INFORMS was present at the annual AAAI meeting in Philadelphia! The Association for the Advancement of Artificial Intelligence hosts one of the top conferences in AI. AAAI invites submissions for "bridge" programs to be held prior to the conference. This year an INFORMS committee organized a bridge session entitled Combining AI and ORMS for Better Trustworthy Decision Making. The organizing committee was @Sven Koenig, @Michela Milano, @Willem-Jan van Hoeve, and @Segev Wasserkrug.
Here is the abstract:
The goal of this bridge program is to unite AI and OR/MS practitioners and researchers to improve trustworthy decision-making in key socio-technical areas such as supply chains, healthcare, crisis management, homeland security, robotics, wildlife conservation, medicine, transportation, and finance. It aims to equip them with better tools by familiarizing them with each other's techniques and domains and bring the disciplines together to advance the research and applications at the intersection of AI and OR/MS so as to improve decision-making.
The organizers put together an outstanding program of plenaries, talks, posters, plus an entire afternoon of intriguing presentations and discussion covering future research themes and projects -- from causal inference and resource allocation to house the homeless, all the way to decentralized POMDP solved autonomously by robots in an Amazon distribution center.
Here's a few personal takeaways
- The "solutions" generated by AI applications are probability distributions. Users collaborate with the application to refine the probability distribution. This style of solution engineering was pioneered by Laplace in the early 19th century.
- ORMS solution engineering can certainly benefit from this perspective! The plenary by @Phebe Vayanos demonstrated how solution engineering that combines probabilistic reasoning with optimization and queueing can satisfy requirements for optimality as well as meeting the goals of a social service agency tasked with implementing a fair and enlightened approach to an intractable human problem.
- Solutions generated by ORMS applications are point estimates. The rigor demanded by industrially implementable solutions highlights the importance of mathematical modeling to satisfy the downstream uses of the solutions. Queuing models for computer networks contribute to sizing of caches and design of channel noise reduction, and so forth.
- AI solutions can certainly benefit from an ORMS perspective! As in the days of Laplace, the great challenge is how to combine the mathematical models of the physics with an honest accounting of data uncertainty. Perhaps the way forward is to extend the "regularization" term as recommended by @Avishai Mandelbaum.
The union envisioned by the organizers is going to have a full slate of challenges, from publication style to education pipelines. But it is perhaps safe to say that finding willing customers is not one of them!
One way to promote this union is data. We envision a common online data service for ORMS-style problem instances to inspire and challenge both communities. While there are certainly many logistical and governance concerns for such a service, there are no technical barriers. I did not manage to keep my hand down during the request for volunteers to help with the data challenge. More to come :)
------------------------------
Alan King
IBM Research
Yorktown Heights NY
------------------------------