Re: Panalogy examples?
- From: curt@xxxxxxxx (Curt Welch)
- Date: 16 Apr 2009 02:03:23 GMT
huibg4@xxxxxxxxx wrote:
I'm reading Marvin Minsky's book The Emotion Machine and am trying to
understand what problems panalogy is good for. It seems that the key
idea of panalogy is that during the course of an activity or "problem
solving" it's useful to switch to a different mental realm when a
difficulty is encountered in the current realm. This seems plausible
and the book gives a handful of sketchy examples.
I want to understand how panalogy can be applied to a wide variety of
tasks that a household robot might perform but I can't come up with
examples. Does anyone have examples of robot tasks where it needs to
switch between several realms because it hits impasses in the realms?
I've still not read his new book and have not heard of the concept of
"panalogy", but after a quick look at:
http://web.media.mit.edu/~push/Push.Phd.Proposal.pdf
I can throw out some ideas.
First, I really dislike classifying "thought" as a significantly different
process than action and behavior. I don't believe the brain is
particularly structured to make such a distinction. I believe it's only
our bias to look at the brain in such a way that creates such a strong bias
in how we tend to talk about it. Whether we are typing on the computer, or
just "thinking about" typing on the computer, I believe nearly the
identical brain processes are at work. The only significant difference is
that in once case, the fingers are moving, and the other, the fingers are
not moving.
The examples of "panalongy" that I quickly ran across in my 30 seconds of
"research" on this subject seemed to be limited to the concept of using
different thought processes to solve a problem. Because I don't believe
thought is different than action, I interpret such an idea to mean the
exact same thing as "we use different patterns of behavior to solve a
problem".
If you can understand how the brain creates patterns of behavior to solve a
physical motion problem, I think you will also understand the basic
processes at work that allow us to solve mental problems (which I believe
can all be described as physical problems where the brain processes is
disconnected from the body parts they would normally control - like the
lips).
In order to act intelligently, the brain needs the power to produce a
sequence of actions directed towards a purpose. As required, the brain
must switch from one sequence, and one set of goals, to another.
As a simple example, lets say the robot needs to pick up a block on the
floor. How many different ways might it be able to do that? If it's got
two arms, it could pick up the block with either arm. If it's been trained
to perform the "pick up with right arm" sequence and the "pick up with left
arm" sequence, it must make a choice about which of these two very
different sequences to use. And if one sequence doesn't seem to be
working, it needs the power to switch from one sequence to the other.
If it has learned more complex behaviors, it might have learned how to use
a tool (other than it's own arm and hand) to pick up the block. It might
use a shovel of some type to scope it up. Doing that would be making use
of a completely different set of learned behaviors in order to reach the
same goal of getting the block off the floor.
Likewise, if the object the robot is trying to pick up has an odd shape, or
weight distribution, it might need to move to a different location in order
to get a good grip on the object. For example, if it has learned to pick
up a suitcase by the handle, it may first have to position itself so the
suitcase is running along it's side, and then grab the handle and pull to
lift. This is just one more skill the machine has learned that it can make
use of.
These are all different learned sequences of behaviors that can be applied
to any situation. A smart robot will need thousands of different sequences
of behavior it has learned. And when it's presented with a new problem
(one it has never seen before), it will have to pick the sequence which is
most likely to work to solve the current problem. The power to pick the
correct sequence of actions to address the current problem created by the
current state of the environment is the foundation of the abstraction
problem. If there's a coffee cup on the floor it needs to pick up, how
should it solve that problem? Should it use the sequence for picking up a
block that it learned? Or use the sequence for picking up a suitcase that
it's learned?
All examples of abstraction, and analogy, can be reduced (I believe) to the
more straight forward problem of picking the correct sequence of actions to
best deal with the current environment.
Where we talk about this action selection problem as a problem of
"thinking" or a problem of "acting" I think it reduces the same brain
processes either way. If we generate a series of words in our mind as we
"think about" what we are going to do, we are reacting to the environment
by talking. We are producing series of sounds we have been trained to
produce because they have been proven in the past to be useful to put our
brain into the right state to allow it to solve some physical problem we
are faced with. If we talk without moving our lips, it's because the brain
has learned to change it's internal state _as_ _if_ we were making the
sounds, and hearing what we said, without having to actually make the
sounds. Just like moving our arm, and seeing what happens, allows our
brain state to change, the brain has learned to make those same state
changes even without moving the physical arm - simply by partial generation
of the states associated with arm movement while blocking those signals
from reaching the arm.
So whether the brain is regulating what sequence of actions to create, or
simply "thinking about" what actions we might take, it's still a problem of
action selection.
In your question you talked about "switching between realms". To me, that
is nothing more than switching between action patterns, and switching
between goals.
If we need to pick up something from the floor, does the robot use the
"pick up suitcase" realm of action, or does it use the "pick up block"
realm of action? I think that's what this idea boils down to.
At a more abstract level, we might use different thought patterns to solve
a problem. Do we use the realm of mathematics to solve it, or the realm of
geometry to solve it? Or some other realm?
If you have read, and learned to talk about, thousands of different "realms
of thought", you have simply trained you mind to follow many different
patterns of behavior. It's no different, (in my view) than having to train
your body to perform thousands of different physical actions.
If a house robot needs to open a jar of pickles, which "realm of thought"
does it use to get the job done? Does it use the "open the beer bottle"
"realm of thought", or does it use the "crack the egg" realm of thought?
Or the "open the box of crackers" realm of thought?
Interacting with the real world means that we (our our robots) never get to
see the exact same problem twice. It means we must select which set of
actions, are mostly like to work best in the current (never before seen)
set of conditions - such as needing to get the pickles out of a jar we have
never seen before.
The more sequences of actions we have learned, the more likely it is that
we will have something in our big bag of tricks, that will work to solve
the current problem. And the better we are at making good choices about
which "tool" (aka sequence of action) to pull out of that bag to try, the
more intelligent we are. If we have already learned 100 different ways to
open a container, the odds are, one of them will work well to open some new
container we have never seen before. Does it have a lid we need to lift?
Or a cap to screw off? Or do we need to use a knife and cut it open? Or
rip with our teath? Or do we get it open by handing it to the spouse and
making the sounds "can you open this for me?"
These are all different sequences of actions we have learned by past
experience, and the job of being intelligent, is the job of determining
which sequence to try at the moment. And that is done by making some
determination about which past environment, is most like the state of the
current environment. Is this jar of pickles more like the bag of dog food
we just opened, or the beer bottle?
An intelligent machine will work by selecting reactions (selecting what
"path of action") to follow, in response to the current state of the
environment, and doing a good job of that, will require the machine to
determine, by some measure of common traits, how close the current
environment is to past environments. The problem of abstraction, and
analogy, is just the problem of selecting the best current action, for the
current environment.
Humans make heavy use of language and thought to guide the state of the
brain to a state that allows us to correctly pick some action. But again,
the selection of which "language" or "thoughts" to produce next, is really
no different than the selection of body motions. It's just more learned
behavior in our huge bag of tricks we pick from as required.
If we need to get to the kitchen, do we think of the problem in terms of
solving a written maze problem (drawing a line on a paper maze drawing)?
Or do we think of it in terms of past memories of what we will see when we
turn right, or left at a point in the hall? If we learn to solve a maze
puzzle in a game book, we can use that experience, to help guide how we
navigate through a large complex building with lots of different connecting
halls. Drawing marks on a piece of paper with a pencil is a very different
physical action than walking though the halls to try and get to the
kitchen. But yet, there are similarities in the events that would allow a
machine, with sufficient powers of abstraction, to make use of what it
learned by making marks on the paper, to help it navigate the halls in the
building. But again, it's just the application of aspects pf past
behaviors (drawing on paper) to solve a current problem (walking down
halls).
So, in summary, you state:
It seems that the key
idea of panalogy is that during the course of an activity or "problem
solving" it's useful to switch to a different mental realm when a
difficulty is encountered in the current realm.
To me, that simply means, humans, and robots, need to learn many different
sequences of behavior, and must have the power to select what course of
action to try, and be able to switch to a different course of action, at
any instant, when there is an indication that the different course of
action might be better.
and you ask:
I want to understand how panalogy can be applied to a wide variety of
tasks that a household robot might perform but I can't come up with
examples. Does anyone have examples of robot tasks where it needs to
switch between several realms because it hits impasses in the realms?
I can give you thousands of examples. If it's trying to open a jar using
the "twist lid" course of action, and that doesn't work, it might try to
switch to the "pull off lid" course of action, or the "give it to spouse to
open" course of action.
If it's trying to get to the kitchen, and is blocked by a closed door, it
might need to switch from the "walk down hall to kitchen" course of action,
to the "open door" course of action. When that doesn't work, it might
switch to the "unlock door with key" course of action. Or the "knock on
door and make sounds like - hey unlock the door" course of action. Or the,
turn around and take other path into kitchen" course of action.
If it's trying to clean the dirt off the floor with the mop, and it's not
working (because the mop is dirtier than the floor), it might have to
switch to the "clean mop in bucket of water" course of action and then
return to the "push mop around on floor" course of action.
You can extrapolate this same action selection problem down to micro level
actions as well. If the current action is raise arm to reach door knob,
then when the eyes detect the hand is too high, it might have to switch to
the "lower arm to reach door knob" course of action.
All complex behavior can be abstracted to: what is the best thing for me to
do at this moment given the current state of the environment? And all
complex intelligent behavior, is a matter of learning how to produce good
answers to that question, in all typical types of environments the agent
might come across.
The "mental realm" in my view, is just the state of the brain in terms of
what goals are currently active based on the current state of the mind, and
that mental state just determines what course of action the system is
likely to select next as the environment changes.
I don't know if Marvin Minsky would agree with anything I've said here
(probably not), but that's how I look at these ideas.
--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.
- References:
- Panalogy examples?
- From: huibg4
- Panalogy examples?
- Prev by Date: Panalogy examples?
- Next by Date: Anyone go to the 2nd annual conference of Artificial General Intelligence?
- Previous by thread: Panalogy examples?
- Next by thread: Anyone go to the 2nd annual conference of Artificial General Intelligence?
- Index(es):
Relevant Pages
|
