On Tue, Sep 4, 2012 at 7:07 PM, Abram Demski <[email protected]> wrote:
> This paper:
>
> http://arxiv.org/abs/0909.0801/
>
> uses a prediction framework similar (but, somewhat different) from PAQ,
> plugs it into a generic planner, and gets decent results. I find it unlikely
> that PAQ would do much worse in the same experimental set-up. (I find it
> likely that PAQ would do just a bit better... could be wrong, though.)

I would expect PAQ to do better, but it depends on the test
environment. In the paper they use CTW, which is described here:
http://mattmahoney.net/dc/dce.html#Section_423

CTW does well on stationary sources but poorly when it has to adapt to
changing statistics. For example, given a bit sequence 0000000001
observed in some context, then CTW would assign P(1) = 1.5/11 (by
adding 0.5 to the counts). PAQ uses indirect context models,
essentially counting 0s and 1s that occurred when the same sequence
was observed in other contexts. If you have 2 different file types,
such as a text file and an image file, then CTW will compress them
better separately than together. For PAQ it won't make much
difference.

CTW and PAQ both use adaptive weighted averaging of log-likelihood bit
predictions. However, CTW is restricted to mixing in a chain from high
to low order contexts with a separate weight associated with each
lower order context. PAQ has no restrictions on the types and the
order of contexts that can be mixed. It associates weights not with
contexts but with models or with models plus a small context separate
from the ones being mixed. Since there are fewer weights, they can be
more accurate because they can be adjusted in small increments.


-- Matt Mahoney, [email protected]


-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com

Reply via email to