6 ms·
Small data, Kalman filter, belief... They're using Bayesian statistic. Probably added some expert belief to their prior distribution (informative prior). The w
by anthony_doan 7y ago
Small data, Kalman filter, belief...
They're using Bayesian statistic. Probably added some expert belief to their prior distribution (informative prior). The whole article made it sound spicier than it has to be. The majority of the ML algorithms out there have too many parameters so it require a lot of data to overcome certain weaknesses like selection bias in tree ensemble, etc... So with small data it seems like ML are turning toward Bayesian. Also article may be missusing the term inference.
- starwars11 7y agothey do not use priors or expert belief. they solve the system id problem directly - which in most problems is unresolvable
- bananaquant 7y agoI'm just working through the math in the paper right now. To me, it looks like all of the formulas could be expressed in a probabilistic PL and MCMC could be used to estimate the probabilities of both latent state and parameters. Would there be some particular roadblocks to this approach?
- yalph 7y agoWhat does solving system id mean?
- starwars11 7y agoestimating model parameters. normally people use heuristics because it’s an unreasonable problem for linear dynamical systems without additional assumptions. here, they’re able to produce analytical results and finite sample analysis
- starwars11 7y ago*unresolvable
- objektif 7y agoDo you have any recommended material to study this topic?