More seriously, I don't totally disagree, but I've never seen stochastic programming problems where the agents' action impacts the environment, changing the distribution of the 2nd stage data. That is, most work I've seen assumes the 2nd stage data is independent of the 1st stage decision. If, by contrast, you assume the 2nd stage data depends on the 1st stage action, doesn't the problem become dynamic programming instead?
There is such a thing as "decision-dependent uncertainty" in stochastic programming. But (IMO) our ability to actually solve the resulting models is so laughably poor that it renders the models essentially useless.
Jim and I are working on some new (I think) ideas to try to handle one specific case of DDU a bit better, where you can decide *when* you want your uncertainty revealed.
Why does stochastic programming with recourse (necessarily) have low impact? It depends on the resource structure modeled in the instance, no?
Jeff! That's just like, your opinion, man...
More seriously, I don't totally disagree, but I've never seen stochastic programming problems where the agents' action impacts the environment, changing the distribution of the 2nd stage data. That is, most work I've seen assumes the 2nd stage data is independent of the 1st stage decision. If, by contrast, you assume the 2nd stage data depends on the 1st stage action, doesn't the problem become dynamic programming instead?
I realize these boundaries are a little loose.
There is such a thing as "decision-dependent uncertainty" in stochastic programming. But (IMO) our ability to actually solve the resulting models is so laughably poor that it renders the models essentially useless.
Jim and I are working on some new (I think) ideas to try to handle one specific case of DDU a bit better, where you can decide *when* you want your uncertainty revealed.
Loving the gradient this blog is following :)