Richard A. Watson: Compositional Evolution


(Andy Walsh) #21

Chapter 3 is a bit daunting to comment on given its length and substance. The following are some initial reactions rather than a comprehensive analysis.

The discussion of compositional mechanisms is worthwhile. I can see the logic of grouping them together as qualitatively similar and qualitatively distinct from substitution; I wonder if all biologists would agree.

Several of them could be considered mutations in a general sense, which can muddy some discussions of evolutionary biology. If one tends to think of mutation as synonymous with substitution, then one will likely think of evolution largely in terms of gradualism (as Watson defines it) with all the limitations and caveats that come with it.

I found the discussion of recombination and the various models of how it might beneficial to be helpful. I’m not sure I’ve seen that particular topic covered so systematically. I also appreciated the discussion of crossover in genetic algorithms and why it can sometimes seem like just another form of mutation. As someone who has played around a little bit with basic GA programming, I’ve puzzled over that myself.

Somewhere around “Beyond Sex” I started to get bogged down more frequently. Possibly because the connections to biology came around less often, possibly because the chapter focuses a lot on surveying more than synthesis. I suspect I will need to revisit some of this material if the later chapters build on it.

(S. Joshua Swamidass) #22

Recombinations make a very big difference.

(Andy Walsh) #23

If we’re talking biology: yes.

If we’re talking computational algorithms: yes, but Watson discusses how it is possible to implement crossover in genetic algorithms such that it is functionally equivalent to substitution with high frequency. He’s not underselling recombination in computational applications, just showing how to use it effectively.

(Andy Walsh) #24

Returning to the book after an unintended hiatus…

Chapter 4 I’m going to guess will be the key chapter, which is interesting because it is not directly about evolutionary processes but instead about objective functions (or fitness landscapes, if one prefers – not that they are exactly equivalent concepts). While I’m not the most qualified to make such an assessment, it seems to me that the most interesting contribution of this work is identifying the conditions that differentiate the success of mutation/hill-climbing methods and compositional methods.

As such, the chapter has something of a chicken-and-egg problem. On the one hand, it has to describe the objective functions of interest to setup subsequent chapters showing the results of simulation studies that demonstrate the performance differences of different evolutionary algorithms when optimizing these functions. On the other hand, it has to motivate the choice of the particular functions without a solid theoretical basis for doing so. Or at least it seems to me that the choice of functions is at least partially empirical; they were found to satisfy the need for performance differentiation and then the relevant properties were worked out. That’s a valid approach, but it means Watson has to give away some of the later results up front.

It will be interesting to see Watson goes into any more detail on how his concepts of systems that are separable and decomposable apply biologically.

(Andy Walsh) #25

Chapter 5 shows how mutation- (substitution-) only methods do not perform well on the objective functions described in Chapter 4. That result was already pretty clear from Chapter 4, so I"m not sure there’s much more to say about it.

Possibly of more interest to this group is that Watson revisits Behe and irreducible complexity. Watson thinks that his HIFF objective function exhibits irreducible complexity, and so the challenge that mutation-only/hill-climbing methods have with such functions suggests they will have difficulty producing irreducibly complex results. In other words, irreducible complexity actually does represent a challenge to strict Darwinian gradualism.

There may be a point there, but I’m not sure how interesting it is. For one, I think Watson may be interpreting irreducible complexity a little too broadly to apply it to the HIFF functions. But more significantly, the challenge of irreducible complexity was brought up in a context where evolutionary biology had already gone beyond models of variation and inheritance that were limited to substitutions as the only sort of change (if it was ever so limited in its scope). Unless Watson thinks a more formal, theoretical response to irreducible complexity is needed, I’m not really sure why it’s worth bringing up.

(S. Joshua Swamidass) #26

Why do we care about strictly Darwinian gradualism? We already know this is false, right?

(Andy Walsh) #27

I don’t know, and yes.

In case it wasn’t clear, the sentence you quoted was my attempt to summarize the point I think Watson is trying to make. But like you, I am a bit mystified as to why he thinks it needs to be made. Perhaps he is trying to, in your terminology, differentiate IC2 from IC3 and IC4. Perhaps he is trying to stake out space for an extended synthesis. I’m not sure.

If it is another instance of EES/Third Way/etc. advocacy, what I find particularly odd about this one is again just how narrow his version of gradualism is. Recombination is one of his compositional methods, and so not included with his mutation-only gradualism. That’s fine from a theoretical point of view in terms of classifying processes and establishing which ones work better in which contexts. But as a critique of evolutionary biology, it strikes me as a critique of a form of biological evolutionary theory that never actually existed.

Now, maybe all of this will be clarified or better contextualized in a concluding chapter. And to be fair, it is a minor element of a book which I think is largely interesting from a computational point of view.