolioT=1 T=2 T=3

INDEX

WRAP

INDEX

WRAP

INDEX

WRAP

Re-Balancing

59

Portfolio Management

Challenges– How to realign with minimum trading costs– How closely to track sector, country..

exposures– When and how frequently to re-align– How to automate the complete process

60

Portfolio Management

• Optimization– Determine trades to re-align with minimum

costs• Business rules

– Determine when to re-align– Generate test cases to determine

–How closely to track–How frequently to re-align

61

Portfolio Management

Want to automate the complete process• Need to combine

Business Rules Optimization

in a ‘single’ application

62

Portfolio Management

• Real Time Management– Rules engine monitors tracked index and

portfolio movement – Calls and runs optimization engine when

index and portfolio deviates beyond acceptable limits

– Rules engine evaluates and executes trades

63

Portfolio Management

Customization – one step further• Determines acceptable deviation limits for

risk-return combinations– Rules generate scenarios – Optimization engine determines action– Enables to predict returns for risk levels

64

Benefits

• Lower monitoring and management costs

• Lower execution costs• Assured compliance

65

Other Applications

• Marketing– State Farm

• Process Industries– AspenTech

• Transport– Schneider

66

Simple Concept

• Business rules monitor and determine when to re-optimize

• Optimization determines best action• Business rules evaluate and execute

actions to achieve desired effect

67

Conclusion

Combine • Flexibility of business

rules and • Robustness of

optimization

Next Generation Intelligent Applications– Automation– Testing– Determining rules ?

68

Workshop

• October 21-22nd

• New York University

69

Informs San Francisco

• Preconference Workshop• 1 Tutorial • 2 talks • Exhibit