Evolutionary algorithms and the Lovelace Test

This article is interesting. It seems AI computer boffs needed a better test than the Turing Test for Hard Artificial Intelligence, should it ever arrive, given the subjective and easily manipulable results of the Turing Test recently. The original publication is here. As you’ll see, the test involves demonstrating a computer outputting something that was not designed into it in the original program – and should that ever happen, the computer will be shown to be truly intelligent, according to those best qualified to say.

But hold on a minute – aren’t evolutionary algorithms already producing new information that wasn’t input in the program, thus confirming the ability of simple organisms to develop indefinitely by natural selection? If so, it seems odd that the computational experts haven’t spotted it and raised the roof about the arrival of AI. Nobel Prizes should have been awarded.

The truth is, of course, that no evolutionary algorithm actually has the ability to generate genuine novelty. Kurt Gödel was absolutely right back in 1966 – natural laws are too lacking in information to generate life in an algorithmic manner, or any computer could simulate it, and it would be an example of hard AI.  GIGO still rules, OK? And that means, as Robert Marks has pointed out, either that the algorithmic resources of evolution are greater than the probablistic resources of the entire universe (and a good number more) so that the impossible becomes possible, or that evolution is a process beyond algorithmic computation (as was suggested by Roger Penrose for the human mind, for example), which makes it anything but a simple scientific theory.

More on the last in a subsequent post.

Jon Garvey

About Jon Garvey

Training in medicine (which was my career), social psychology and theology. Interests in most things, but especially the science-faith interface. The rest of my time, though, is spent writing, playing and recording music.
This entry was posted in Science. Bookmark the permalink.

14 Responses to Evolutionary algorithms and the Lovelace Test

  1. Lou Jost says:

    “The truth is, of course, that no evolutionary algorithm actually has the ability to generate genuine novelty.”

    Genetic algorithms regularly produce novel solutions that the human programmers did not know. An example is the satellite antenna design produced by NASA’s Evolvable Systems Group (http://www.nasa.gov/centers/ames/research/technology-onepagers/evolvable_systems.html). The programmer does need to be able to evaluate a solution’s success at achieving some goal. But he or she does not have any idea about what those solutions will be. In the case of the satellite antenna design, the photos on the linked page show a design that can only be described as bizarre, and nobody in their right mind would have thought of it. But it works well. In what sense is this not the production of novelty?

    This does not mean the computer is conscious, of course. That problem seems to have little to do with evolution.

    • Jon Garvey Jon Garvey says:

      The question is not whether the programmers knew the answer, but whether the answer was inherent in the program, and (as the original Bringsjord et al paper says) could be reproduced by the programmers manipulating symbols.

      If so, the information was inherent in the program: if not the program has produced an original idea (and so your evolutionary algorithm would pass the Lovelace Test and win its originator fame and renown).

      Darek Barefoot’s theistic conception of evolution, I believe, holds that God put all the information necessary for life into the universe at creation. Theoretically it could compute all that we see – but the information would be inherent in the creation equations.

      Conceptually Darwinian evolution is different: brand new information (in the form of new lifeforms) are formed bottom up from blind primary forces. Therefore, unlike the “frontloaded” version, it must be creative of complete novelty. And to emulate it a computer algorithm must be creative too, in the terms that the Lovelace Test suggests. Their section on oracles echoes what I said in the previous thread.

      The authors suggest that, perhaps, only a conscious intelligence would have the kind of creativity necessary to pass the test, but that is a secondary issue: creative novelty is an empirically measurable thing: consciousness is not.

      Now if you’re right, and evolutionary algorithms genuinely create information not inherent in the program, that’s fine: I’m not an authority on computing. But I will still prefer the authority of those who are.

      • Lou Jost says:

        “if so, the information was inherent in the program”

        No, as we discussed before about biological evolution (and I thought you agreed, and so did the physicist whose video you showed in your top-down causality post–he almost repeated my words exactly) the information is not inherent in the program, but rather comes from the environment.

        The Lovelace test has nothing to do with evolution. The fact that all the steps of evolution could be retraced is completely irrelevant to the question of whether evolution can produce the kind of novelty that we see in real life.

      • Lou Jost says:

        “Now if you’re right, and evolutionary algorithms genuinely create information not inherent in the program, that’s fine: I’m not an authority on computing. But I will still prefer the authority of those who are.”

        I am not an authority on computing either, but I have written genetic algorithms and have worked as a consultant in artificial intelligence for the Computer Science Dept at the University of Texas decades ago. I’m not saying the authorities you believe in are wrong. I am saying the question they are addressing is irrelevant to arguments about evolution. You do not need to be an expert to see this.

        • Jon Garvey Jon Garvey says:

          Lou the logic is this: if computer algorithms are successfully modelling evolution, it’s because the process is capable of being computed, ie it’s a mathematical process (thus you can’t model composing a symphony or the history of America on a computer).

          If it’s infeasible for computers to model genuinely creative processes (as per Lovelace Test) then either (a) the process being modelled is not genuinely creative and the information is inherent in the system or (b) the computer is not genuinely modelling evolution.

          If you say the information is provided by the environment, I say that that is an insufficiently coherent source of information, and in fact the organism has to “match” it by organised changes to its form (variation), which is the real source of the information.

          Otherwise it would be like saying that the properties of the atmosphere and gravity provide the information that makes a Boeing 757 fly. In fact the environment merely supplies the parameters which some kind of teleological process apprehends and organises into a vehicle that can uitilise those properties. It is life that the algorithm models, not the environment.

          • Lou Jost says:

            Let’s suppose, for the sake of argument, that something called “genuine creativity” cannot be produced by a mathematical or computable process. In order for your conclusion to follow, you would still have to show that evolution produces that kind of “genuinely creative” result, rather than merely the kind of novelty exhibited by my antenna example. But the products of evolution are indeed like the antenna; you have given us no reason to think that they are uncomputable. On the contrary, since we could in principle retrace their generation step by step, they are computable.

          • Lou Jost says:

            “…the organism has to “match” it [the environment] by organised changes to its form (variation)…”

            No, in genetic algorithms the variation can be random with respect to fitness; the environmental screening process creates nonrandomness in the population. Natural selection provides a mechanism by which aspects of the organism become correlated with the environment.

          • Lou Jost says:

            In relation to your Boeing 757 comment, it is interesting that the blog-o-sphere claims Boeing has used evolutionary algorithms to help design some of their wings…..

      • Jon

        A technical correction. If there is no hidden variables solution to the quantum measurement problem (as there does not seem to be) then the collapse of the wave function might be interpreted as an introduction of information–depending on exactly how information is conceived.

        • Jon Garvey Jon Garvey says:

          Darek

          Ian Thompsom’s your best man on that question. But I would say that you’re right.

          In fact Robert J Russell (Cosmology from Alpha to Omega), one of the better science-faith academics, suggested that God’s direction of quantum events would be a way of his governing nature without the indiscretion (as far as the “science-faith community” is concerned) of “interfering with nature”.

          That assumes, of course, that quantum events have no cause within nature, so God is allowed to add a cause from outside. It also assumes, with some justification, that individual quantum events, by making a difference to molecular mutations, could produce a result in evolution over and above the statistical.

          My objections are not that this couldn’t happen – I’m pretty sure it must, as theologically quantum events cannot be exempt from providence, but that it assumes that God is subject to the very laws that science says he has made inviolable. He himself has made no such concession to nature’s inviolability.

          That’s one of the things that first got me asking why one should commit oneself to a view that God works predominantly or exclusively through natural laws at all, since that view is in itself merely a product of 18th century Deism.

          And that in turn led me to prefer (no stronger) a view of God’s action both through regularity and choice contingency.

          At the same time, I assume some version on methodological naturalism in assessing the sufficiency of mechanisms of evolution, such as their computability: if a proposed means appears not to work mathematically, either a better mechanism or a non-mechanistic route must be involved.

        • Lou Jost says:

          Darek, a small correction to your correction: only local hidden variables have been ruled out. Nonlocal hidden variables are possible. Also, as you know, there is a solution to the measurement problem which do not involve a collapse of the wave function: the Everett-Wheeler or Many-worlds interpretation, which in fact is the most popular current interpretation among workers in the field, according to a recent informal poll. In spite of its apparent strangeness, it does seem to be the most direct interpretation of what the equations are telling us, just as the mathematical form of the solutions to Dirac’s equation suggested the existence of the bizarre concept of antimatter.

          (Dirac’s equation, by the way, is engraved on his grave in the floor of Westminster Abbey, the only equation in the whole place.)

          • Lou

            Unlike antimatter, the infinitely proliferating branches that allegedly fill out the Schrodinger equation are, I believe, in principle unobservable. I’m not a theoretical physicist and may be wrong about that so I’m more than willing to be corrected. In any case, there is still the question of what picks out the particular branching path our own world takes, isn’t there?

            • Lou Jost says:

              Darek, that’s the beauty of the interpretation: every world is equally real, and there is no picking out of our particular world. Every world has its observers.

              You are right that Dirac’s prediction was easily observable, and that the many-worlds interpretation of the Schrodinger equation is not like that. I’m not sure about this, but perhaps one could consider interference effects as the effects of the alternative worlds on each other before an observation forces them to split off. For example, in the famous double-slit experiment, the full interference pattern remains even if the density of photons is so low that only one of them is going through the apparatus at any given time.

              • Lou

                The subject is more complicated than it first sounds, I gather. However, there seems to be no practical experimental test on the horizon that could favor MWI over CI. Both have their own interpretive ways of accounting for superposition, etc. Also, there are varying opinions over MWI realism–is MWI literal truth or the best representation the human mind can make of a deeply mysterious aspect of reality?

Leave a Reply