On Fri, Jan 31, 2020, 3:24 PM <immortal.discover...@gmail.com> wrote:
> On Friday, January 31, 2020, at 2:04 PM, Matt Mahoney wrote: > > Compression is a highly experimental process. Most of the stuff I tried > either didn't work or resulted in tiny improvements. > > Last I checked I did 1 thing and shaved off 76MB of the 100MB wiki8. Yes I > studied it but whoever figured it out first got the big shave. The trick is > not to code, but to philosophize. You need a plan. You need to think. > There's another trick to shave more off I say. :-) > The trick is to study what others have done and come up with improvements by writing code to test your ideas. We need less philosophy and more actual data. My very early version of a neural network compressor, P12, a precursor to paq1 did that in 2000. I called it P12 because it was my 12th experiment worthy of a version number. I was inspired by a 1996 neural compressor by Schmidhuber and Heil that took several days to compress 10 KB of text. My version ran at a practical speed and got better compression. You can read about it here. http://mattmahoney.net/dc/mmahoney00.pdf Out of 200 compressors I tested, about 70 compress enwik8 to under 24 MB. You can get about 25 MB with LZ77, 20 MB with BWT, 19 MB with PPM, and 15 MB with context mixing. Maybe you can come up with something better? ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T409fc28ec41e6e3a-Mdac4f31f1051bac2406d5305 Delivery options: https://agi.topicbox.com/groups/agi/subscription
