Presentation delivered during Data Science Rzeszow meetup: I will present reasons for optimization being superior to predictive algorithms in data science practical applications. I will cover exemplary case studies, tools and hints from my experience on delivering hybrid solutions that exploit both prediction and optimization

1. recommendation =
optimization(prediction)
Wit Jakuczun, PhD

2. Once upon a time Wit met a Customer that needed
demand forecasts to ...

3. Customer: I need better demand forecasts.
Me: I understand. Can I have a simple question?
Customer: Yes
Me: Imagine I created a demand forecasting model and provided you with
desired 5M forecasts (numbers). What are you going to do with them?
Customer: Well… I will take the forecasts and optimize my logistics decisions using
the numbers.
Me: I see. Why don’t we talk about the whole decision problem? Maybe the
inefficiency is not in demand forecasts but in optimization part?
Customer: Can you create a math model for such complex business problem with
many constraints and exceptions? I thought it was impossible.
Recommendations are calculated in “Sheet” and it is a bottleneck.
Me: It is possible to build decision support system that uses mathematical
optimization for your problem
Customer: Great, let’s talk about the details.

4. What has happened in the past?
What is an optimal course of
actions for the past?
What can happen in the future?
What is an optimal course of
actions for the future?
DataAnalyticsOptimization

5. Forecasts
Expert
(guts mainly)
Decisions
Forecasts
Recommendations
(optimization)
Decisions
Expert
(guts mainly)

6. Forecasts
Expert
(guts mainly)
Decisions
Forecasts
Recommendations
(optimization)
Decisions
Expert
(guts mainly)
Notreproducibleandnotscalable
Reproducibleandscalable

7. Forecasts/predictions are “just” a tool for better decisions.
And better decisions are based on right recommendations.
And right recommendations are result of optimizing business KPIs that are linked
to business decisions.
Optimization model deals helps to improve robustness of decision making
process robust.

8. Automation of the complex business process.
Transition to central/global planning.
Learning from best (optimal) decision.

9. I was selling vehicle routing solution to a logistics company.
I managed to persuade manager/owner to meet and talk with the team.
After 1 hour presentation of the solution to the company I got one question
How can I create an invoice in this solution?
This is real story that happened to me. Similar story can be read in “Being wrong with Clarke & Wright” by Robert E.D.Woolsey

11. There are three commercial highly efficient solvers for mixed integer problems.
Solutions have been on the market for 25+ years…. and are still in development.
Very resistant to parallelization and distributed computing techniques.
Very sensitive to data.
Tightly coupled with business.
Only “auto” for mip.

12. There is only one good library for mixed integer programming that is open-source.

13. Graphics taken from SCIP solver webpage. Benchmarks from Hans Mittelmann

14. Most common ML libs are open-source.
Open-source is very efficient compared to commercial solutions.
Easy to create distributed implementations
Fairly insensitive to data.
Less tightly coupled with business.
Quite a few AutoML solutions that work.

15. Problem is infeasible.
Explanation is extremely difficult.
Customers expect any-time feature.
Debugging is hell :)

17. Mixed-integer
programming
black-box
hard to customize
limited applicability
(still wide!)
Constraint
programming
white-box
easy to customize
not limited applicability
Metaheuristics
custom-box
easy to customize
not limited applicability
Easy Difficult

18. Mixed Integer Programming
IBM CPLEX
Gurobi
Fico
Local Solver
Constraint programming
Sicstus
IBM CP Optimizer
Commercial world
Mixed Integer Programming
CBC (solver)
MIP (wrapper)
Or-tools (wrapper)
Constraint programming
ECLiPSe
Choco
Gecode
Or-tools
Open-source world

19. Easy
I can declare model using existing solver.
Fairly difficult
I can solve problem with a sequence of easy problems.
Very Difficult
I must implement custom solver.

20. Business requirements are almost impossible to be collected upfront.
Performance is not satisfactory.
Solution quality is not satisfactory.
No solution found is not acceptable.

21. Validator
Solver
Optimization engines
Integration layer
Load balancer
Validator
Solver
Optimization engines

22. Start with business process and decisions
Start with small and iterate.
Use real data since the first day of the project.
Assume problems are infeasible or internally contradictor.
Deal with must vs nice to have requirements.

24. Saving up to 20% of cash management costs in Deutsche Bank

25. Challenge
• Factory throughput was too low
• Upgrade or not to increase throughput
Solution
• Integrated planning and scheduling optimization model
• Scenario generation to support investment decision
• Tailor made optimisation model
Effects
• Ability to support investment decision with numbers
Based on academic work by Roman Barták

26. Challenge
• Dynamic and unpredictable orders flow
• Complex tasks
Solution
• Automation by optimisation
• Tailor made optimisation model
Effects
• In progress - feasibility tests of the
working solution
Collect Pay Drive Deliver Drive
Collect Pay Drive Deliver Drive
Collect Pay Drive Deliver Drive

28. Contact info
● Private: wit.jakuczun@gmail.com
● Business:
○ wit.jakuczun@fourteen33.com
○ wit.jakuczun@wlogsolutions.com