--- On Mon, 9/22/08, Ben Goertzel <[EMAIL PROTECTED]> wrote:

> http://goertzel.org/agiq.pdf

Some of the problems you describe in intelligence testing also apply to data 
compression testing. For example, an AI could "cheat" by being tuned with the 
knowledge needed to pass a specific test. This is a common problem with 
compression programs, which are often tuned to do well on popular benchmarks. I 
once wrote as a joke a program that compresses the Calgary corpus files to 1 
byte each.

http://cs.fit.edu/~mmahoney/compression/barf.html

Usually cheating is not so overt as hiding a copy of the benchmark (i.e. the 
test answers) in the program, but developers do test their programs on popular 
benchmarks. But it is hard to avoid unintentional bias in the form of shared 
information, such as external English dictionaries used to tokenize the input 
text. In earlier tests, I would distinguish such programs, but there are other 
ways around this.

A common solution is to make the test data private, which many benchmarks do. 
However, since I run a benchmark and also write compression software, it would 
be unfair not to give competing software the same advantage I would have.

You mentioned the "Nintendog" effect, where an AI could pretend to learn what 
it already knows. Your solution is to start with a base test, give different 
copies different training data, and test them on the newly learned data. In 
effect, we do this when we test a compressor with many different text files. It 
is much harder to hide knowledge about all of the input files in the model than 
about just one file.

In compression testing, there is another option, which is to include the size 
of the decompression program. Any pre-programmed knowledge has to take up space 
which can't be compressed beyond its algorithmic complexity. This is the 
technique used in two benchmarks with prize money: Calgary challenge and the 
Hutter prize, and also in my large text benchmark.


-- Matt Mahoney, [EMAIL PROTECTED]



-------------------------------------------
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription: 
https://www.listbox.com/member/?member_id=8660244&id_secret=114414975-3c8e69
Powered by Listbox: http://www.listbox.com

Reply via email to