Using Weisberg’s Framework to Evaluate Optimality Models

In Discussion on February 28, 2010 by Jenny

In Potochnik’s article “Optimality Modeling and Explanatory Generality,” Potochnik argues that even if we had perfect information regarding the genetic information of species, we would still have reason to use an optimality approach to model natural selection, rather than use models that include such genetic information.  Her reason for claiming that an optimality approach is useful regardless of our access to genetic information is that an optimality approach tells us about the “fitness-conferring interactions between organisms and their environment” in a way that more complex models could not.  And in certain contexts, explanations that capture such fitness-conferring interactions are better for our scientific explanations than ones that do not.

While I won’t go through Potochnik’s argument for why this is so, I think it is interesting to note that her argument seems to be easily framable in terms of Weisberg’s categories of idealization.  Here’s how it would go in Weisbergianese:

Lewontin (and others) think that the optimality approach to modeling natural selection is a Galilean idealization, in that it introduces simplifying distortions to gain computational traction that will later be de-idealized as we gain more information about the target phenomenon.  For Lewontin, as we gain more genetic information about the target phenomenon, we will be able to correct many of the distortions we make in optimality models, until we eventually dispel with such models altogether.

Potochnik counters by saying that the optimality approach shouldn’t be understood as a Galilean idealization, but as a minimalist idealization, where the goal is to include in our model only the primary causal factors that give rise to our explanandum.  Thus, even as we progress in our understanding of genetics, our model won’t change or be obviated.

If this re-casting of Potochnik’s argument is correct, then we would expect the representational ideals of the two forms of idealization to be different.  And indeed, we see that this is so: Lewontin’s preferred form of explanation (the Galilean idealization) aims for the representational ideal of completeness, while Potochnik’s preferred form (the Minimalist one) aims at describing the primary causal factors, given certain fidelity rules (which say how precise our explanation has to be).

If Weisberg’s idealizations can be applied, we would also expect for the two forms of explanation to be non-competing, in that they account for different phenomena. Again, this is also true.  As Potochnik concludes, both models of explanation have different explananda: the optimality approach explains “long-term phenotypic evolution by natural selection with a particular interest in the fitness effects of organism-environment interactions” while models that take into account lots of genetic information do not.

So what I want to know is whether anything is lost in translating Potochnik’s framework into Weisberg’s idealizations?  Can everything she explained about the optimality approach be explained in terms of Weisberg’s framework?  I would suspect that it cannot, but I don’t really have a good reason why this might be so.  My hope is that in demonstrating where Potochnik’s framework comes apart from Weisberg’s, we can gain some traction on which sort of framework better captures the debates between optimality modelers and their opponents.


7 Responses to “Using Weisberg’s Framework to Evaluate Optimality Models”

  1. I’m *sure* Collin will have a lot to say about this! The latest Philosophy of Science has a paper by Potochnik that goes over some of the ground we did on reductionism. Check it out (you will need to be on campus or use a VPN client to link):

  2. Jenny, I like your suggestion. Especially since I thought Potochnik’s arguments against the principle of maximal inclusion (or ideal of completeness) were not very convincing. She says that this principle is “wrong”, but her reasons seem to be 1) because it would yield an explanation that is too complex for our small brains and 2) normally complex phenomena are not explained in this way. As Potochnik notes, “the strategy of maximal inclusion is not effective for such complex causal processes, for it would result in exceedingly involved explanations. Even when constructing such an elaborate explanation is even possible, we seldom or ever explain this way” (682). Just because it is true that the principle is cognitively taxing or that many explanations don’t use this principle doesn’t mean that the principle is wrong. The principle may very well be of interest to a modeler depending on what the target of explanation is. So, it seems that pulling MW categories of idealization out would hurt her assessment of the principle of maximal inclusion.
    However, as to your particular application of the work of MW, I am not sure if Lewontin’s preferred explanation counts as Galilean. This is because it seems like Potochnik’s analysis of what the best explanation is can capture Lewontin’s preferred explanation. Lewontin’s context of inquiry might just be different from that of a theorist who is employing optimality models. Perhaps Lewontin just wants to know why, from the perspective of gene transmission, will we get population X instead of population Y or Z. So, it might be that Lewontin is not using a different ideal from the optimality modeler; rather, he just has a different context of inquiry. I take this to be a real virtue of her view. Arguments that claim that specificity is sometimes preferable over generality do not affect her characterization of best explanation. The reason for this is because the context of inquiry will determine the degree of generality. Hence, her characterization of what constitutes a best explanation captures both so-called precise explanations (à la Lewontin) and general explanations (à la optimality modelers). What do you think?

  3. I am of the opinion that models are not explanations and they are not theories, and the only way to save Potochnik’s core position is to morph her work into something along the line of Weisgerg’s. At its core, her argument is one of Weisberg’s.

    Optimality models have utility as idealized models of the minimalist variety because they help us learn about relevant causal forces. The models are not explanations in and of themselves. For Potochnik optimality models have utility because they help in this precise kind of understanding and illuminate what she feels are explanatory forces. The model itself explains nothing.

    She sets out criteria for selecting a model, not an explanation. The only ideal explanation that can be considered best is one that completely explains everything for all time, preferably in a single character.

  4. I think Jenny is in general correct about the kinds of idealization that Potochnik is discussing in her article. It’s one of the reasons I really liked that paper by Weisberg. However, I think that both approaches can be captured by the 1-casual representational ideal simply by allowing the context of inquiry to change. This is similar to Yasha’s point above. Some contexts will favor the causes represented by an optimality model, others will favor the inclusion of different sets of causes. Lewontin is certainly interested in a very specific context of inquiry and I think that Potochnik has simply pointed out that there are other contexts that might favor optimality models. Thus, I see the debate as not so much over completeness (Galilean) vs. 1-causal (Minimalist), but rather over the contexts in which evolutionary biologists find themselves now and in the future.

    Although the mapping onto Weisberg’s categories is helpful in some respects, it is misleading in others. I have two major worries about doing so. First, it suggests that a dynamical model is simply an optimality model with more causes added in. This is false; they are two very different types of models. Optimality models describe an equilibrium point (the local optima) that is claimed to be independent of the censored information. Dynamical models describe the step-by-step trajectories that lead to evolutionary outcomes (i.e. they are not equilibrium explanations). Thus, it is misleading to suggest that the explanations offered by optimality models are just dynamical explanations (what Lewontin prefers) with some of the causes left out. My second worry is that the real controversy surrounding optimality models is not whether their explanations would benefit from the maximal inclusion of causal factors, but rather (1) are they adequate explanations in the first place, and if so, (2) are their assumptions satisfied frequently enough to justify the way adaptationists rely on them (e.g. assuming a priori that there will be an optimality explanation for a trait)?

    In addition, unfortunately, I think Potochnik has grossly misrepresented Lewontin’s position so as to have a suitable target for her sort of view. Lewontin is not committed to anything like the principle of maximal inclusion that Potochnik associates him with. In fact I can’t think of any evolutionary biologist who defends such a view of maximal inclusion as objectively preferable when studying such complex systems. No one is advocating such a principle concerning what makes for better explanations, some biologist simply object that optimality models are unable to capture enough of the causes to do the job (or at least not enough to assume they will do the job). Including more isn’t always better, but when explaining (most) phenotypic traits – they will argue – more information is usually required. Therefore, although Weisberg’s categories are useful for distinguishing between the representational ideals Potochnik discusses, those ideals are not really what is at issue in the debate over optimality models.

    What Lewontin actually argues is that that we will have to investigate genetic dynamics first in order to even test whether the explanations of optimality models are correct. The reason this test is needed, Lewontin argues, is because the frequency with which the assumptions that underlie optimality models are realized is not known (although he is certainly skeptical of their ubiquity). Nowhere does he claim that including more causal information will make for better explanations. Rather, he suggests that in order to determine if optimality models are the correct explanations of traits, we will have to investigate the genetic dynamics anyway.

    Optimality models assume (among other things): (1) that other evolutionary forces were not important to a trait’s evolution (something Lewontin doubts is a safe assumption to make without first investigating those other forces); (2) that we are able to separate out specific “problems” that organisms face as well as discern the “best” “solutions” to those problems (something Lewontin also doubts is actually possible); and (3) that the right kind of genetic variation was available (which Lewontin provides several examples against).

    What Lewontin actually says in his 1979 paper is, “We may sum up the difficulties of optimality arguments by saying that they are useful only when the solutions they propose correspond to the outcome of a dynamical process of genetic evolution, but that each case must be separately verified. If, however, it is necessary to solve the dynamical problem in order to validate the optimality argument, then the optimality technique has added nothing.” In other words, I read Lewontin not as making the claim that including extra causal information will lead to better explanations (as Weisberg’s Galilean form of idealization and representational ideal of completeness would suggest). Rather, he suggests that optimality models ought to be phased out within evolutionary biology because in order to verify them we will have to investigate the outcome of the step-by-step dynamics. Lewontin’s point concerning how biologists must proceed is independent of any claims about what makes for a better explanation in the end. He may very well agree that, as Sober argues, whether we want a more general explanation or a more detailed one will depend on the context of inquiry. The two aren’t competing; science has room for both. Lewontin objects to the adaptationists’ use of optimality models (without checking the dynamics), but he does not say that those models cannot be explanatorily useful once we know they are true.

    That said, I think there are a number of very interesting questions that come up in Potochnik’s article that are independent of her misrepresentation of Lewontin and the debate over optimality models. I will save my own thoughts on them until Tuesday, but here are some topics that will perhaps stimulate some discussion:

    (1) Potochnik builds in explanatory generality as an objective feature of the best explanation of an event E. The context determines how general the best explanation is able to be, but across contexts we are supposed to prefer the most general one allowed. Is this correct?
    (2) What is the relationship between predictive accuracy and explanatory adequacy when it comes to optimality models?
    (3) What kinds of explanations are actually offered by optimality models (e.g., causal, unifying, equilibrium)? Can we single out one kind of explanation for them? And how does this perhaps inform us about the contexts in which they will be the best explanations available?
    (4) A functional explanation is “forward-looking.” It emphasizes the adaptive significance of a trait observed in present individuals in a given environment. An evolutionary explanation is historical; e.g., the trait in question is an adaptation. It attempts to explain the trait in terms of its past evolution. To explain the maintenance of a trait, one gives a functional explanation; to explain its origin requires an evolutionary explanation. Are optimality models even able to provide evolutionary explanations for the existence of phenotypic traits (this comes up in the Brandon & Rausher article)?
    (5) Let us grant that in some contexts of inquiry optimality models are the best explanation. A further important question is how often is that context recognized within evolutionary biology (e.g., in studying human behavior)? In other words, there is another debate over where and how often the contexts of inquiry obtain now and in the future, not their mere existence.

  5. I find this entire adaptationism, optimality, and modeling business a bit troubling. To start with I am troubled by simply calling the change in rate of appearance of a trait the evolution of the trait’s fitness. Reading this over and over, I found myself beginning to agree with Andre that this statistical model has nothing to do with Darwin who wrote of the evolution of species. If we are just tinkering with traits, that is within a single species, and not evolution, in my view, but that is another argument.

    Setting aside metaphysical questions, as I think science does, when confronted with modeling as the beginning of creating a hypothesis, there are certain trades that will be made. On the one hand we have the totality of the phenomena. Clearly the best explanation would explain everything, but that is impossible with our limited understanding and limited brains.

    As this understanding is limited, we look for generalities. Generalities allow us to use more basic understanding over and over in different situations. This is where we begin to idealize, either to gain more understanding that may lead to hypotheses regarding central causal mechanisms or to break the problem of the phenomena into manageable parts.

    Optimality models are a peculiar thing indeed and appear to be serving as a potential method for arguing for a rather peculiar hypothesis that seems a bit absurd on its face. The adaptationism hypothesis, even in the modified version of Brandon & Rausher, is oddly vague and hardly likely in every case. The ‘evolution’ of the ‘fitness’ of a particular phenotypical trait is known to have been caused by a mix of natural selection and a host of other forces, but the adaptationists argue for the primacy (sufficiency) of natural selection and that this evolution had reached ‘optimality’.

    To test this optimality, it is proposed that the phenomena is compared to an optimality model that is idealized to exclude other factors in statistical evolution other than natural selection. Brandon & Rausher do well in showing that there are serious problems with this as the model could quite likely be wrong, and quite possibly the trait has not reached optimality yet. As the environment and other factors are quite dynamic, one would think this would happen more than a handful of times, especially given the increasingly disruptive impact of human activity.

    Brandon & Rausher point out that an idealized optimality model does not even explain the fitness of the trait, as the other factors given in the actual fitness will have an impact. They suggest that one should perhaps verify optimality by being sufficiently close to what has actually happened. This is because only a detailed historical account explains what actually happened in the ‘evolution’ of the fitness level of the particular trait. Every causal influence was necessary to completely explain how and why the fitness level turned out the way it did. The difficulty here is that adaptationists are attempting to make a dubious argument regarding sufficiency and optimality.

    This is not to say models based solely on natural selection to not have informative utility. This is where we use Weisberg’s one-causal approach. While the details of what actually happened in any particular case are necessary to explain the exact outcome, generalizing will (potentially) demonstrate a common factor involved, in this case natural selection, that left to an idealized model of its own might get pretty close to the final fitness ratio. This might have utility for either rough predictive purposes, or investigative purposes as to the causes as to why something might deviate even further from the model.

    I was a bit harsh before in saying that models explain nothing, but idealized models do not explain what went on- they illuminate common causal factors. Ultimately to explain why there is the exact fitness at time t we need to have complete knowledge of the universe, but to understand the common mechanisms between generalizable situations, we need to idealize. Use of optimality models to test adaptationism just seems silly, as the hypothesis as formulated may be practically, if not logically, untestable. Sorry for the so unsytematically approach Collin’s questions, but I’ve been strangely annoyed my thoughts on this all day, and typed them out without direct reference to his very good questions.

  6. Todd: your last paragraph reveals the problem, I think, in your thinking. You are expecting optimality models to explain “what went on…illuminate common causal factors”. But, as Potochnik points out (and the point is echoed in this thread) there are other ways that a model might explain.

    The rest of the paragraph also reveals problems with your thinking. First, why is the problem of explaining evolution any different than explaining other phenomena? On the view you espouse, to explain, say, why this ball fell–to provide all the causal details–would I have to cite all the factors that lead back to the beginning of the universe? No, of course not. The problem is you aren’t thinking about alternative ways to explain some phenomenon. Second, logical testability–what’s that? We need empirical testability. This issue was raised by Orzack and Sober. They show how optimality tests provide empirical testability about adaptationism in indirect ways. To make your “silly” label stick you have to address their arguments.

    Remember: philosophy is about evaluating arguments (determining whether an argument is valid or not, or, alternatively strongly inferred or not). It isn’t about asserting one or the other conclusion.

  7. Second point first, “most” or “a lot of the time” is likely not testable is what I was getting at, but upon a closer examination I see that it might be, being very charitable. It is empirically impossible to test everything that adaptaitionism would cover. But there is potentially a large enough sample size where one could create a confidence interval that might not be huge, and, being generous, we might define ‘most’ as more than half the time. So, with a large enough sample size, one could potentially create a confidence interval that had a sufficiently high enough probability that one could be safe in predicting that adaptationism was likely true or false more than half the time. This, of course, ignores the difficulty that arises if the optimality model is inadequate, as pointed out by B&R. Thus I perhaps overstated that suggesting it was logically impossible to test such a hypothesis, but what I mean was that the ‘most’ descriptor was so loose as to (potentially) make it logically impossible to test adaptationism- that is- it is so vague as to defy verification. But upon closer examination, in this case that was probably unfair, as it is possible (and exceedingly difficult) to demonstrate whether the statement is true or false.

    As to adaptationism being ‘silly’, well, that is a longer argument and I was trying to keep my post focused. If I have time after lunch I can make such an argument for the blog, but a lot of it has to do with the ‘most’ thing. That is just weak. I think B&R do a good job taking it apart, though that is not their intent.

    As to the first point, I think my position is similar to Weisberg, but not necessarily Potochnik. I think models can serve as a utility in exploring and discovering causal factors, but B&R point out quite well that optimality models do not exactly reproduce the state of the trait, B, but, when done correctly, a state sufficiently close to B, namely B’. It is this closeness that might allow one to contemplate that natural selection as the ‘sufficient causal factor’ or ‘primary’ factor in the perceived ‘optimality’.

    The difference here, between the ball falling, and having fitness B, is distinct. Fitness B is a precise real number value. A number that differs in the slightest from fitness B is not fitness B, but a sufficiently close real number B’ could be of interest. If you want to explain B’, fine, your model is sufficient, but it does not explain B. If you want to explain the prominence of a trait, rather than its actual fitness, your model might be sufficient, but again, it does not explain the fitness of B.

    The difference is the same in explaining why the ball fell and why the ball had the exact instantaneous velocity it did the moment before impact with the ground. If you want why the ball fell, a very broad model might get you there. If you want why the instantaneous velocity was what it was, you must describe the universe. Simply applying Netwon’s universal law of gravitation will produce a rough approximation, but not explain the actual instantaneous velocity.

    The obvious problem here is that if we take Newton’s approach as a ‘model’ we have not actually explained why the ball fell in terms of the current relativity paradigm. The model will need additional verification. Models that are general can point to potential common causal factors, but in and of themselves they do not necessarily explain.

    Does an optimality model show natural selection is a significant and common factor in the fitness of a trait, maybe, but we need to verify (but, again, getting back to the silly, weren’t we already pretty sure natural selection played an important part in fitness already, so what is the model doing for us? Showing its relative ‘importance’? B&R do a good job of showing it is not verifying the adaptationist hypothesis.). Does it explain the actual fitness of a trait, not unless the model accurately produces the fitness B.

Leave a Reply

Fill in your details below or click an icon to log in: Logo

You are commenting using your account. Log Out / Change )

Twitter picture

You are commenting using your Twitter account. Log Out / Change )

Facebook photo

You are commenting using your Facebook account. Log Out / Change )

Google+ photo

You are commenting using your Google+ account. Log Out / Change )

Connecting to %s

%d bloggers like this: