AI for Operations Research

Location: Boole, 36 (Mekelweg 4).

Speaker: Wouter Kool (ORTEC)
Title: Data vs. expertise: learning to optimize
Abstract:
Where does Machine Learning (ML) meet Operations Research (OR)? I will discuss some interesting ideas on the scale from 'classic' OR/human designed algorithms to data/ML driven approaches to solving optimization problems. I will then dive into our recently developed method, which is at one end of the scale and uses a graph neural network and reinforcement learning to learn from scratch a policy that constructs solutions to routing problems such as the TSP.

Speaker: Joaquim Gromicho (VU / ORTEC)
Title: Does data beat the optimum?
Abstract:
Mathematical optimization is based on finding solutions with proven optimality for models consisting of an objective function and constraints defined on the same decision variables.
Nothing beats an optimum, and that is mathematically proven.
The Operations Research community has always been relying on mathematical optimization, whose power is well-know from both theory and practice.
The simplest of such models, Linear Optimization (also known as Linear Programming) has increased efficiency of virtually all sectors of human activity.
Increasing efficiency has always been important but nowadays is a matter of survival: mankind needs to lower her footprint otherwise we are orchestrating our own extinction, see the latest and alarming report from the UN at https://www.un.org/sustainabledevelopment/blog/2019/05/nature-decline-unprecedented-report/.
Solving linear optimization models can be considered as a commodity that scales very well, huge models are no problem.
That is not the case for all mathematical models.
I can see at least two "Achilles' heels" in mathematical optimization:

1. Is the model correct? A relevant question even for linear models.
2. Can the model be solved in practice? Not all models are linear.
   Both question clearly challenge the value of 'proven optimality'.
   The optimum of a wrong model is useless, in as much as an optimum that cannot be found in a realistic amount of time.
   For years, dealing with these questions was a matter of art and expertise: models are handcrafted and their complexity is well understood by those properly educated.
   Until the dawn of big data analytics and the wonders from artificial intelligence we needed to become better in the art of modeling and in the methodologies for solving or approximating the optimum of models.
   Nowadays, the public opinion seems to prefer processing raw data with AI above the use of traditional OR. Game over? No. I am convinced that together we can bring the game to new and exciting levels, and maybe even save the world.