On Sat, Jan 2, 2021, 2:47 PM Ben Goertzel <[email protected]> wrote:

>
>
> "Paraconsistent Foundations for Probabilistic Reasoning, Programming
> and Concept Formation
> "
>
> https://arxiv.org/abs/2012.14474


Extensions to Boolean logic are mathematically interesting. But it isn't
clear to me how "both true and false" and "neither true nor false" differ
in real life, or what is to gain by having both as separate concepts.

Have you done any experimental evaluation of different logics? I have, and
it is quite a bit of effort to optimize both candidates to get a clear
answer. In PAQ6 I use probabilistic logic with both a probability and a
confidence. In PAQ8 I used a single valued probability with an implied high
confidence for values near 0 or 1. It took a couple of years with people
competitively optimizing both methods to establish the latter as superior.
The former had already achieved top rankings in several benchmarks
including the Calgary challenge before being topped by the latter.

In PAQ6, the assertion "the next bit is 1 in context X" has probability r/n
and confidence n, where a 1 bit was previously observed r out of n times in
context X. To combine predictions in contexts X and Y you add the counts,
like probability (rx + ry)/(nx + ny) and confidence nx + ny. These can also
be adaptively weighted. There are practical considerations like how do deal
with large counts overflowing.

In PAQ8, a probability p = r/n has an implied confidence of 1/(p (1-p)). To
combine predictions you average them in the logistic domain, x =
ln(p/(1-p)) using adaptively weighted averaging. The result is converted
back to a probability by the inverse, p = 1/(1 + e^-x). The adaptive
weighting is effectively a neural network trained by gradient descent in
coding cost space.

PAQ8 is simpler, with nice mathematical properties. But that doesn't
necessarily make it better. You can only determine that experimentally with
a lot of work.

>
>

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T425c68f0cea319cd-M013558b6102ddf4bd14abb92
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to