Master of AI

John Creighton · 2004-08-21 16:01:22

There has been numerous discussions about artificial intelligence. The concept being that we send robots there first to create the industry and then move the humans into a relative luxury. Clearly many will challenge the possibility of this scenario and I am not advocating that we should wait for this technology. However, we should use what ever tools we available from tellerobotics to artificial intelligence. The propose of this thread is to discuss how to create self sustaining industry and human outputs with as little with as little human input as possible.

John Creighton · 2004-08-21 16:10:52

Some objectives in creating self sustaining industry are to reduce mass and complexity of the initial seed, reduce the replication time and allow a large diversity of industry to flourish. The full fledge industry and the initial seed is composed of a number of robots. Tradeoffs must be made between versatility and simplicity. The simplicity can often be improved by reducing the complexity of the problems each robot must deal with. For instance there may be roads made by one robot to reduce the number of bolder the other robot must traverse. Another example is each tool may have a bar code placed on it so all robots don't need advanced patter recognition capabilities. I think this is where tellerobotics may come in. In a simplistic system the human may tell the robot what bar code to place on each tool.

John Creighton · 2004-08-21 16:13:56

Taken from ]http://www.newmars.com/forums/viewtopic.php?t=2081]the need for moon direct

An AI emulating the thinking process is basically a series of "If then or else" commands along with other types of commands to run our robot that can fix things. Now that robot can only operate with the perimeter of that AI emulator to fix something and anything that out side that AI emulator it can't fix. Just to drill a simple hole on an NC/Machine it might take ten to twenty commands to drill that hole or lines of code and this is on a fix location too.
But, back to our NC/Machine when they proof a tape to do a new job which may have hundreds of thousand of lines of code, which I admit is one of the bigger programs that they write to do a job and it on a fix perimeters too or boundaries. It will generally take several months to program it and take it out to the machine and run and modify the program and run it again until they get right. Your probably talking about a tenfold increase in lines of code for a mobile unit to do the same thing and that just one project too. Your going to have thousands or tens programs to do different things. So you will have to add that to the list of programs that you have to keep on line.
In AI terminology a program that solves a problem by a large amount of human input is said to have a low A to I ratio. Read http://www.genetic-programming.com/book/]Genetic Programming IV: My “Routine Human-Competitivve Machine Intelligence” John R. Koza, Martin A. Keane, Mathew J. Streeter, William Mydlowec, Jessen Yu & Guido Lanza
Your suggestion of program composed of hundreds of thousands of if then statments is such a program. Another example of such a program is the Chess program big Blue. Although the program learned how to play chess very well it was the result of many man years of human input. If we can solve our AI problems this way that is fine. It works and won’t be two difficult to program if there is a limited number of cases.
“Now that robot can only operate with the perimeter of that AI emulator to fix something and anything that out side that AI emulator it can't fix.” Well if this were true it wouldn’t really mater to much. The robot would just have to waits for the upgraded to be downloaded from earth. However, this style of programming is not the objective of artificial intelligence.
The goal of artificial intelligence is to obtain a high A to I ratio. That is the program can solve the problem with as little human input as possible. Genetic Algorithms are one example of algorithms that have a very high A to I ratio. They can solve a large class of problems with very little human input. All the user has to specify are the objectives. The algorithm then figures out how to solve the problem. So for instance if the objective is to screw in a bolt all the user tells the program is it is bad to bump into things, It is bad to cross thread, it is good to be fast. The program would then simulate several algorithms (thinking) until it found a good enough algorithm (stopping condition) to screw in the bolt.
Thus in a primitive AI the user could write many IF then statements. However in more advanced AI the human first designs the architecture and then decides the objectives and lets the computer program take care of the rest. BTW the computer thinking can take place both on the mars and on the moon. Genetic algorithm have the advantage of being highly parallel. If each robot had enough free processing power they could work together to solve common problems. Anyway, I think I am going to start a new thread.

Grypd · 2004-08-21 16:44:34

To program into a robot enough knowledge to have the intelligence of say a dog and a stupid one at that would take about 500 people 500 years. Not to mention they would run out of room in the robots brain to fill in all this code. The alternative is called artificial learning where you try through teaching get to the robot to learn similar to babies.

One such attempt was the robot called Cog but though a lot was learned in the attempt it remained that the robot could learn very very slowly unlike the intuitive learning of a human baby. It may be that we are incapable of getting robots to be truly intelligent as we have enough trouble ourselves in actually stating what is thought.

Oh and our robots are morons

John Creighton · 2004-08-21 16:55:37

To program into a robot enough knowledge to have the intelligence of say a dog and a stupid one at that would take about 500 people 500 years. Not to mention they would run out of room in the robots brain to fill in all this code. The alternative is called artificial learning where you try through teaching get the robot to learn similar to babies.
One such attempt was the robot called Cog but though a lot was learned in the attempt it remained that the robot could learn very very slowly unlike the intuitive learning of a human baby. It may be that we are incapable of getting robots to be truly intelligent as we have enough trouble ourselves in actually stating what is thought.
Oh and our robots are morons

In artificial intelligence there is something called learning how to learn. The most fit program is the program that learns the quickest. Clearly a dog is going to learn slowly if only how the dog physically responds to actions is used as the performance measure. Ideally some kind of simulation should be done where many dogs programs are simultaneously explored to find out which dog is the best. We can mix human intelligence with artificial machine learning. For instance we may create a filter whose output tells us if something is an orange or not. Cog could then use this as one of the functions in his final program to decide if something is or is not an orange or Cog could alter this filter to produce a filter that is better for identifying if something is or isn’t an orange.

John Creighton · 2004-08-21 17:02:08

To program into a robot enough knowledge to have the intelligence of say a dog and a stupid one at that would take about 500 people 500 years. Not to mention they would run out of room in the robots brain to fill in all this code. The alternative is called artificial learning where you try through teaching get to the robot to learn similar to babies.

I suppose it all depends on how clever they are but you are just making up numbers. Anyway clearly the machine learning alternative would be used as you describe later. BTW you make it should like that was the only experiment ever done in AI.

Grypd · 2004-08-21 17:25:39

The statement it will take 500 people 500 years is a quote from a magazine article where an expert in robotics had been trying to tell us the layman just how hard it is to make a form of intelligence in robots. And Cog was one of the most successful attempts at learning robots ever created but like everything we are still at an early stage. But with computing power getting better at the exponential rate it is, I wonder if it could be bettered very soon.

The test to prove artificial intelligence is to have a person in one room and the subject in another. The tester then asks the subject various questions and it will be when the tester can not be sure if it is a robot or another person in another room that we will have reached one of the milestones of artificial intelligence. These questions are similar to what is your favourite color? how do you feel about it? We are a long way from that sort of capability. pity!

John Creighton · 2004-08-21 17:31:37

Some Links:

http://www.handshake.de/user/blickle/Tr … html]Truck Control Problem

http://www.genetic-programming.org/]Genetic Programming

http://freshlinks.net/odp.aspx/Computer … Artificial Life

http://www.genetic-programming.com/book/]Genetic Programming IV: Routine Human-Competitive Machine Intelligence

John Creighton · 2004-08-21 17:41:15

1.1 Genetic Programming Now Routinely Delivers High-Return
Human-Competitive Machine Intelligence
Focusing on this book’s first main point (i.e., that genetic programming now
delivers high-return human-competitive machine intelligence), the next four subsections
explain what we mean by the terms
* human-competitive (section 1.1.1),
* high-return (section 1.1.2),
* routine (section 1.1.3), and
* machine intelligence (section 1.1.4).
Then, four additional sub-sections outline the evidence that supports the claim
genetic programming now delivers results with these four characteristics.

Samuel’s statement reflects the common goal articulated by the pioneers of the
1950s in the fields of artificial intelligence and machine learning. Indeed, getting
machines to produce human-like results is the reason for the existence of the fields of
artificial intelligence and machine learning.
To make this goal more concrete, we say that a result is “human-competitive” if it
satisfies one or more of the eight criteria in table 1.2.
The eight criteria in table 1.2 have the desirable attribute of being at arms-length
from the fields of artificial intelligence, machine learning, and genetic programming.

1.1.7 Routineness of the Results Produced by Genetic Programming
This book demonstrates the generality of genetic programming by solving illustrative
problems from several fields, including problems involving
* control,
* analog electrical circuits (including six post-2000 patented circuits),
* placement and routing of circuits,
* antennas,
* genetic networks, and
* metabolic pathways.

from: http://www.genetic-programming.org/gp4c … apter1.pdf

Rxke · 2004-08-21 17:46:54

Grypd,

the Turing test.... I never understood why succeeding that test would be considered a breakthrough...
An 'ELISA' or other chatterbot on steroids could pull this off, IMO, w/o much intelligence.

And... our fixation with human intelligence/self-awareness... Isn't that in many cases overkill for a lot of repititive jobs? Tilden's BEAMbots are w/o central nervous system/CPU, but more like insects, a body that is th system, and they do remarkable stuff... 7 transistors and some servo's beating a robot with a big CPU, thousands of lines of code...

Heck, when I was working on a transportband, I sometimes wished I could've left my 'CPU' at home, the only stuff needed was eye-hand coordination motorskills..

Rxke · 2004-08-21 17:49:45

To add to your list: a program that finds/ correctly diagnoses tumors on scans with a better 'hit-rate' than experienced doctors; programs that go through sci.papers and find new correlations everyone overlooked,....

John Creighton · 2004-08-21 17:51:56

You could add automated theorem proving but I don’t know if that is considered human competitive.

Vir Stellae · 2004-08-22 06:39:24

The man who commmented on the difficulty of programing AI is true. To program a Robot with the intelligence of a man would probably take a team of one million programmers decades to write all of the minutia, that is AFTER hey have R&Ded enough the processes of the mind to actually *know* what to write, which could sstill taake cenuries!

The pace of programming has not keep even with the pace of hardware advances in the slightest.

Rxke · 2004-08-22 08:50:31

Not so, IMO...

I mean, the analogy doesn't make sense, a lot of what we learn goes 'automatically' etc, future AI would be like that too, construct a learning system, largely 'Tabula Rasa,' and feed it data.

Already there are programming environments that are self-learning, and the results are higher complexity-levels than originally programmed.

It will be a process of self-learning, self-'programming'
Like the abovementioned patented circuits, invented by AI's, no programmer programmed those circuits, only gave some rules to play with, and a means to self-assess 'success' of new designs.

Of course, good language-parsing could be one of the keys here... Imagine a budding AI system able to distill meaningful data, that would lead to new 'mental' logical trees, from the gazillion webpages, electronic encyclopedia etc...
A good (intermediary) language would be http://www.lojban.org/]Lojban, as interface to computers, IMO, it is possible to program a system to use Lojban w/o the innuendo's, ambiguity etc current systems have huuuge problems with. Of course, then you have the 'sayings' etc... ("It's raining cats and dogs," is hard for an AI to understand, for instance, because it's database would rule out cats and dogs habitually falling out of the sky, but these pesky humans keep saying that they do... Confusing. )

Lojban is a logical language, English is not, but even the ability to parse languages accurate is not much help if an AI doesn't have an internal image of *what* the world is. A reference.

What is a table exactly? or a stone, a cloud, how do these things relate to eachother, to me, or the AI as entities. What is the difference between you and I etc etc... Hard stuff for 'humanlike AI.'
Note, that children dont make a difference between 'you' and 'I,' initially, everything is 'I.' But how to give an AI a sense of 'I,' and how, once that would be achieved, wean it off the sense that 'I am all,' It would be a clumsy child, but maybe more powerful, if not controlled tightly during its upbringing...

(Sorry, too much coffee, rambling wildly... and widely..)

John Creighton · 2004-08-22 11:31:21

It is kid of interesting if two uniquely developed AI programs are evolved we can bread the programs together to hopefully find a better program. Edit by that way that language Logban sounded cool. I wonder if it would be a fairly dry sounding language though. What about the expression in someone’s voice. Is that included in the language?

John Creighton · 2004-08-22 11:41:50

And... our fixation with human intelligence/self-awareness... Isn't that in many cases overkill for a lot of repititive jobs? Tilden's BEAMbots are w/o central nervous system/CPU, but more like insects, a body that is th system, and they do remarkable stuff... 7 transistors and some servo's beating a robot with a big CPU, thousands of lines of code...

I was thinking about the beam bot concept. A BEAMnot has a number of voltage levels that can be set that control its dynamics. If the beam bot had some kind of sensors, these sensors could be used to identify the model parameters that govern the dynamics of the BEAMbot. Once these parameters are identified the voltage levels could be reset to give the beam bot a desired behavior. There may be several voltage configurations for different modes of operations. Anyway there is a discussion about BEAMbots in the http://www.newmars.com/cgi-bin/ikonboar … ;f=16]Mars Rover folder called: http://www.newmars.com/forums/viewtopic … Simulation of Intelligent Robotic Colony.

Grypd · 2004-08-22 13:33:39

I like BeamBots they have been used by Robot enthusiasts and industry for decades. And they get better each year. But the one thing they are not is intelligent, they may have terrain object avoidance but no real complicated logic problems if they get stuck they just shout for help.

Im currently working on one to mow my lawn it will follow a magnetic wire around my garden and do my most hated garden chore. I could program the dimensions of the garden in and let it go but I believe the way im doing will give the straight lines I want.

One efficient use of Beambots is to provide fire and security services these Robots simply do not get bored while patrolling.

The Turing test is one of the few means we have to test a Robot or computer that would make a non biased way to see if we have an intelligent computer. But it has problems it relies on the skill of the human operator. And frankly I have run into certain people who would have trouble being classed as Human in this test.

edit I forgot to mention my beambot though wheeled had to get rid of the grass it had collected then return back to the point it had stopped at, needless to say im not having too much luck.

John Creighton · 2004-08-22 13:43:57

I like BeamBots they have been used by Robot enthusiasts and industry for decades. And they get better each year. But the one thing they are not is intelligent, they may have terrain object avoidance but no real complicated logic problems if they get stuck they just shout for help.

And of course that is really the point. A lot of tasks don't really require that much intelligence. For instance hitting a ping pong ball requires little intelligence. Trying to fool your opponent might require a little bit of intelligence.

Grypd · 2004-08-22 13:56:22

You do know that when we get a decent high AI capable device it will probably be something similar to the HAL type computers in 2001.

Id thank my makers that im not called Dave

John Creighton · 2004-08-22 13:59:34

You do know that when we get a decent high AI capable device it will probably be something similar to the HAL type computers in 2001.
Id thank my makers that im not called Dave

Unfortunately I didn't see 2001. However, designs should be penalized for undesirable personality characteristics. Probably AI should be designed with as little personality as possible. Unfortunately not all humans will agree with this notion.

Grypd · 2004-08-22 14:14:18

In 2001 HAL was not what I would call exactly a very personable computer, Its voice was monotomous and its simple overiding drive was efficiency. It was only when given direct orders to lie and decieve that it decided that the crew where interfering with the programs goals. It then attempted to kill all the crew.

John Creighton · 2004-08-22 14:30:41

In 2001 HAL was not what I would call exactly a very personable computer, Its voice was monotomous and its simple overiding drive was efficiency. It was only when given direct orders to lie and decieve that it decided that the crew where interfering with the programs goals. It then attempted to kill all the crew.

Well, I guess give the AI a conscience. But then again even people with a conscience sometimes do bad things. I guess any technology can be abused and even without technology we still have rocks. The practice of science must be accompanied with the appropriate ethics. However, more then just ethics are needed. A way needs to exist for the everyday person to discover the abuses of technology, and mechanisms need to be in place to rectify those abuses. It is really funny though. I dream of an ideal place and time that hasn’t yet come but hopefully will, where the truth is absolute and manipulation does not exist.

Rxke · 2004-08-22 14:45:24

Grypd, about your mower...

I know someone who has the huskvarna (sp?) robomower, it doesn't need to collect the grass, because it keeps mowing days on end, silently, unobtrusively. If you mow your grass very repeatedly, there's no need to collect the clippings, they are small, and 'sink' to the ground, added benefit: less nutrients go to waste...

My parent's neighbour, a bit of a nutjob, used to mow his grass every day with a big noisy gasoline 'sitmower' (English? those mowers you can drive, like a little tractor...)
He didn't need to collect it either, and the grass looked really good.

(But we hated him, heehee...)

About the bots being stupid: exactly, but they excel in errr... (sorry, when tired, my English vocbulary goes to bits...)

good in fluid movement, motoric adaptibility, negociating terrain. But they have no brains. Insects.
But you *could* give them brains, classical computerstuff, giving 'metacommands' like: "go to yonder hills, fill your trunk with ore, and when full, return to the refinery".... The Beampart fills in the 'blanks,' the actual way of *how* to go to the collectingplace...

Just like we humans walk, pretty uncounscious of *how* we walk, we don't think 'left foot up, forward, down, right foot up forward down...etc'

Both systems combined can be a very strong platform, IMO.

Grypd · 2004-08-22 14:48:13

John Creighton you are a dreamer, but im glad you are. If it had not been for dreamers mankind would still be living in caves and going uuugh uuugh. But some ancient dreamer went Hmmm if I hit these two rocks together wonder what would happen....Fire.

Robots are still at an early stage in development but to make them intelligent is at an even earlier stage. We can do so much but it can be abused too easily. Everyone knows the Terminator and no one hopes this happens.

http://www.spacedaily.com/news/uav-04zzl.html]Military air power

this is the way some people think robots will go

edit sorry I cant make the link work
edit 2 sorry again managed to get it to work major apologies

SpaceNut · 2004-08-23 06:54:57

While this may not be the start of AI it is in the right direction for Human and Machine interfacing in future missions.
I find it interesting to see the use of a segway for mobility.

NASA Engineers Free Robonaut with Wheels, Leg

http://www.space.com/businesstechnology/robonaut_ugrades_040823.html] story.nasa.robot.jpg

New Mars Forums

Announcement

#1 2004-08-21 16:01:22

Re: Master of AI

#2 2004-08-21 16:10:52

Re: Master of AI

#3 2004-08-21 16:13:56

Re: Master of AI

#4 2004-08-21 16:44:34

Re: Master of AI

#5 2004-08-21 16:55:37

Re: Master of AI

#6 2004-08-21 17:02:08

Re: Master of AI

#7 2004-08-21 17:25:39

Re: Master of AI

#8 2004-08-21 17:31:37

Re: Master of AI

#9 2004-08-21 17:41:15

Re: Master of AI

#10 2004-08-21 17:46:54

Re: Master of AI

#11 2004-08-21 17:49:45

Re: Master of AI

#12 2004-08-21 17:51:56

Re: Master of AI

#13 2004-08-22 06:39:24

Re: Master of AI

#14 2004-08-22 08:50:31

Re: Master of AI

#15 2004-08-22 11:31:21

Re: Master of AI

#16 2004-08-22 11:41:50

Re: Master of AI

#17 2004-08-22 13:33:39

Re: Master of AI

#18 2004-08-22 13:43:57

Re: Master of AI

#19 2004-08-22 13:56:22

Re: Master of AI

#20 2004-08-22 13:59:34

Re: Master of AI

#21 2004-08-22 14:14:18

Re: Master of AI

#22 2004-08-22 14:30:41

Re: Master of AI

#23 2004-08-22 14:45:24

Re: Master of AI

#24 2004-08-22 14:48:13

Re: Master of AI

#25 2004-08-23 06:54:57

Re: Master of AI

Board footer