Comment by peoplewrong | Hacker News Reader

Comment by peoplewrong | Hacker News Reader

Parent Full thread

> The kind of system described (though only very vaguely) in the article is a rule-based system, possibly an expert system. These hand-crafted systems were the norm in AI up until the '90s or so and are still in widespread use in various domains (e.g. in finances, or air travel bookings etc).

this is wrong at least for UNOS (united network for organ sharing)-which is the main system in the article-the current computer program used to find matches is an integer linear program. I think it is an interesting question if some kind of rule based system could replicate their ability to find the matches. I tend to think not.

https://pdfs.semanticscholar.org/0837/80a157a7f8ebfa44dce0bc...

page 16 "To our knowledge, there is no solver that would scale to the nationwide steady-state size— including the CMU solver used by UNOS. This solver is based on the work of Abraham et al. (2007), with enhancements and generalizations by Dickerson and Sandholm, and uses integer linear programming (IP) with one decision variable for each cycle no longer than L (in practice, L = 3) and constraints that state that accepted cycles are vertex disjoint. With specialized branch-andprice IP solving techniques, Abraham et al. (2007) were able to solve the (3-cycle, no chains, deterministic) problem at the projected steady-state nationwide scale of 10,000 patients"

edit: wait wait this comment could be out of date. The google scholar for this paper says 2018 but the paper says 2013??

Cheers. Just as you were posting this I read your comment and noticed, in the link, the reference to MIP. However, it does sound like early matching systems where simply hand-crafted rule-bases encoding experts' decision-making process, so expert systems:

One night in 2000, tired of delivering the heartbreaking news to patients and their loved ones that no suitable kidney could be found, a US nephrologist named Michael Rees lugged home several crates of files and spent the next few hours scrutinizing blood, antibody, and tissue data, and comparing patient charts. The work was mentally grueling. Eventually, he realized he had no viable matches—but also, that if the pool were bigger, pairs could be made. Working with his father Alan Rees, a computer scientist, Michael Rees created a simple computer program that did the work of pairing up donors and recipients, introducing AI to the matching process.

I bet the first version of this was some kind of maximum bipartite matching. like hopcroft-kraft

but on the topic of experts check out this paper by the same research group. This is much more of an 'AI' paper that does include expert domain knowledge and learning

"FutureMatch: Combining Human Value Judgments and Machine Learning to Match in Dynamic Environments" from AAI2015

http://jpdickerson.com/pubs.html

>> edit: wait wait this comment could be out of date. The google scholar for this paper says 2018 but the paper says 2013??

Chill! Your refs are up-to-date :)

There's two versions of the paper, one published in the Proceedings of the ACM Conference on Electronic Commerce, in 2013, the other published in 2018, in, er, Informs PubsOnline (??) in 2018. The latter appears to be an online journal- sorry if I messed up the name, I don't know the journal.

Reply on news.ycombinator.com