Thanks for providing this format. I've enjoyed viewing it very much. One crucial thing in my opinion is when it comes to automated-decision making there are algorithms which produce "bad" decisions and algorithm which produce "good" decision. Unfortunately the assessment good/bad may come months or even years after the decision. While human communication is slow you may have strong counterparts in the arguments. So major weaknesses can be (theoretically) found early. When it comes to modelling I haven't seen a error free implementation. So my question is: How can you achieve reviews of automatically made decisions with respect to the correctness of the underlying assumptions in the modelling.
The issue of not being able to access the adequacy of a decision until long *after* the decision was made is equally present for people. Having a human making the decision does change anything with regards to this challenge. Humans have no special powers in this area. If at the time the decision is generated (by an algorithm), a human can object to the long-term viability of the decision, and if this objection is well-reasoned; then the algorithm must be modified to immediately take this into account. Automation doesn't preclude keeping human intelligence around to adjust or improve the automation itself. Furthermore , if the decision is automated, it is possible *after the fact* to modify the logic so that the same mistake isn't made again. This property is in sharp contrast with employees who may or may not comply, who may or may not learn from the mistake, or who may arrive fresh and ignorant due to turnover. Joannes
While modeling a S&OP system is there any easy way to tackle innumerable constraints in say production processes, customer priorities, technical difficulties etc. or do we have to map each one piece by piece into the algorithm?
In order to mechanize the decision-making processes, there is no alternative, but to "map each one piece by piece into the algorithm" as you very correctly suggest. This is one of the core responsibilities of the Supply Chain Scientists at Lokad. Cheers, Joannes
Thanks for providing this format. I've enjoyed viewing it very much. One crucial thing in my opinion is when it comes to automated-decision making there are algorithms which produce "bad" decisions and algorithm which produce "good" decision. Unfortunately the assessment good/bad may come months or even years after the decision. While human communication is slow you may have strong counterparts in the arguments. So major weaknesses can be (theoretically) found early.
When it comes to modelling I haven't seen a error free implementation. So my question is: How can you achieve reviews of automatically made decisions with respect to the correctness of the underlying assumptions in the modelling.
The issue of not being able to access the adequacy of a decision until long *after* the decision was made is equally present for people. Having a human making the decision does change anything with regards to this challenge. Humans have no special powers in this area.
If at the time the decision is generated (by an algorithm), a human can object to the long-term viability of the decision, and if this objection is well-reasoned; then the algorithm must be modified to immediately take this into account. Automation doesn't preclude keeping human intelligence around to adjust or improve the automation itself.
Furthermore , if the decision is automated, it is possible *after the fact* to modify the logic so that the same mistake isn't made again. This property is in sharp contrast with employees who may or may not comply, who may or may not learn from the mistake, or who may arrive fresh and ignorant due to turnover.
Joannes
It would be amazing if you could start adding this to Spotify! 😀
That's a good idea. It's something we'll look into. We do like the video format, but perhaps we should expand the listenership via Spotify.
While modeling a S&OP system is there any easy way to tackle innumerable constraints in say production processes, customer priorities, technical difficulties etc. or do we have to map each one piece by piece into the algorithm?
In order to mechanize the decision-making processes, there is no alternative, but to "map each one piece by piece into the algorithm" as you very correctly suggest. This is one of the core responsibilities of the Supply Chain Scientists at Lokad. Cheers, Joannes