Pharyngula

Posted by: Eamon Knight | July 30, 2010 10:11 AM

Genetic Programming and Evolvable Machines

Mice try. That has to be the cleverest filter-avoidance euphemism for "penis enlargement" I've yet seen.

Posted by: theswede | July 30, 2010 10:27 AM

Good catch, thanks! I'll be taking an AI course this fall to get some advanced credits, so this'll be nice bedside reading before that starts.

Posted by: Adamvs Maximvs | July 30, 2010 10:35 AM

Sweet! Thanks for the link PZ, it's been a while since I've read bio journals, so it's nice every once and a while to broaden the horizons a little bit.

Posted by: Nic McPhee | July 30, 2010 11:57 AM

It's definitely a "weird-and-wonderful" computing journal :-). Here "genetic programming" is about evolving computer programs to solve particular problems, so essentially evolution as a machine learning/artificial intelligence/software development/problem solving tool.

Posted by: HertfordshireChris | July 30, 2010 1:51 PM

Climbing the Mount Improbable of Evolving Programs

I entered the computer industry in 1965 and retired from research into the human/information interface over 20 years ago. Recently I decided to look at recent developments and have examined several of these papers. May I suggest that as an “outsider” I may be able to get a glimpse of the wood, while everyone seem so involved with the trees that they are blind to some of the underlying assumptions and ignorant about humans tendancy to mould their actions to the established technology however ill-designed – for instance the QWERTY keyboard.

My conclusion is the whole academic research world seems to have been brain-washed into thinking that the “stored program” concept is a good starting point for research into intelligence and its evolution.. The whole philosophy of the stored program computer requires there to be an intelligent designer to write the program – which is unlikely to be as intelligent than its designer. If intelligence is the result of an Intelligent Designer (and I understand many of the people working in this field reject the religiously inspired view of human intelligence) why is everyone using as a starting point for their models a technology which requires one?

If intelligence evolved without the aid of an Intelligent Designer using intelligent designer style tools to try and find out what happened is like trying to climb the Intelligence Mount Improbable. Bearing in mind that evolving a personal information model of the world is so simple almost every child can do it one wonders how many millions of man-years have been spent trying to climb the most difficult rock face, when there must be a simple route round the back.

Part of the problem seems to the development of the stored program computer meme into what can be considered to be an all-powerful religion – which must be right because it delivered the goods. From the first working systems there was a mad stampede to make money from the electronic calculating system technology, and there was no time for any blue sky research into whether a system based on a liner store of numbers and an intelligently designed program was the best way to get a human friendly information friendly machine. Getting something to work on well-defined applications was the priority and the only human factors research was to develop better tools for the programmers and to modify the processor hardware to make programming easier. By the mid 1960s some genuine human factors research was starting – but only to find out better ways to hide the unfriendly interior of the black box. By the 1970's schoolchildren were being brainwashed (sorry I mean taught) computer programming – and virtually everyone knew that if you were very very clever you could become a high priest (sorry – I mean programmer or systems analyst) and be one of the oracles in charge of the mysterious black box. Since then the black boxes have become more powerful, and many very useful specific tools have been developed – but for most of the people that use, for example, a modern word processor, the human has no real understanding of what the black box is doing – while the black box has even less understanding of the meaning of what is being typed.

By now virtually every science researcher takes it for granted that if you want an electronic box to do something useful it MUST be programmed – and of course this observation is true for the stored program computer.

But who did the blue sky research into alternative electronic processors that might be inherently human friendly? In the 1940s, 50s and 60s everyone was too busy in the rat race to exploit the technology – and by the 1970s the religion (sorry I mean technology) was so deeply established that anyone questioning the underlying principles was deemed to be committing sacrilege – and research funding applications would have to get by the now firmly established body of high priests.

The problem I was finding in the 1980's was that trying to get people thinking about viable alternative solutions which did not need a significant program was almost impossible. Even in areas such as neural nets and genetic programming, most of the research seems to be carried out by people who mentally take the “stored program” philosophy for granted. Even if this does not apply to the key idea it creeps in through the highly formalised approach they take looking for a solution.

I don't expect you all to agree with my arguments but I will leave you with three blue sky questions?

Do you have any evidence that the stored program philosophy (i.e. a number-based system which requires an intelligent designer to write programs) is the best foundations for working symbiotically with humans?

Bearing in mind that humans intelligence pre-dates modern civilization, what relevance do you think the stored program philosophy with its intelligent designers (and the associated highly formalised research techniques) has to do with the evolution of intelligence?

If intelligence is so straight forward thata child can “evolve” its own individualised world model from scratch, why do you think it takes so many people so much effort to make so little progress?

Posted by: jcmartz.myopenid.com | July 30, 2010 3:54 PM

Grab those pdfs while you can!

Posted by: John Morales | July 30, 2010 8:35 PM

HertfordshireChris, thanks for your comment.

I found it thought-provoking.

--

For one thing, I immediately thought about why self-modifying code is strongly discouraged in production systems (here be monsters!)

(Yeah, I know it's not relevant to your main point.)

Posted by: Nic McPhee | July 30, 2010 11:05 PM

@HertfordshireChris Without getting into all the issues that you raised, one of the cool potentials of evolutionary computation and genetic programming is the ability to work outside the box. Several examples are in Koza's article on human-competitive results. A very cool example from nearly 15 years ago was evolving circuit designs for unclocked circuits. Nearly all human designed circuits use a clock to allow the wire voltages to settle down before being read by the next logic gate. Adrian Thompson did some cool work evolving circuits that didn't use a clock, and ended up depending on some weird electronic field interactions. I could rattle off several other examples (Lee Spector's nifty work on evolving programs for quantum computers, Jason Lohn's evolution of satellite antenna designs), but it's clearly possible for evolutionary systems to develop solutions for (very) non-standard systems.

Posted by: HertfordshireChris | July 31, 2010 12:15 PM

The idea of my post was to stimulate thinking about the evolution of intelligence - and I am very interested to see the comments it has caused.

I would agree that the computer is currently successfully used as a tool for a wide range of tasks, as is the QWERTY keyboard. The QWERTY layout remains with us because to many people have modified their way of working to conform to the layout for it to be economical to change - and we all use it regularly without continually thinking "there could be something better". There are many tasks which involve formal mathematical models for which the stored program computer (with suitable software) may well be ideal. There may be a theoretical question as to whether the stored program computer, or a particular piece of application software, are the best for the job but if the users are satisfied, feel that they are getting good value for money, and can see not better alternative, they are likely to continue as at present - and good luck to them.

However real problems can occur when there are real humans, with ill-defined real world problems in the system. We have seen in England with the attempts to get a universal online medical records system, involving real patients, real medical staff with real medical problems, and continual changes in medical remedies, diagnostic equipment, and even the redefinitions of diagnoses.. I was interested in this kind of real world human related information processing problem before I retired - to see if there were alternative more human-friendly processor architectures.

#13 melissa.b.elliot

We have wonderful symbiotic relationships with animals but we cannot get them to do EXACTLY what we want at all times, for example.

I note you emphasise the word EXACTLY. If you cbn exactly pre-define a complete task it may well be sensible to write a program to do it. If animals (or human beings) do not do exactly what you expect it is your fault for having unrealistic expectations ...

#14 John Morales

I found it thought-provoking.

Thanks - That was the idea.

#14 John Morales   I immediately thought about why self-modifying code is strongly discouraged in production systems

#15 melissa.b.elliott  and is strongly reminiscent of virus behavior????.... .....

Well if we think of how a child's memory evolves as it absorbs new ideas we all know that they can absorb intellectual viruses and end up as, for example, fervent young-earth creationists. Any successful model of the evolution of human intelligence must allow for individuals whose view of "logic" and "the truth" differ widely from our own. We must not assume that because all humans have a nose in the same place and with a roughly similar shape therefore all adult minds (particularly across different cultures) will be equally similar.

#16 Nic McPhee

A very cool example from nearly 15 years ago was evolving circuit designs for unclocked circuits. Nearly all human designed circuits use a clock to allow the wire voltages to settle down before being read by the next logic gate. Adrian Thompson did some cool work evolving circuits that didn't use a clock, and ended up depending on some weird electronic field interactions. I could rattle off several other examples ...

I think it is appropriate to stand back and think what one is trying to do.  To take a simplified view the clock in a computer processor is there because it is a single complex processor working sequentially and it it is important that the operation of fetching program instructions and data from the memory, processing it and returning the results to the memory, are carried out as fast as possible. I am unaware of any research which demonstrates that the brain works by streamlining information through one (or perhaps more) complex "central" processors. To me it seems far more likely that the information remains "resident" in the relevant nerve cells, which work in parallel, with some kind of locus of attention moving between cells using a far simpler processor architecture - perhaps repeated in every cell.

This comment is not in anyway intended to caste doubt on the validity of your techniques for solving electronic engineering problems. What you are talking about may well be generally considered as Artificial Intelligence, with the emphasis being on Artificial. I doubt its relevance to the problem of designing human-friendly computers or the the problem of understanding human intelligence and its evolution.

#17 Haruhiist

Basically, communication (interface) that comes naturally to humans is too ambiguous to be very useful to use as a tool.

How insulting to humanity. The logical conclusion to what you say is that if two humans communicate they are too ambiguous to get anything useful done. The point I am trying to make is that there are many applications which require humans in the loop and in order to handle these you need a system with which they can easily interact. If we can understand how human intelligence evolved we have a better chance of building tools which will interact with people, warts and all.

You assume that because a child can do it, it is an easy task. This is not necessarily true: children are learning machines, and the human race has evolved such that its learning mechanisms have 'intelligent' features built in at birth. We know nothing that is better at learning than children, and they obtain their intelligence by learning constantly, for years on end.

I won't answer your point directly, although I have my views on the matter. Instead I will tell you a real life story with a moral.

In 1980 I was faced with introducing a first year undergraduate class of 125 onto the first fully interactive teaching computer in the university. It had only been installed a few weeks before and I had effectively no technical support. About half of the students were unfamiliar with a typewriter keyboard; about a third had very limited contact with a computer at school; about a dozen had taken a final year school module in computer science, and four had their own personal computers. It was quite obvious that those with computer experience would be busy boasting of their skills, showing off, and to some degree intimidating, the ones who had never seen a computer before. So what kind of first assessment could I set which would be fair to all?

I actually took a small demonstration program which had been prepared for a television show where an 11 year old had asked that a computer would help him with his homework.  The coursework was to run the program (which included 7 or 8 different parts) and write an essay on how suitable they thought it would be for an 11 year old. 124 out of 125 students successfully ran the package and some of the incipient computer junkies who thought computing was all about showing off by writing programs in BASIC had the most obvious difficulties.

There was a sting in the tail - most of the packages had a hidden feature which few of the students found. In the debriefing I explained that we were all getting older and that we became more knowledgeable but less imaginative with age. The things most of them missed would have been found by a  bright 11 year old but that they, at 18 or 19, had reached a stage where academic learning was replacing imagination.

The moral - the older we get, and the more exposure we get to formal academic teaching, the harder it is to think ourselves back into the position of a child who starts surrounded by unknown unknowns, progressed to known unknowns and unknown knowns - finally reaching the pinnacle of known knowns. There is a very real danger of assuming that human intelligence is brilliant because we know nothing better. I note that much "intelligence" research starts with a simple (i.e known know) situation and/or take a view that one will find a clue by trying to model tasks which university academics find difficult. Yet such tasks are far removed from the starting point as seen by a new-born child.