Decision Optimization - From Theory to Practice

Overview of decision optimization: challenges, lessons learned, and future outlook

1. Decision Optimization –From Theory to
Practice
Eray Çakici, Ph.D.
Senior Operations Research Engineer

2. BS, Industrial Engineering, Baskent University
MS, PhD, Industrial Engineering, University of Arkansas
20 years of industry experience
(IBM, Transplace(Uber Freight), Zero Gap Analytics)
Adjunct Faculty at Bogazici, Koc, Baskent Universities
Author and Referee of many academic articles
Introduction
2

3. Agenda
1
Introduction
2
Theoretical
foundations
4
Conclusions
3August, 2025/ © 2025 IBM Corporation
Motivation
What is Decision
Optimization?
Problem definition(s)
Solution methods
State of the art
Lessons Learnt
Future Expectations
3
Practical
Applications
Use cases
What really matters?
Deployment challenges

4. Motivation
4
•Years of collaboration with the ILOG team, pioneers in practical optimization
•Decision Optimization / Mathematical Programming / Operations Research still
do not receive the well-deserved attention and underutilized
•Many AI pilots fail to transition to deployment — decision optimization is no
exception
August, 2025/ © 2025 IBM Corporation

5. What is Decision Optimization?
5
•Managing limited resources in the most efficient way
•Resources can be: Machines, Vehicles, People, Capital, Materials
•Keywords to look for:
•minimize/maximize
•decide, assign, choose, plan, schedule
•how many, how much, which, when, where
•limited capacity, limited supply, limited budget
August, 2025/ © 2025 IBM Corporation

6. A Simple Analogy
Machine Learning
Machine learning can be compared to teaching a
baby to distinguish between apples and
pineapples, where the goal is to train a model to
recognize patterns in data
Optimization
Optimization is like picking the juiciest apple from
a basket, where the goal is to find the most
optimal solution within a given set of constraints.
August, 2025/ © 2025 IBM Corporation 6

7. Business concepts
Business goals:
•Maximize profit
•Maximize service
•Minimize cost
•…
Business rules/constraints
•Capacities
•Compatibilities
•Priorities
•…
Business Concepts
Data
•Cost
•Capacity
•Supply
•Demand
•Price
•…
Data
Mathematical
model
Plan or
schedule
with metrics
Optimization
engine
The Structure of Optimization Models
August, 2025/ © 2025 IBM Corporation
7

8. Agenda
1
Introduction
2
Theoretical
foundations
4
Conclusions
8August, 2025/ © 2025 IBM Corporation
Motivation
What is Decision
Optimization?
Problem definition(s)
Solution methods
State of the art
Lessons Learnt
Future Expectations
3
Practical
Applications
Use cases
What really matters?
Deployment challenges

9. Optimization problem involves:
Decision variables (x): A vectorof variables
Objective function (f(x)): A function to minimize or maximize
Constraints (C(x)): Conditions that limit possible combinations of values for
decision variables (feasible solutions)
Problem:
min/max f(x)
subject to C(x)
Problem Definition
9
August, 2025/ © 2025 IBM Corporation

10. Operations Research
•Decision Optimization mainly relies on Operations Research (OR) techniques.
•The field of OR was first developed during World War II, when scientistshelped
the military to solve complex logistical and strategical problems.
•The name, Operations Research comes from its application to military
operations.
•After the war, scientists started applying Operations Research techniquesto
similar problems in industry.
10
August, 2025/ © 2025 IBM Corporation

11. George B. Dantzig and the simplex method
•Linear Programming (LP) is the best-known mathematical optimization
technique - many other optimization methods are derived from LP.
•Dr. George B. Dantzig is considered the "father" of linearprogramming.
•While working at the Pentagon, in the 1950s, as mathematical advisor tothe
comptroller of the U.S. Air Force, he formulated the general LP problem and
developed the simplex algorithm for solvingit.
•He continued his pioneering work as a researcher and academician.
11
August, 2025/ © 2025 IBM Corporation

12. Solution Techniques
•Linear Programming (LP) / Nonlinear Programming (NLP)
•Integer Programming (IP) / Mixed-Integer Programming (MIP)
•Variables take integer values (partially or fully)
•Constraint Programming (CP)
•Focuses on satisfying logical constraints (not just optimizing an objective)
•Heuristics & Metaheuristics
•No optimality guarantee
•Dynamic Programming
•Solves problems by breaking them into smaller subproblems
•Stochastic Optimization
•Handles uncertainty in the data
12
August, 2025/ © 2025 IBM Corporation

13. State of the Art
•LP/IP/MIP Solver performance improvements continue (problem complexities
grow faster)
•i.e. Gurobi, NVIDIA (GPU-accelerated)
•Advanced heuristics/metaheuristics-based solvers
•i.e. Hexaly, Timefold
•Non-linear solvers
•i.e. FICO
•Constraint programming
•i.e. CPLEX
•Strong performance and easy to model special constraints in scheduling
•Of course, LLMs & AI-Assisted Modeling for development, debugging, and tuning
August, 2025/ © 2025 IBM Corporation
13

14. Agenda
1
Introduction
2
Theoretical
foundations
4
Conclusions
14August, 2025/ © 2025 IBM Corporation
Motivation
What is Decision
Optimization?
Problem definition(s)
Solution methods
State of the art
Lessons Learnt
Future Expectations
3
Practical
Applications
Use cases
What really matters?
Deployment challenges

15. Optimization Landscape
15
Manufacturing
•Production scheduling
•Supply chain design
•Predictive maintenance
optimization
Transportation
& Logistics
•Routing
•Railway timetabling
•Empty container
repositioning
Energy &
Utilities
•Unit commitment
•Waste management
•Market Bidding
Finance
•Portfolio optimization
•Cash management
•Trade matching
Retail &
E-commerce
•Replenishment planning
•Marketing systems
•Pricing optimization
Healthcare
•Staff scheduling
•Operating room planning
•Radiotherapy optimization
15August, 2025/ © 2025 IBM Corporation

16. Store-2-Store Replenishment Planning
August, 2025/ © 2025 IBM Corporation
Use Case
•Store-to-store transfers
•When?
• Central items are under a certain stock level.
•Why?
•Sales Increase, Customer Satisfaction, Space Allocation for New Items
Challenges
•400+ stores, 1000+ products (with
different sizes)
•Storage & transportation limitations
•Business rules i.e. major size
completeness
•320M+ decision variables
16
Benefits
•ROI (Return of Investment) is less
than 2 weeks
•Sales ratio of transferred products
has increased from 61% to 66% in
the first 3 weeks
•7 million USD revenue increase is
expected per year

17. Procurement Planning
August, 2025/ © 2025 IBM Corporation
Use Case
•Optimize usage of sugar – A mixing/recipe optimization problem
•Satisfy sugar demand by reducing total costs including handling,
transportation, processing and procurement
–monthly and as-needed basis.
•Different suppliers, different sugar types
Challenges
•Specific business rules and multiple
goals
•Dissolution, Warehousing Capacities
•Supplier Availabilities
•Quotas
•Supply & Distribution cost structure
17
Benefits
•Planning period reduced to
minutes from days
•Agile and flexible process
•1-2% cost savings

18. Sales and Operations Planning
August, 2025/ © 2025 IBM Corporation
Use Case
•Global supply chain network optimization
•An optimization solution that is sustainable to changes in operating costs,
procurement costs, currency exchange rates, taxes, demand and possible
manufacturing capacity changes
Challenges
•A complex network of facilities and
customers
•20 grinding mills, 12 kilns, 40 silos in
15 Locations (6 domestic, others
abroad)
•Fluctuations in currency exchange
rates, taxes, demand and
manufacturing capacity changes
18
Benefits
•Supply chain cost reduction by
5-15%
•Better customer service
•What-if scenario analysis
capabilities

19. Only Real when Deployed
•More than 80% of AI projects fail, more than double the failure rate of traditional
IT projects
RAND Corporation (2024)
•Only 48% of AI projects make it into production; up to 80% fail overall due to
poor data, unclear value, or scalability challenges.
Gartner & Informatica (2025)
•Over 70% of AI projects fail to move from pilot to production.
Ayadata & CIO Dive (2025)

•75% of AI initiatives fail to scale, and 69% don’t make it into live operational
use—most due to messy/inconsistent data.
TechRadarPro (2025)
19
August, 2025/ © 2025 IBM Corporation

20. 20
Applicable >better, cheaper, faster
What really matters?
August, 2025/ © 2025 IBM Corporation
20

21. 21
Visualization Recognition Confidence
Marketing Matters
“Machine learning has been extraordinarily successful at marketing itself—and many
businesses eagerly adopted predictive analytics—while decision optimization, despite
its decades of proven impact, has remained hiding in plain sight.”
— Thomas H. Davenport, The AI Advantage: How to Put the Artificial Intelligence
Revolution to Work, MIT Press, 2018
21
August, 2025/ © 2025 IBM Corporation

22. 22
Challenges
Collection
Validation
Preparation
Generation?
DATA
Optimize ALL ?
Start small, win big
SCOPE
Explainability?
How do I know your solution is optimal?
TRUST
Longer cycles than typical AI projects
DURATION
Political Movements i.e. Arab Spring 2010-2012
Floods i.e. 2011 Thailand
Pandemics i.e. 2020 Covid
Team Changes i.e. Ginni Rometty retirement
August, 2025/ © 2025 IBM Corporation
22

23. 23
Challenges cont’d
120! = 6.689e+198
Earth has roughly (and we're speakingveryroughly here) 7.5 x 10
18
grains of
sand, or seven quintillion, five hundred quadrillion grains.
Many businesses are complex and uncertain than they look
23
August, 2025/ © 2025 IBM Corporation

24. 24
Challenges cont’d

Over-expectations & Over-promises
A trend over the last 10+ years
August, 2025/ © 2025 IBM Corporation
24

25. Agenda
1
Introduction
2
Theoretical
foundations
4
Conclusions
25August, 2025/ © 2025 IBM Corporation
Motivation
What is Decision
Optimization?
Problem definition(s)
Solution methods
State of the art
Lessons Learnt
Future Expectations
3
Practical
Applications
Use cases
What really matters?
Deployment challenges

26. Some Mistakes
26
•Engaging the wrong people, asking the wrong questions, solving the wrong problem
•Assuming everyone shares the same goals
•Too much focus on details
•Not showing early results
•Wrong or avoided feasibility analysis (in terms of people, processes, culture,
infrastructure etc.)
August, 2025/ © 2025 IBM Corporation

27. Lessons Learnt
27
•If you do not co-create, do not create
•Not speaking the business language = delays & misalignments
•Story telling is a must requirement
•"Le mieux est l’ennemi du bien. " -Voltaire
•“The better is the enemy of the good.”
•In certain cases, “good enough” is quite sufficient
•Good assumptions don’t break the model — reality is already an approximation
•Once again, applicability > better, faster, cheaper
August, 2025/ © 2025 IBM Corporation

28. Future Outlook
28
•Back to reality
•Enough shiny apps, enough over-expectations & over-promises
•Reframing the conversation
from “Can we use AI for this?”
to “How can we empower our people through AI?”
•Greater focus on process mining, process automation, people-centric designs
•The rise of Decision Optimization
•Growing visibility and “measurable” impact
•Systems that “act”
August, 2025/ © 2025 IBM Corporation

29. Data Science & AI Elite Team / © 2020 IBM Corporation
Thank you !