In the usual usage, I think it also needs to be capable of state estimation. My impression is that it is often aspirational, just as simulators for self-driving cars do not cover the entire state space.
I really like your requirement that a simulator must support counterfactual (or conditional) predictions. That insight is new to me.
From my very incomplete reading of the papers from Waymo and Waabi, they claim to have created simulators primarily by fitting deep nets to data. Waymo has published papers where they generate counterfactual queries against the learned simulator in order to improve the controllers. Doing this right would seem to involve satisfying Pearls' causality requirements for the third rung. Do you agree?
Regarding sim2real, my robotics colleagues make heavy use of "domain randomization", which I guess should be viewed as an ensemble of simulators intended to represent uncertainty about the true system dynamics. Reinforcement learning is then supposed to work well in all of those simulators, similar to robust control.
Curious if you are planning to cover some of the fuss about "Digital Twins" in your lectures.
Is "Digital Twin" anything more than an oddly popular buzzword for simulator? I have not found any evidence to the contrary.
Precisely! I say it all the time and folks get all defensive and attack me---now I'll cite your comment ;)
Thanks!
In the usual usage, I think it also needs to be capable of state estimation. My impression is that it is often aspirational, just as simulators for self-driving cars do not cover the entire state space.
Projections, commonly used in economic and energy policy analysis are a particular kind of simulation, often confused with prediction
https://theconversation.com/please-no-more-projections-what-we-need-are-predictions-and-theyre-harder-126734
What's the full citation for Zakka et al.? Thank you for yet another insightful post, Ben.
Ah, it's the survey I linked earlier in the post.
https://arxiv.org/abs/2502.08844
I should have called out Kevin by name there.
Thanks, Ben.
I really like your requirement that a simulator must support counterfactual (or conditional) predictions. That insight is new to me.
From my very incomplete reading of the papers from Waymo and Waabi, they claim to have created simulators primarily by fitting deep nets to data. Waymo has published papers where they generate counterfactual queries against the learned simulator in order to improve the controllers. Doing this right would seem to involve satisfying Pearls' causality requirements for the third rung. Do you agree?
Regarding sim2real, my robotics colleagues make heavy use of "domain randomization", which I guess should be viewed as an ensemble of simulators intended to represent uncertainty about the true system dynamics. Reinforcement learning is then supposed to work well in all of those simulators, similar to robust control.