Evolution Cubed? June 1994

Evolution has been going on a long time, and in that time it has got pretty smart. The question is, just how smart has it got, and how smart is it going to get?

In his book 'The Wisdom of the Genes', Christopher Wills suggested that the mechanisms of evolution had themselves 'improved' so that evolution could work 'better', or faster, or something (Richard Dawkins also gave a paper on the 'Evolution of Evolvability' at the Artificial Life bash at Santa Fe, but I haven't read it - those proceedings are too darn expensive for an amateur like me) . From memory, Wells cited such examples as 'jumping genes', which are DNA base sequences which can copy themselves up and down the Chromosomes, rather like computer viruses can now copy themselves round and round the memory / disk environment.

The traditional evolutionary view was that mutations, or changes in the DNA base sequences, arise only from dislocations caused by high energy particles (or some other physical shrapnel), or by errors inherent in the DNA copying mechanism itself. (You might think that the 'perfect' copying mechanism would have zero defects, but since this would seriously inhibit evolution, then the 'evolutionary perfect' mechanism must have some low controlled error rate). The traditional view is being extensively modified as more complex effects in DNA expression are discovered. It turns out that 'copying' is a pretty loose term for the way DNA base sequences are processed (that's probably a better word); some 'missing' sequences are inserted, corrections are made, all sorts of stuff is going on. It sounds more like the low levels of the OSI data communication model, with built in data compression at that. It is also quite possibly (or even probably) true that other agents are responsible for the insertion of base sequences into the genome, such as viruses or viral like vectors. The point that Wills was making is that all these mechanisms haven't arrived by accident, they have themselves evolved, along with the other genes in the genome that are expressed in the Phenotype (things like arms, eyes, blood cells and all the other bits).

It doesn't surprise me greatly that this kind of 'second order' evolution should happen, this kind of 'gear shift' is common to many other aspects of evolving dynamic systems. For example, mathematics has often made steps forward by shifting from, for example, first order predicate calculus to second order. The kind of shift I'm talking about is when some process or mechanism (like a mathematical function) can start processing others of its own kind, instead of just processing simpler creatures. This kind of Godelian self swallowing trick is also what enabled great steps to be made in machine tools (which can make other, better machine tools), and even in computers (I know all about that, because I worked in Computer Aided Design before anyone used it - it was fun in those days). You have to be a bit careful with the word 'order', because it can mean different things - a second order equation isn't quite the same kind of change (I think). The logic theory people have this horrible word 'arity' which might be closer, but I don't like it, so I'm not going to use it.

Just as a car has more than two gears, there is no reason why this trick can't be carried out more than once, though it does get a lot harder. In his book 'Towards an Ecology of Mind', Gregory Bateson talks about human learning, and gets all metaphysical trying to relate it to Bertrand Russel's theory of Logical Types. The thing that interested me was that he produced curves which showed that humans not only learn , but they definitely learn how to learn (in other words, their learning rate improves), and they probably learn to learn to learn, maybe learn (^4), and some really smart guys may even learn(^5) - sorry about these ^'s, but I've got to abbreviate somehow.

This ties in with some stuff I read ages ago from another Cyberneticist whose name escapes me, but who had a little table showing 'levels of intelligence'. A simple feed back servo was level 1, a more complex control mechanism was level two. Humans came in around level five, and he went up to seven or eight (so when those neural networks really get going, we'd better watch out).

I also remember reading somewhere that people go on learning indefinitely, the progress just gets slower. If you measure someone doing a simple repetitive task in a factory, they may improve by a factor of two in a year, they will then take ten years to improve by another factor of two - in other words, it's a log scale. No one has yet found out if you improve another factor of two in a hundred years. Seriously, though, I recently saw Sir John Harvey-Jones (ex head of ICI) on BBC TV in a pottery; there was a woman there whose job it was to stamp plates with a rubber stamp (I kid you not). She was obviously a star turn, and ran through a stack of plates so fast you could hardly see what she was doing. It turned out she had been doing this job for about eighteen years (or some such), and could lick any machine into second place.

So what has all this got to do with evolution and genes, I hear you ask (well, you should). I am suggesting that there is a natural sort of progression in any evolving dynamic system (I won't get bogged down here with what that is - does Gaia evolve and all that sort of thing). This progression involves a gradual shift up in 'order' every now and again, and I would expect these shifts to come on a log time scale. Nope, I haven't got any other evidence, it just seems kind of plausible to me. As I'm not a scientist, I can hide behind my quite unjustifiable faith in my own intuition. Perhaps I should add that I don't think these kinds of shifts are the same as 'phase changes' that also emerge naturally from such systems.

So the question is, how far has natural evolution got? Are we at level two (as Wills was suggesting), or even higher? This is not an easy question. I guess you can make out a case that we are currently in the middle of a 'gear shift' - from Genetic Engineering. Dawkins would say that our genes, having created very effective carriers, have now evolved them to a state where they can start tinkering with the genes using mechanisms totally outside the body (an example of Dawkins 'Extended Phenotype'?)

Even then, its not clear whether there are other 'third order' mechanisms. I don't think we know enough about the mechanisms yet to answer that. Its not as simple as just getting the raw data either, its a question of how its interpreted. How do you know what order a system has reached? This is tricky stuff, the answers are possibly to be found in areas like Chaitin's complexity work, I don't know enough about that, but roughly speaking it says that the complexity of something is the minimal description of an algorithm that can generate it. This still doesn't cover the way in which the algorithm itself is 'coded' - does it have subroutines? are they nested, recursive? This is the kind of stuff computer scientists are struggling with when they consider the complexity or difficulty of a software system.

The connection between these ideas and the whole evolving (!) field of the study of non linear dynamic systems is obviously close. I am currently reading Lewin's book 'Complexity', but its mainly about Artificial Life, not Chaitin's work. However, I originally started thinking along these lines when I got into the whole Evolution / Chaos / Alife area. I certainly think that Alife (and Genetic Algorithms) will teach us a lot about how evolution works. It may be that the only way we can answer questions like the ones I pose here will be by simulation rather than investigation - although of course the two go hand in hand, much like pure mathematics and physics.

In particular, I like the work Tom Ray is doing with Tierra, and I think it has great potential for answering some of the tricky questions. Tierra consists of a virtual machine environment wherein software 'animals' can be born, compete, reproduce, mutate, and die. The crucial element is that the means of 'evolution' must be able to evolve itself in order for it to increase in order (if you see what I mean!). I think Ray has many of the pieces in place for this to happen, but possibly there needs to be more changes yet. For example, I think it may be necessary for the evolution mechanism to be able to change the 'mutation rate' that controls how fast changes can happen. I don't see any reason why this can't be done - although it probably implies keeping some 'data' with Ray's animals, which he has so far avoided. (I am suggesting each animal has its own mutation rate embedded, and so liable to change, but otherwise passed on to its descendants). At first I assumed that the 'shrapnel' mutation rate should be an outside constant - after all, we can't change the rate at which the sun fires nasties at us - but maybe even this should be amenable to evolution, because we can 'control' the amount of pigment in our skin. Also, I can't help feeling that subroutines have got a large part to play, after all, its the introduction of subroutines, and nested/recursive ones, that gave computers a 'gear shift' (or maybe it was Lisp, OK you AI guys).

Anyway, I'm tired of writing, and you've probably read enough to get the idea. You can read about Tierra in various documents on CompuServe (Cyber Forum).

Next Essay

Back to List of Essays

Back to Home Page