AI Experiment With Strange Results

OK. Since, I can not post in the Research Forum, this is the next best forum to post this thread...

A challenge to you all.

When I was in college, I interned at a bank's corporate heardquarters. I made a friend

with a guy who used to work at NASA as a computer engineer... I wish I can remember his

name. But sadly I can't. Let's call him Bob.

Bob told me back in the 70s, he and the NASA team did experiments on AI. One experiment

was they created these basic mouse robots made from spare parts. They were simple small

robots on wheels with "whiskers" used as sensors. Programming was very, very simple:

If sensors hit wall, randomly turn Left OR Right

The NASA engineers ran these mouse robots in a huge maze with a beginning and end. At

first, the robots couldn't get far. Eventually, the robots with simple programming learned

how to navigate thru the maze. They learned! After thousands of trials, the robots were

zipping thru the mazes and found the exit quickly.

Does anyone have an explanation about this? Has anyone had a similar result in a similar

experiment? If not, would anyone like to accept the challenge and try to duplicate this

experiment?

Thanks in advance.

I am certainly interested..but cannot offer any information. I am also an engineer and memory of course was very limited in those days. You would need a couple gated logic chips and battery as well. Interesting hopefully someone will have some information

I've studied Cognitive Sciences extensively. I am still a novice in this field, like all other AI researchers.

However, I know that the software needed to learn how to efficiently navigate a maze through trial and error is well known. It is probably the same algorithm needed to auto-route printed circuit boards using CAD/CAM software.

#

I've recently heard of a truly amazing AI experiment dealing with high speed random number generators. Does anyone know what I'm talking about here? Has anyone heard about this?

they could have "learned" any number of ways, but my first guess is that it was using some type of distributed processing / network computing. Essentially, all the mice form a collective unit, which educates itself. So if Mouse A runs into a wall at 5,8 - all the other mice know that 5,8 is not valid. It's a form of quasi-intelligence. It can look pretty amazing, but it's really not a big deal.

reply to post by scientist

---

Yeah but in 70s..how did the first mouse pass that info along..to all other mice..

but the maze runs were not supposed to be a 'learning experience'
for the robot mice..
their program was intended to Intentionally bounce their 'whiskers' off of walls...


there is something either unsaid or there is a slanted analysis being
foisted on the reviewers.

the spontaneous generation of intelligence...is the same b.s. as the
spontaneous generation of Life arguement

[edit on 12-12-2007 by St Udio]

reply to post by Buck Division

---

Okay – not much interest in my earlier question regarding this cool new AI project I heard about. It's off topic, but I want to get this into ATS -- here are the details that I know about:

#

According to what I heard from a friend at [ University Name Deleted ]: a few years ago some researchers hooked up a large network of high speed random number generators around the world, for the purpose of predicting natural disasters, and fed the info into some sort of correlator. The results of this particular experiment were inconclusive.

HOWEVER, a few months ago, as part of separate project, they fed these random number generators into some sort of immense “semantic concordance” software, part of a language translation system, and have achieved what appears to be real-life “qualia” from the resulting software.

Although this system was not designed to appear self-aware, it took on many characteristics of self-awareness, including a desire not to be switched off! It also seemed to have some pretty interesting forecasting capabilities.

In particular, the software could be used to reliable predict whether a joke was going to be funny, and could rewrite poor jokes to make them more humorous!

I asked my friend at [ University Name Deleted ] to corroborate this with some details, but she is reluctant to say more, because this is part of some [ Organization Deleted ] project.

She gave me a link that seemed somewhat relevant, available here:

www.scieng.flinders.edu.au...

This friend of mine is deep into I Ching. Except for that oddity, she is very down to earth. Most people would accept her as a bonafide genius, especially in Information Theory.

I was wondering if anyone else has heard about this? She said that the secret about this project is pretty much been outed by other people on the project, so maybe others have heard about this from other sources...

[ Note - edited to delete details, to protect my source's anonymity. ]


[edit on 12-12-2007 by Buck Division]

Yes AI is extremely interesting..I will check your link out. I am interested in I ching..I always wanted someone who is into it to sort of help me get started. Is your friend on ATS?

ah, i overlooked the year. interesting, but since it's pretty much all speculation, over 30 years ago - we'll never really know.

reply to post by BlackProjects

---

She's not a member of ATS, and now I am afraid to invite her -- I think she would be upset with my previous post about her

Regarding I Ching -- This whole AI subject was revealed to me as part of a discussion with her as to why she thinks I Ching works. She is convinced that I Ching is valid! Her discussion is filled with hardcore Information Theory references. The way she backs things up is quite believable. It makes sense, because she is definitely into the mathematics of randomness, which (to my limited understanding) is the basis for I Ching.

BTW -- I was just now researching the original project, which she called the E.G.G. project, and found an amazing hit! This project is also called the "Global Consciousness Project" -- search for it in Google. I was astounded by this!

I can see how this AI story might fit now. Has anyone else ever heard of the "Global Consciousness Project"?

Her story just got a lot more plausible to me.

Sorry, but my story is by hearsay. I wish I was an engineer in NASA who actually participated in the experiment. But, those were the details my friend gave me.

If the experiment's results were true as he said they were, what could it mean? Ghost in the Machine? What could this mean about our computers now or the internet? Could something be "learning"? Sounds science fiction. But I've had years to think about this story.

I wonder if we can just program a software that runs a simulation of this experiment. The conditions will not be the same. But, perhaps the experiment may come up with similar results. Doubtful.

In an off subject, the same guy from NASA told me a story of the experiments they did on cockroaches. They placed them into enclosed environments. Environment A was similar to Earth's conditions. Environment B was made similar to Mars. The results state cockroaches thrive a lot more on Mars than Earth. Creepy.

i put a very simple test in ~5 min to see if the mice "learn"

assumptions:
if the mouse reaches a deadend, it automatically returns to the previous step
if the mouse reaches the last step, it reaches the exit
you can't have a right step followed by a left step and vice versa, so a forward
must follow a right/left step

requirements:
java version "1.6.0_03" run time environment

unzip the file using winrar/winzip or whatever is your favorite extract utility iz
open the console and navigate to AITest/dist/

e.g. assuming you put and extracted the file to desktop


usage:
enter the size of the maze in steps or enter 0 for a 5 step default one
the maze will be automatically generated
enter the number of trials

tested on linux

sources under /src

rapidshare.com...

too simplistic I know but maybe later i'll make a more complex one

Thanks for the program DarkSonic.

I should have added one important detail to my initial post. The programming should meet the following pseudo-code:

Go forward
if hit wall then
randomly turn 90-degrees left or right
end if

No other conditions or behavior should be added to the mouse. The programming is something I would like to preserve. If it is not too much of a bother, could you modify your program?

Deadends and logic for turns should be omitted for the sake keeping the original experiment's parameters. One day, I would like to try with an actual robot made from spare parts. For your program, I intend to run it for at least a two weeks long.

Who knows the military robots being deployed and developed may start behaving "peculiar" after being in constant-action for long periods of time.

requirements:
java version "1.6.0_03" run time environment

unzip the file using winrar/winzip or whatever is your favorite extract utility iz
open the console and navigate to AITest/dist/

e.g. assuming you put and extracted the file to desktop





changes:
>modified to implement the algorithm in the post above this post
>added the option to select whether to run the test for a certain # of trials
or run the test until the specified date/time.
>added file output for every x number of results for the second option
>optimized some parts for memory management
>added exception handling to some of the code.

the assumptions that still remain:
>if the mouse reaches the last step, it reaches the exit
>you can't have a right step followed by a left step and vice versa, so a forward
must follow a right/left step to make the maze more simple and avoid CIRCLES and OVERLAPPING.
e.g:
RIGHT RIGHT RIGHT RIGHT

NOTES:
>for the time option, it is accurate to the minute rather than to the second
>you SHOULD NOT run the program for something like 2 weeks because you
will probably get 1000000000000000000000000000000000000000000000 results
>for the time option, the program will write every 10 million results produced
to a new file AIresultsx.txt, where x stands for the extension (e.g. AIresults2.txt, AIresults3.txt etc). The created file will be in the same directory as the proggy
>if the results<10 million and the test ended, it will write the results to file anyway
>the program must be have write access to the folder its folder so it can write the
file to disk.
>you can copy/paste or somehow import the results into excel and such, then graph them .

typically it should be a bell curve

, but if the mice actualy learn then it should be a decreasing curve or line


>don't use something like -9989 for the year and such, because not all of the variables
are checked and we're making the assumption that the people who will use this are sane
>sources are in /src
>tested on linux

FILE:
rapidshare.com...

reply to post by DarkSonic

---

Again, DarkSonic, thank you for this program. I will run it on an old machine that can support the correct Java version. There shouldn't be a problem finding one.

Hopefully, someone or myself can replicate as much as we can the experiment with real mouse-bots. But funds are a bit short. Luckily the advent of personal computers allows us to accomplish so much.