On: Applying machine learning to the freight industry

[platz]

What happened to the gold standard of logistics industry, which is Operations Research (OR)?

OR has essentially been customized to this domain long before generalized ML appeared on the scene - I can't imagine some of-the-self sci-kit learn libraries have improved on this much.

[_raoulcousins → platz]

It's being used in the industry heavily. The term is extremely unsexy. In my mind, I'm an OR analyst. But if people ask me my job, I try to keep a straight face and say I'm a data scientist with an expertise in prescriptive analytics.

[R_haterade → _raoulcousins]

There are literally dozens of us.

[yummyfajitas → platz]

Operations Research is totally unsexy. It still works better than ever before; I'm shipping a work project based on stochastic programming pretty soon. But we'll probably call it "AI" in marketing materials.

[SixSigma → yummyfajitas]

This.

I work in OR, specifically Computational Logistics.

There is plenty of room for operational improvement in all industry, all the time. Articles like this paint us as big and stupid because that's how the "disruption" people see all industry.

Often it is not the optimization that is lacking but the will of disparate companies to co-operate.

An example - the Port of Rotterdam barge system is horribly inefficient for the barge operators because the individual shippers see minor gains. The challenge is getting the shippers to agree and co-operate, not to machine learn the best routing for the barges.

[zump → platz]

I think OR only works if the models are linear.

[srean → zump]

Huh! what ?

I think you are thinking of linear programming, and OR is way more than that. OR is the first chapter of an OR book.

[srean → srean]

Too late to edit, meant LP is the first chapter in an OR book.

[raverbashing → zump]

Yes, and there are several ways of linearizing models for OR

(however I think the MIP people waste too much time looking for the perfect linear model and the best solution instead of using things like simulated annealing which might give a very good solution in 10% of the time)

[kyllo → zump]

There are nonlinear solvers available but they are slower and not guaranteed to converge on a globally optimal solution.

[rokob → kyllo]

Your statement is not universally true, it depends on the nonlinearity.

[kyllo → rokob]

It's correct in the general case. A solver algorithm can find a globally optimum solution for some nonlinear functions, but it cannot be guaranteed to find a globally optimum solution for arbitrary nonlinear functions. That would be like solving P = NP or the Halting Problem.