I don't think you've answered my point - which perhaps wasn't put well
enough.
All you propose, as far as I can see, is to apply *values* to behaviour - to
apply positive and negative figures to behaviours considered beneficial or
detrimental, and thus affect the system's further behaviour - reinforcing
it, for example.
It is more or less like a value approach to investing on stocks on the
stockmarket - when their value goes up or down, a formula determines whether
the system buys more or less shares.
But this is a purely numerical approach to altering behaviour. There is
nothing a priori wrong with it - although, in fact, (although this is a more
complex argument which I won't really go into), it would never actually work
for AGI, which has to deal with problems where it is impossible to apply
precise or reliable values.
But the important point here is that these *values* are not *emotions* at
all. They're fundamentally different entities and affect behaviour in
fundamentally different ways - your values, for example, will not cause any
pleasure or pain to a self, or have a corporeal, hormonal nature, or
conflict. You, like others, are trying to invest your value system with a
complexity and dignity that it simply hasn't got and has no right to. It's
absurd - you might just as well talk of every plus or minus sign in a
mathematical calculation as conferring pleasure or pain.
It also shows a very limited understanding of emotions.
Matt: Mike Tintner <[EMAIL PROTECTED]> wrote:
Matt: I don't believe that the ability to feel pleasure and pain depends
on
> consciousness. That is just a circular definition.
> http://en.wikipedia.org/wiki/Philosophical_zombie
Richard:It is not circular. Consciousness and pleasure/pain are both
subjective
issues. They can resolved together.
Both of you, in fairly standard fashion, are approaching humans and
animals
as if they were dissected on a table with consciousness/ emotions/
pleasure
& pain lying around.
The reality is that we are integrated systems in which -
a self
is continually subjected to
and feels (or to some extent may choose not to feel)
emotions (involving pleasure/pain)
via a (two-way) nervous system.
The questions Matt has to answer is:
1) are the systems you envisage going to have a self (to feel emotions) -
and if so, why?
No, I am proposing a measure of reinforcement for intelligence in general,
whether human, animal, or machine, all of which fall under Legg and Hutter's
universal intelligence ( http://www.vetta.org/documents/ui_benelearn.pdf ),
which is based on Hutter's AIXI model (
http://www.hutter1.net/ai/aixigentle.htm ). In this model, an agent and an
environment are modeled by a pair of interactive Turing machines exchanging
symbols. In addition, the environment sends a utility or reinforcement
signal
to the agent at each step. The goal of the agent is to maximize the
accumulated utility. The paper on universal intelligence (UI) proposes
defining intelligence as the expected accumulated utility for a randomly
chosen environment (from a Solomonoff distribution of environments, i.e.
self
delimiting Turing machines chosen by coin flips). Hutter's AIXI model shows
that the most intelligent strategy is to guess at each step that the
environment is simulated by the shortest program consistent with the
observed
interaction so far. However, AIXI is not computable.
In humans, it is natural to think of positive utility or reinforcement as a
"reward" signal or pleasure, and negative utility as a penalty, such as
pain.
In this respect, humans seek to maximize expected accumulated utility. But
this is not quite right because utility has no scale in the AIXI/UI model.
If
you double a reward (e.g. food or money) or punishment (e.g. electric shock)
to a human or animal, you approximately double the change in behavior. But
in
the AIXI/UI model, if you double the utility signal, the agent's strategy
does
not change.
I propose a measure of a bound on reinforcement which is more consistent
with
our intuitive notion of pain and pleasure. The strength of a signal is
bounded by the change in the state of the agent, the amount of information
learned, as measured by Kolmogorov complexity. This bound is consistent
with
intuition. For example, a person under anesthesia feels no pain during
surgery and also has no memory (learning) during this time. Drugs that
increase the rate of learning (synaptic changes), such as hallucinogens,
also
heighten sensations of both pain and pleasure. Children learn faster than
adults, and also react more strongly to pain and pleasure.
Allow me to distinguish between utility and reinforcement as follows. An
agent's goal is to maximize utility, but utility is independent of the
agent's
behavior, and has no scale. Reinforcement depends on the agent, such that
if
an agent's state changes from S1 to S2 as the result of reinforcement R,
then
|R| <= K(S2|S1), the number of bits needed to describe the state change.
If you accept this definition then you could say that a human has 1000 times
more capacity to experience pleasure or pain than a mouse because a human
brain is 1000 times larger and therefore can learn 1000 times more.
Likewise,
if humans can learn 10^9 bits and autobliss (
http://www.mattmahoney.net/autobliss.txt ) can learn 10^2 bits, then
autobliss
experiences 10^-7 as much pain or pleasure as a human.
You can interpret this how you wish. I make no claims about the morality of
inflicting pain on animals or programs. Morality is an evolved cultural
belief. We believe in compassion to other humans because tribes that
practiced this belief (toward their own members) were more successful than
those that didn't. Likewise, we eat animals.
-- Matt Mahoney, [EMAIL PROTECTED]
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&;
--
No virus found in this incoming message.
Checked by AVG Free Edition.
Version: 7.5.503 / Virus Database: 269.17.1/1181 - Release Date: 12/11/2007
5:05 PM
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=8660244&id_secret=75488151-54c57f
