time-series, simulations and games
Mar. 16th, 2009 01:19 am![[personal profile]](https://www.dreamwidth.org/img/silk/identity/user.png){kind=small}
guslQ: How can we simulate things that we don't see/understand completely?
A: Imperfectly! I imagine that weather simulations, and this simulation of the 2004 tsunami (actually a reconstruction i.e. a "postdiction") are explicitly based on probabilistic models, and that the movie shown to us is something like the maximum-likelihood path.
If we treat non-observable things (including micro level, as well as unobservable macro variables) as noise, then these systems are approximately causal Markovian. (Of course, the fact that super-resolution works refutes this Markov assumption, but let's call that a micro effect)
This feels very trivial, but it also feels insightful.
From a (sufficiently) complete causal model, we can create games, in which players can mess with the state of the world at any given time, and watch the consequences of their actions unfold.
I am struck by the idea of automatically generating simulation games for any given domain, by simply plugging in a time-series model, learning a transition function, and slapping on causal assumptions.
--
I can imagine some useful educational simulations. What if instead of memorizing processes the usual way, biology students had games to play? and the higher you advance, the more detailed the simulation gets!
--
Of course, physical simulations probably need physics too! I trust the real specialists to be right about this. But it's tempting to think of encoding the laws of physics as merely a prior. :-)
A: Imperfectly! I imagine that weather simulations, and this simulation of the 2004 tsunami (actually a reconstruction i.e. a "postdiction") are explicitly based on probabilistic models, and that the movie shown to us is something like the maximum-likelihood path.
If we treat non-observable things (including micro level, as well as unobservable macro variables) as noise, then these systems are approximately causal Markovian. (Of course, the fact that super-resolution works refutes this Markov assumption, but let's call that a micro effect)
This feels very trivial, but it also feels insightful.
From a (sufficiently) complete causal model, we can create games, in which players can mess with the state of the world at any given time, and watch the consequences of their actions unfold.
I am struck by the idea of automatically generating simulation games for any given domain, by simply plugging in a time-series model, learning a transition function, and slapping on causal assumptions.
--
I can imagine some useful educational simulations. What if instead of memorizing processes the usual way, biology students had games to play? and the higher you advance, the more detailed the simulation gets!
--
Of course, physical simulations probably need physics too! I trust the real specialists to be right about this. But it's tempting to think of encoding the laws of physics as merely a prior. :-)