Reader Comments
Post a new comment on this article
Post Your Discussion Comment
Please follow our guidelines for comments and review our competing interests policy. Comments that do not conform to our guidelines will be promptly removed and the user account disabled. The following must be avoided:
- Remarks that could be interpreted as allegations of misconduct
- Unsupported assertions or statements
- Inflammatory or insulting language
Thank You!
Thank you for taking the time to flag this posting; we review flagged postings on a regular basis.
closeThere is competition here!
Posted by bostman on 01 May 2013 at 19:34 GMT
It is not passive drift - even if all organisms have the same fitness - when you limit the number of organisms to five per niche/phenotype/genotype. The increased values over time of the trait that you call evolvability is a direct effect of competition for space. The carrying capacity of each niche of five creates a selection pressure for offspring with greater mutational effects, and thus competition for being different. This is not passive drift, and it is not without competition.
Additionally, evolvability is a parameter that has everything to do with the genotype-phenotype map, which here is just one-to-one. In other words, I do not think it is sufficient to name the second 'evolvability' and say that that is what you are studying. You might as well have called it "parameter that controls the effect-size of mutations". Modeling evolvability as a mutable parameter ignores the genotype-phenotype map, which does not conform to how people think about the cause of evolvability.
RE: There is competition here!
kenstanley001 replied to bostman on 02 May 2013 at 06:33 GMT
Hello from the authors. The commenter may wish to revisit the article to review that there are two different sets of experiments and multiple genetic models. The first set of experiments is indeed entirely passive with no limitation on the number of organisms per niche. In these cases (which are called “Passive Drift Models” in the article), evolvability is observed to go up on a per-niche basis, that is, the average niche becomes more evolvable over generations, even in this entirely passive case. So that case is indeed passive drift.
Regarding the concern with genotype to phenotype maps, as noted above, there are multiple models. Only the “abstract” model instantiates evolvability as a parameter. However, the “evolved robots” models actually involve a complex mapping between genotype and behavior, and its results nevertheless echo the abstract model’s results, which is one reason we provided both such models: the abstract model makes a theoretical point, while the evolved robots model is more concrete.
It may be useful to elaborate a bit on our response to the type of criticism in this comment because it is likely a point of concern for others as well. Our interpretation of the results in these studies is not that they are a precise model of nature that tell us in some sense “what is really going on.” Rather, this article shows that we should be careful about assumptions regarding how evolvability can come about theoretically. In other words, the article is a cautionary note, not a proof of how something works in nature one way or another. It shows both that even if evolvability appears to go up on a per-species basis in nature, it may not actually be going up overall, so we need to be cautious about assuming that it even does go up overall in specific branches or lineages. Furthermore, it shows that one very simple condition (limited capacity) can accelerate evolvability even further, which means we need to be cautious about assuming anything more is necessary to yield increasing evolvability. You can of course argue that nature is different from the models in the article, that there is something special in the genotype/phenotype map or in some other aspect of nature, but the point is that the onus is back on the one arguing such a point. That is, these results highlight that we can no longer simply assume such a factor is essential to increasing evolvability.
RE: RE: There is competition here!
bostman replied to kenstanley001 on 02 May 2013 at 15:42 GMT
Regarding the result from the passive drift model, I gather you are referring to figure 5. The average evolvability averaged over niches does indeed here show an increase (even though you do not supply error-bars, so I can't be sure if this is significant). However, do you allow evolvability to go below zero? I suspect you do not (that wouldn't really make sense), and if not, then this result can be explained purely by diffusion.
I do understand that the abstract model is to supplement the robot model, but I just don't think it shows what you claim.
RE: RE: RE: There is competition here!
jal278 replied to bostman on 03 May 2013 at 04:30 GMT
This is the authors again. Note that figures 1-4 illustrate the results of the passive drift model more clearly (e.g. the label to figure 1 reads "Evolvability in the abstract passive drift model") than figure 5 (which focuses mainly on the niched model). Also note that the differences between shown averaging methods (in figure 1) and between the models and initial evolvability level (in figure 5) are indeed significant after 500 generations. The reason error bars are omitted is that the variance in the abstract passive drift results is small enough that the bars would be uninformative and only complicate the plot.
While it is true that the abstract passive drift model enforces the constraint that evolvability remains non-negative (because a negative probability is not meaningful), the difference between averaging over niches and averaging over the population shown in figure 1 illustrates that the apparent evolvability increase over niches in that model is not due to the triviality of such a fixed lower bound. Such a slow-increasing trend might indeed become apparent if the simulation continued indefinitely; however the interesting effect the model demonstrates is that the space of niches acts like a filter to separate more evolvable from less evolvable genomes significantly more quickly. Supporting this idea, figures 2 and 3 indeed show that the niches farther from the starting niche contain more evolvable organisms on average. Figure 4 serves to show that the same trend is visible in a more concrete domain where evolvability is an emergent property of the evolved genotypic architecture rather than a purely abstract property.
Hopefully this response clarifies how the results from the drifting models do in fact support the claim that an apparent increase in evolvability across niches can result from a pure drifting process across genotypes.
RE: RE: RE: RE: There is competition here!
bostman replied to jal278 on 03 May 2013 at 18:47 GMT
Okay, looking at the figures again (1 and 5) I realize my mistake. There is indeed no effect (even from diffusion) on evolvability in the passive drift model averaged over the whole population. But there is when you average over niches.
However, this is of course just an artifact of your particular averaging scheme. Those individuals with low E will sit close to the origin, and per "niche" there will be many of them. Those who have high E will spread out more, and per "niche" there will be few. But you are averaging per "niche", so you are biasing the measure by giving more weight to those with high E. Thus your result.
I do not think that a phenotype is a niche. It is certainly not a genotype, and here they are all the same thing.
RE: RE: RE: RE: RE: There is competition here!
jal278 replied to bostman on 04 May 2013 at 07:40 GMT
This is the authors. As we mention in the paper, in the abstract passive drift model "evolvability across the population remains unchanged on average; it is only evolvability's distribution over the space of niches that becomes biased during evolution," so the commenter's remark that the difference in evolvability in the passive drift models is due to an artifact of the averaging scheme is correct, but is exactly the point that model is designed to illustrate. In particular, the abstract passive drift model shows that the deceptive appearance of overall increasing evolvability (when averaging over niches) can result from a purely drifting process over genotypes (where evolvability on average remains constant). In other words, the results suggest caution when seeking additional explanation for evolvability beyond such a drifting process because our intuition about increasing evolvability in nature may result from just this kind of biased average. That is, our impression about evolvability may come from surveying the wide diversity of nature rather than averaging evolvability over all organisms.
However, beyond the abstract passive drift model, the commenter should also review the other models in the paper (as noted in our first response). The concrete evolutionary robotics experiments address the concern about overly simple mappings between genotypes, phenotypes, and niches in the abstract models, and the limited-niche models demonstrate true genotypic bias towards evolvability (and not only the appearance of increasing evolvability when averaged over niches).
Focusing on the genotype/phenotype/niche mapping issue, supplementing the abstract models (where genotype, phenotype, and niche are closely related) are the concrete evolutionary robotics models (where there is a complex relationship between the genotype, the artificial neural network it encodes, the behavior the neural network induces in the robot when the neural network controls it, and niches that consist of distinct subspaces of possible behaviors).
As noted above, the overall idea is that the abstract models are an idealization that clearly illustrate the mechanism by which evolvability may increase (either passively through drift or with limited-capacity niches), and that the more concrete models are intended to show that these effects can hold in more complex situations.
We hope that highlighting the different models and their motivations addresses the commenter's additional concerns.