On Sun, Jul 9, 2023 at 10:20 AM John Rose <[email protected]> wrote:

> I do wonder though what criteria would be used to discern amongst various 
> computable models of AIXI assuming there is spectrum of them. I already have 
> a model for that but I wonder what the specialists in the field say...

Good predictors (including AIXI, other AI, and lossless compression)
are necessarily complex. The simple proof: Suppose you have a simple
program that can learn to predict any computable sequence of bits.
Then I have a simple program that generates a stream you can't
predict. My program runs a copy of your program and outputs the
opposite bit.

Two examples:
1. SINDy, mentioned earlier, predicts a time series of real numbers by
testing against a library of different functions and choosing the
simplest combination. The bigger the library, the better it works.
2. The top compressors on most benchmarks have a lot of code because
they look for all the special cases and different file formats to pick
from. For example, PAQ searches for JPEG images embedded in other
files and applies a special model to compress them. PreComp searches
for deflate streams and unzips them so they can be recompressed with
better algorithms. But these are common formats. The best compressors
could look for hundreds or thousands of obscure formats and never get
all of them.

For example, how would you compress the bit sequence x[i] = x[i-24]
XOR x[i-55]? It looks like a random bit sequence (until it repeats
after 2^55 - 1 bits) unless you know to look for this particular
pattern. How would you compress an RC4 generated stream if it appears
random unless you know the key? AIXI^tl would eventually discover the
key by brute force search, but not before the heat death of the
universe.

-- 
-- Matt Mahoney, [email protected]

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Taf667527679b18c3-M60401287acbf6e66404b0660
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to