I believe I have discovered a way to save a bunch of computation time. This is actually something I already knew but I just hadnt applied it to this specific computation. Basically, the smoother takes about 8 hours to do 500 particle over 100 time steps...the filter can do the same computation in less than 2 minutes.
The cost is that the smoother has a better fit of the data, how much better of a fit? Not much...maybe a few percentage points better. This makes the filter desirable to at least test the accuracy of the computations. I am not sure if I can do parameter learning with it because it goes against the theory, but I will still try. I am currently testing the filter with 5000 particles...this would no doubt take about a month to do with the smoother. I will probably have very similar results in about an hour with the filter.
This is the accomplishment of the day. Once again, just scaling back the ambition for a while, making sure all my expectation computations are correct.
No comments:
Post a Comment