Skepticism and Rational Belief
by DeuceKrauze's recent post on Skepticism And The Origin Of Life gave me a good excuse to write on a topic I'd been meaning to for a while. Namely, what is the difference between a belief and a rational belief? Put another way, what does it take for a particular belief to be held on the basis of reason, rather than assumption or desire? The obvious answer is that in order for a belief to be held on the basis of reason, you need reason(s) to believe it. And by that I mean not reasons for wanting it to be true, but reasons for thinking that it is true. That is, you need evidence for it.
Now, there's an interesting idiom of our speech (or maybe it's universal to other languages too, I don't know). We talk about evidence being "for" or "against" things, as if evidence itself had motives or preferences, but this is technically not right. Rather, evidence is data that the individual has analyzed, and found more compatible with a particular conclusion than with the logically possible alternatives to that conclusion. Hence, to say "X is evidence for Y" is actually shorthand for "X, taken alone, makes me more for Y relative to Y's alternatives after rational consideration". The bottom line to all this is that a rational belief is one that the individual holding it has compared with the logically possible alternatives, and found to be the most logically compelling among them, based on the data. An uncontrasted belief is one that is not held on the basis of reason. A rational belief is one that is held skeptically.
In Krauze's post, one responder said "I guess to me to be skeptical that life on this planet derived from the chemistry that existed back then is a bit too close to being skeptical of science in general." I don't mean to single out the poster in question (he is, after all, in good company), but note that this is a reason for wanting it to be true, not a reason for thinking that it is. Mike Gene responded with "Which would then mean we have defined science such that geochemistry must have given rise to cellular biology."
One can define science that way if they wish, but it is important that we accept the logical implications of doing so: namely, science (so defined) can't be considered a way of acquiring true, rational beliefs on its own, since conclusions are never compared to their logical alternatives. If one wishes only to have the most pragmatic view allowed within a particular predefined paradigm, they can settle for this. But if one wishes their beliefs to have a foundation in reason, a bit of self-skepticism is always on the menu.
December 14th, 2005 at 10:59 pm
Deuce,
Rational belief must be an inference from the data.
Bayes' Theorem tells you how to do that. To be a valid inference, your theory has to meet two requirements.
First, it must preferentially predict the observations, it cannot simply be consistent with them. If the theory fails to do this, then there's no inference from the data. (P(B|A) = P(B)).
Second, your theory cannot simply be a paraphrasing of the data. How do you know if your theory paraphrases the data? If your theory has more free parameters in it than you have observations, then you have just paraphrased your observations. If your theory can always be tuned to fit any data set, then it's a tautology, not an inductive inference. (P(B|A) = P(A|B) = 1.)
Because a valid inference has fewer parameters in it than the original data set, you can use the inference to deduce one subset of observations from the the rest of the observations. This means your theory must be predictive.
Inferred theories usually raise new questions, but they must also answer some. If your theory doesn't make predictions within the data, then you don't have an inference from the data (nor, for that matter, an explanation of the data).
A corollary is that belief in the truth of a proposition which is consistent with anything you might ever experience is irrational. Indeed, it is questionable whether such a proposition is even meaningful, i.e., can ever be assigned a truth value.
Comment by doctor(logic) — December 14, 2005 @ 10:59 pm
December 15th, 2005 at 12:25 am
question for doc log
how many predictions does a theory have to make before it be considered a theory?
Comment by willo — December 15, 2005 @ 12:25 am
December 15th, 2005 at 12:28 am
You know… that's actually brilliant.
I've been thinking about how to dialog lately, and it's difficult because I criticize you when you use circular reasoning, but I'm blind to it when I do… A little judicious application of self-skepticism in every thought would go a long way toward having actual communication. When we fail to do that, we may talk, but we're not communicating…
Comment by havoc — December 15, 2005 @ 12:28 am
December 15th, 2005 at 10:23 am
Blah, blah,blah, Bayes' Theorem, blah, blah, blah.
You sound like one of those analytic philosophers whose been described as a blind man in a dark room searcing for a black hat [Bayes' Theorem in the brain] that isn't really there.
Mistaking models for human rationality. Should one even proceed in discussion with a person who moves from the nature of rational belief to the ridiculous pronouncment that "rational belief must be an inference from the data." - If you want to find the flaw in dr.logics arguments, look for what follows the words "must" and "necessarily."
So, wtf, before Bayes' Theorem people didn't know how to properly form rational beliefs?
Here's another model to consider: a belief is rational (or a theory is to be preferred) in cases where the product of disbursement and probability of predictive success (use your Bayes' Theorem here) is high.
This model for rationality emphasises that it is perfectly rational for human beings to pick a theory or form a belief by weighing the significance of truth against one's life. So, if I have two theories (maps) about fifty islands, and both theories have 75% predictive success, but the first theory tells me which islands have cannibals and the second theory tells me which islands have palm trees, and you can only pick one map, which one am I going to pick?
Rationality has to take value into account. It is not just an inference from the data.
Otherwise you are describing some imaginary process that some imaginary being performs. Human beings do not *just* do Bayes Theorem when they form rational beliefs.
Comment by bipod — December 15, 2005 @ 10:23 am
December 15th, 2005 at 10:39 am
I should note that I'm not endorsing the second model mentioned in the post above.
My main points here is that it is rational to be selective (on some value) about theory selection.
My secondary point is that dr.logic is way too enamored with his formalisms to see the actual world. Thus, he says something as innane as this "Rational belief must be an inference from the data." and proceeds to tell us how to do this by way of Bayes' Theorem. Very good doctor. Very good. I'll make sure to check that there aren't any bugs in the code up in the grey matter.
Comment by bipod — December 15, 2005 @ 10:39 am
December 15th, 2005 at 2:18 pm
Hi bipod,
LOL! This is reminiscent of a quote from one Gordon Sumner: "Da doo doo doo, da daa daa daa, is all I have to say to you."
Flattery will get you nowhere!
I don't think you understand my comment. I'm not trying to create a model of how humans think. I'm not saying that human brains are pure Bayesian engines. They can't possibly be. Humans aren't fully rational.
What I'm trying to determine is what constitutes a rational inference from observation. I don't want to define rational to mean "the way the average human thinks."
Agreed.
What you are talking about is rational method for making choices, not inference.
For example, I can choose to assume one theory is true rather than an alternative based only on the possible outcomes predicted by each theory, without any validation of my theory from observation.
Suppose I reach my front door at night and find that I have dropped my keys. Where do I search? I could have dropped them anywhere along my path, and I have no observations to tell me where. However, I will first assume that I dropped the keys under a street lamp, because otherwise I will not be able to find them until morning.
There's no reason to adopt the "dropped under the lamp" theory over the "dropped in the dark" theory. However, it is the predictions from these theories that make the adoption of the theories rational.
BTW, your maps example and your expectation value concept both rely on prediction. They are both fine examples of rational choice.
Can you think of any rational choices that are not facilitated by a predictive theory of experience?
If you base a strategy on a theory which fails to predict experience, then your strategy will be blind to outcomes.
Comment by doctor(logic) — December 15, 2005 @ 2:18 pm
December 15th, 2005 at 3:15 pm
drlogic wrote:
I don't think there's anything particular about the bayesian approach which makes this argument more or less valid. In fact, generally, a Bayesian analysis allows the scientist to import additional non-data information into an analysis. When such information is available, its use can also dramatically improve the quality of inference.
It's true that it is trivial to find a model that fits data perfectly, and such a model is likely to have zero information when it comes to predicting the future behaviour of a process. The objective is to capture as much structure in data as possible (because future observations are also likely to conform to this structure), but using as few parameters as possible.
However, at the heart of theory/ practice are criteria that enable rational decisions on whether the gain obtained in making a model more complex has value despite having less predictive power. Additionally, in the particular case of extreme value modeling, it is usually the case that models with many parameters do just as well in predictive terms as models with much fewer parameters. Quite what criteria are adopted and how the value is measured vary from application to application, and are indeed different in Bayesian and non-Bayesian approaches, however the issue is foundational.
Comment by Guts — December 15, 2005 @ 3:15 pm
December 15th, 2005 at 6:26 pm
Hi Guts,
I agree with this statement. You're choosing between two theories that both have at least some predictive power.
The question is, when is a model just a tautological paraphrasing of the data? (When this happens P(T|O) = 1.)
My claim is that a theory is paraphrasing observations when it makes no predictions. None of what I am saying prohibits things like fitting a polynomial function to points on a graph (this is predictive). However, my claim does prohibit just drawing dots over the data points and calling them an inductive inference.
Comment by doctor(logic) — December 15, 2005 @ 6:26 pm
December 15th, 2005 at 10:06 pm
Doc Logic,
So is holding a belief that Bayes theorom tells you whether you hold a rational belief a rational belief? I.e. Does Bayes theorom meet it's own criteria?
Comment by alangrey — December 15, 2005 @ 10:06 pm
December 16th, 2005 at 1:07 am
alangrey:
Actually Bayes Theorem does not really imply the position that Doclogic holds. Bayes Theorem seems to imply that there is no difference between prediction and model adjustment, because the conditional probabilities used in Bayes Theorem are neutral with respect to time. For example, Einstein's theory of relativity got support from the fact that it correctly "predicted" the precession of the perihelion of Mercury, even though that information was known way before Einstein constructed his theory. Mike Gene "predicted" proofreading during transcription even though it was already known (although, not commonly). There is a lot of debate about this going on. For more about that see here .
A brief excerpt:
Doclogic also says that that if a theory is only consistent with data that P(B/A) = B and that it must preferentially predict the observations. However, there will usually be some hypotheses with non-zero prior probability that are inconsistent with the data. Bayes' Theorem will require that their probabilities be reduced to zero after the observations. For example, you have the sum of the prior probabilities of all hypotheses inconsistent with the data. That sum will, in general, be distributed over all the hypotheses consistent with the data (though not equally), so all hypotheses consistent with the data will generally have a higher posterior probability than prior probability. That would be, for any such hypothesis B and data A, P(B/A)>P(B).
Doclogic writes:
As to whether a theory must be predictive in order to be valid, actually, this can get pretty complicated. However, say that when P(B/A) = P(A/B) = P(A/A), then the data A maximally confirm B, because they raise its probability from whatever its prior probability was to one. Even though it is possible that the theory B is just paraphrasing the data, that does not prevent B from being validated by A.
Comment by Guts — December 16, 2005 @ 1:07 am
December 16th, 2005 at 2:52 am
Hi Guts,
Thank you for your insightful reply.
You covered several points. The first was about what constitutes the data set and what constitutes prediction.
Let's look at General Relativity. Of course, General Relativity made many predictions beyond the then-known data, but let's look at how it "predicted" Mercury's precession. GR did this without being fitted to Mercury's precession. Instead, it made the prediction after being fitted to data from other observations. GR predicted the precession and many other data points based on other small parts of the then-known data set. If for every observation, it had to be totally recalibrated, the theory would have been a useless restatement of the data.
I appreciate that Mike Gene's prediction was inspired by ID, but it was not a prediction of ID. The need for proofreading is a consequence of physics, and it had to be present no matter what the cell's origin. We can't remotely put ID on the same footing as GR.
You also wrote about theories that are merely consistent with the data:
I think I understand what you are saying. However, by this logic, what prevents me from doing research on the eating habits of Pacific sharks, and using the resultant data to raise the posterior probability of a wholly irrelevant theory, e.g., that Greenday will release a new single tomorrow? Greenday's single release schedule is consistent with all shark eating habits. Is Greenday slightly more likely to release a single tomorrow because sharks don't eat sea cucumbers?
If my Greenday theory is more probable after observing the sharks, is it not more probable only by an extraordinarily small (infinitesimal?) amount?
What conditions (if any) do I place on my candidate theory such that its probability is enhanced by the failure of other inconsistent candidates?
You then wrote of paraphrasing:
You appear to be saying that a theory is validated by data even when the theory is the data. I suppose it is tautologically validated by the data, but is that what you mean to say?
Bayes theorem appears to say nothing in the case of restatement of the observations.
P.S. I wasn't able to access that Bayes blog post this evening. I'll give it a go in the morning.
Comment by doctor(logic) — December 16, 2005 @ 2:52 am
December 16th, 2005 at 8:04 am
If the need for proofreading is a consequence of physics, then why did many scientists use a Darwinian explanation to explain why RNA pol didn't have the ability to proofread when it was thought to lack this ability?
Comment by MikeGene — December 16, 2005 @ 8:04 am
December 16th, 2005 at 1:46 pm
MikeGene,
What did they say was their rationale? Was it was based on a belief that additional proofreading had no selective advantage? Starting from that premise, I would have expected ID proponents to concur.
You predicted a second layer of proofreading based on the existence of the first layer. I think one might make this prediction without reference to any theory of origins (just based on dynamical similarities between the transcription processes), however, that in itself does not rule out the inference you were trying to make.
You cite an intelligent design rationale for the additional proofreading. Does that rationale still hold if there is no advantage or necessity for the additional proofreading?
It seems to me that if the proofreading is necessary, then any theory of NDE and ID would have provided it. The same is true if the proofreading were just advantageous. If the proofreading were of no value or were a detriment, then I might expect both NDE and ID to prune it.
In other words, I don't see either NDE or ID preferentially predicting proofreading (or lack thereof), but both are compatible with it (or the lack of it).
This leads me to another question for you. I think most ID proponents would argue that evolution has been a combination of ID and NDE. In that case, wouldn't NDE erase any design rationale footprints that might otherwise be predicted by ID?
If the second proofreading step were unnecessary or detrimental, then NDE would have pruned it even if it had been designed into the prototype.
Or, for another example, what do you think ID predicts about modularity and isolation? When humans design machines, we deliberately create non-interacting modules to make the system more maintainable. Yet, living things are highly non-linear, and changing one little thing (like maybe eliminating some junk DNA) might change the expression of important genes. Living things are like spaghetti code. If you apply NDE to a system that was initially modular, do you end up with spaghetti code anyway? If so, a design footprint has been brushed away by NDE.
I suppose that if this obscuration occurs, then one probably has to go back to looking at whether or not certain structures cannot be generated by NDE (or are unlikely), and we're back at the IC debate.
Comment by doctor(logic) — December 16, 2005 @ 1:46 pm
December 16th, 2005 at 8:24 pm
Drlogic wrote:
It's debatable whether this is an example of accomodation or prediction, however, my point by bringing up Gene's prediction and Einstein's prediction was simply to point out the temporal neutrality and not really to bring up examples of the former.
Drlogic:
There is no reason, other than for biological life, that it "had to be present" and it is no "consequence of physics". There is no physical law that requires proofreading to occur, in fact there are situations where proofreading should not occur. Also, Gene's prediction was not about proofreading in general but about proofreading during transcription . As a matter of fact, before it was discovered, there was a time when it was thought that proofreading does not occur during transcription.
Both Gene's and Einstein's predictions are examples of "postdiction", where the predictive hypothesis was made despite the fact that the data was already known, due to other known data.
Drlogic:
Not yet.
Drlogic:
There is a more formal way to say this but I think it is simply a matter of the data itself constraining the theories. Greenday releasing a single tomorrow is not inconsistent or consistent with shark feeding habits, it is simply irrelevant to it. Bayes theorem serves to update to P(B|A) from P(B). Usually the two probabilities are different, with either being bigger than the other. If P(B|A) = P(B) then this is still a perfectly good use of Bayes theorem, A could still be informative in an overall sense because the probability of some other hypothesis–such as C–could change. For example, P(B|A) = P(B) but P(C|A) ≠P(C).
Drlogic:
What I showed was a case of a basic accomodating theory vs. a predicting theory. It is not necessarily true that accomodating theories are basically just redescriptions of the data, theories are too complex themselves to be described as a "tautology", statements can be tautologies, but not theories, as Andrea Bottaro at PT has pointed out. I'd say you need both in order to have a true well rounded explanation of the data. However, if the theory begins making too many false predictions then you have a "Kuhnian crisis". Even in the worst case scenario, when there are more observations than parameters , the inferential situation is not hopeless. Consider the real-world problem of deciding what type of patients will respond to a drug based on microarrays of 20,000 genes that have been assessed on 100 patients. Actually, the Bayesian view is enormously helpful in such a problem because it allows for formally introducing prior information concerning biological phenomena such as genetic pathways.
Comment by Guts — December 16, 2005 @ 8:24 pm
December 16th, 2005 at 8:31 pm
Dr. Logic:
"We can't remotely put ID on the same footing as GR."
Incidentally, I doubt we can put any proposition in historical biology on the same footing as general relativity.
Comment by Krauze — December 16, 2005 @ 8:31 pm
December 17th, 2005 at 2:25 pm
Guts,
I don't mean to slight "postdiction" at all. I think postdiction is a very valid way to check an inference under certain conditions. In particular, I think you have to be able to subdivide your existing data, fit your theory to that subset, then predict (or postdict) the rest of the set. This process is like simulating initial observations and confirmatory experiments using subsets of known observations. I think this may be related to the time independent nature of Bayes' theorem you alluded to.
I agree with your statement, but you're only saying that experiments/experiences are useful in constraining theories.
My claim is that hypothesis B is never informative/useful until P(A|B) <> P(A), i.e., until B predicts something about observation. Otherwise, P(B|A) is never updated, and believing B never helps you predict the next data point. B is not an explanation of A unless P(A|B) <> P(A).
I think it is completely consistent with it, but irrelevant to it.
Let's look again at what it means for the data to constrain a theory. A theory has some mathematical structure which relates possible observations. That structure may have a large number of parameters in it that can be fixed by observations.
But what can we say about a theory that makes no predictions, and can therefore accomodate any observation? I think it says that the theory has an infinite number of degrees of freedom, so that any observation can be accomodated by fixing a new parameter. In essence, the parameters of the theory become the observations themselves. There are extremes of prediction and accomodation, and this is one of them. So my claim is that a theory that has no predictions and infinite accomodation isn't a theory at all. It is just a tautological restatement of the data.
If a theory does make predictions, then either it has the same or fewer number of parameters as the data set, or else the theory also states that the effects of the unfixed parameters are small. (My previous statement about parameter counting was too strong.)
Consider your genetic microarray example. If each unique genome had a radically different response to the drug, then there would be no point in doing the exercise (everyone's reaction would be effectively unique). The theory works because it implicitly says that there are far fewer important factors in drug response than there are possible genomes (or dots on the microarray). Without this assumption of the theory, it would fail to be predictive.
I think we basically agree on the requirements of a good theory. What I am trying to do is come to a more formal criteria for deciding when a theory is either irrelevant or a tautological restatement of the data. I think that criterion is the failure of the theory to make any predictions.
The question is, what are the predictions of generic ID, if any?
Do you think that generic ID is predicting that there will be proofreading in RNA transcription even if it isn't advantageous (i.e., that the designer overdesigns)? Or is the additional proofreading not a prediction of ID?
Comment by doctor(logic) — December 17, 2005 @ 2:25 pm
December 17th, 2005 at 5:18 pm
Not sure if this falls within your predictive criteria Dr logic, but check out Jonathan Witt's article on the research Jonathan Wells is conducting thanks to the ID paradigmhere.
Comment by willo — December 17, 2005 @ 5:18 pm
December 18th, 2005 at 9:29 am
doctor:
First, it must preferentially predict the observations, it cannot simply be consistent with them.
What predictions have NDE given us? Either you or someone else linked to a talk origins article (once) but not one "prediction" on that page was due to random variations culled by NS or any other NDE mechanism.
Dennett (and the PBS series Evolution) has already told us that there is no way to predict what will be selected for at any point in time.
With "evolution" we can "predict" minor variations, stasis, and great transformations. IOW anything can be accomodated by "evolution".
BTW ID predicts IC and CSI…
Comment by Joe G — December 18, 2005 @ 9:29 am
December 18th, 2005 at 8:12 pm
Joe G,
I think the talkorigins page did a perfectly good job of explaining the predictions of evolution. What kinds of predictions were you hoping for?
You claim ID predicts IC and CSI, but it doesn't. How is designed complexity researched? Someone points to something they don't understand and say "Hey! That looks too complex to be evolved!" Then, actual scientists figure out an explanation of how it works and how it could have evolved, and then it's not complex anymore. There's no formula for designed complexity that you can compute based only on observations. Designed complexity is determined by the absence (i.e., gaps in)of compelling NDE theories.
However, this is not to say that a specific ID theory can't make predictions. ID theories predict designed utility, not designed complexity. In SETI, forensics, and archaeology, there's an assumed utility to the pattern of intelligent design, e.g., communication, weapons, tools, etc. All you have to do to show utility is to be specific about why the design is useful to the designer. It's not enough to show that it's useful to the survival of the device under study.
The problem ID has is that NDE also predicts "designed" structures, but those structures have only one purpose: survival. And guess what? For almost all of the last 3.8 billion years, the only apparent purpose to life has been survival.
Maybe there is some designer out there who had a hand in creating life on Earth, but it's hard to make a compelling case for this when there's no apparent purpose to life here. There's no obvious utility in life on Earth to any designer capable of designing it. Who plans ahead 4 billion years?
Comment by doctor(logic) — December 18, 2005 @ 8:12 pm
December 19th, 2005 at 9:42 am
doctor:
I think the talkorigins page did a perfectly good job of explaining the predictions of evolution.
I am sure you do. However evolution isn't being debated. This is all about the mechanism and TO's "predictions" could also fit in with any number of evolutionary models.
doctor:
You claim ID predicts IC and CSI, but it doesn't.
ID absolutely predicts IC and CSI. The only way to deny that fact is via ID ignorance.
doctor:
How is designed complexity researched? Someone points to something they don't understand and say "Hey! That looks too complex to be evolved!" Then, actual scientists figure out an explanation of how it works and how it could have evolved, and then it's not complex anymore. There's no formula for designed complexity that you can compute based only on observations. Designed complexity is determined by the absence (i.e., gaps in)of compelling NDE theories.
Yeah right. If we listen to you then there is nothing that can be said to be the product of intentional design unless we directly observed it. It is a good thing reality demonstrates otherwise. Ya see we have strict criteria that has to be met BEFORE coming to a design inference.
doctor:
The problem ID has is that NDE also predicts "designed" structures, but those structures have only one purpose: survival. And guess what? For almost all of the last 3.8 billion years, the only apparent purpose to life has been survival.
The NDE "predicts" any and everything. And how did/ do organisms know to survive?
doctor:
Maybe there is some designer out there who had a hand in creating life on Earth, but it's hard to make a compelling case for this when there's no apparent purpose to life here.
Again your ignorance is not a refutation. Science has given us an apparent purpose- scientific discovery. As in "The same narrow circumstances that allow for our existence also offer the best over all conditions for making scientific discoveries."
It would serve you better if you actually read some ID literature written by IDists. Arguing from ignorance is not a very good position.
Comment by Joe G — December 19, 2005 @ 9:42 am
December 19th, 2005 at 10:46 am
Joe G,
Look, I'm a reasonable guy. Just make a compelling case for your side. That's all I'm asking. If you have a point, I'll concede it. Calling me ignorant is just bad form.
If I observe a structure, what tests do I do to determine that it is IC or CSI? Go ahead and explain to me (in a short paragraph) how it's done without using a process of elimination of mechanistic theories.
Comment by doctor(logic) — December 19, 2005 @ 10:46 am
December 19th, 2005 at 11:39 am
ID is based on three premises and the inference that follows (DeWolf et al., Darwinism, Design and Public Education, pg. 92):
1) High information content (or specified complexity) and irreducible complexity constitute strong indicators or hallmarks of (past) intelligent design.
2) Biological systems have a high information content (or specified complexity) and utilize subsystems that manifest irreducible complexity.
3) Naturalistic mechanisms or undirected causes do not suffice to explain the origin of information (specified complexity) or irreducible complexity.
4) Therefore, intelligent design constitutes the best explanations for the origin of information and irreducible complexity in biological systems.
As I said if you had read ID literature (the type written by IDists) your questions would have been answered. Then when you say the following:
"Go ahead and explain to me (in a short paragraph) how it's done without using a process of elimination of mechanistic theories."
It truly exposes your lack of understanding on things are done. Using a process of elimination is only way to come to a design inference without being biased towards that end. That is in the absence of direct observation or designer input.
Wm. Dembski pg 36 of The Design Inference:
"The principal advantage of characterizing design as a complement of regularity and chance is that it avoids committing itself to a doctrine of intelligent agency.
Defining design as the negation of regularity and chance avoids prejudicing the causal stories we associate with the design inference."
Why we call it the design inference:
pg. 91 of The Design Revolution:
"The prospect that further knowledge will upset a design inference poses a risk for the Explanatory Filter. But it is a risk endemic to all of scientific inquiry. Indeed, it merely restates the problem of induction, namely, that we may be wrong about the regularities (be they probabilistic or necessitarian) which operated in the past and apply in the present."
As for "bad form"- bad form is discussing a topic which you have no knowledge of or false knowledge of.
CSI and IC have been well defined. Not knowing about them is evidence you don't know ID.
Comment by Joe G — December 19, 2005 @ 11:39 am
December 19th, 2005 at 7:04 pm
Drlogic wrote:
Sorry for the delay , I'm cramming for finals. I am sympathetic to the importance of distinguishing between prediction and accommodation. I think we agree on that. The main issue is whether confirmation is understood entirely as a logical (or quasi-logical) relation among propositions or whether it is understood more pragmatically. Those who favor a logical relation typically argue for the equivalence of prediction and accommodation. Ironically, your reply to me about Einstein and Mercury seems to make out a logical distinction. I am more pragmatic about confirmation, so I don't accept the equilvalence.
Because a Bayesian tends to make confirmation relations logical ones, Bayesians generally advocate the equivalence. There is a huge literature on this topic. For a good overview of the issues, see: Prediction versus Accommodation and the Risk of Overfitting. Hitchcock, Christopher; Sober, Elliott. British Journal for the Philosophy of Science, vol. 55, no. 1, pp. 1-34, March 2004
Drlogic wrote:
No, irrelevant theories are not potential explanations of the data, nor are they favored because they were compatible with the evidence. Maybe this example would illustrate what I mean, suppose you show me what looks like a fair coin and begin tossing it. The first three tosses are heads. One potential explanation is that you substituted a two-headed coin without my being aware of it. Call that hypothesis h1. As I interpret it, h1 implies heads on every toss. I think that h1 deserves some small, non-zero probability. Suppose the fourth toss is tails. Then the probability of h1 goes to zero, but I would expect the probability of the following hypothesis to increase ever so slightly: h2 = you substituted a two-headed coin for the first three tosses and a two-tailed coin for the next three.
It seems to me that, no matter what the actual sequence of tosses might turn out to be, there will some hypotheses of these kinds that earn some slight increase in probability on the basis of the data. It was a mistake for me to suggest that all hypotheses compatible with the data have their probability increased. I think it would be a mistake even to claim that all theories that would explain the data have their probability increased by it. But my example is meant to make it plausible that some potential explanations of the data do have their probability increased by the data, largely (if not entirely) because they are compatible with the data.
Drlogic wrote:
However, nature was not constructed like this. A finite number of genes control such important functions as apoptosis. Even in such a complicated and heterogeneous disease as cancer, there are commonalities within subsets of patients. A person's cancer may be unique, but it has similarities with some other cancers.
Drlogic:
Yes, but without order in nature there could be no science at all.
Drlogic:
Lets call your criterion the "parameter counting criterion". I don't believe this is a very good criterion for distinguishing theories from non-theories. For one thing, it does not allow for probabilistic theories (you can never derive data points from probabilistic theories, but only probabilities of data points). Second, your parameter counting criterion is one way that philosophers have tried to work out an account of theoretical simplicity. But there is a lot of disagreement about how to understand theoretical simplicity.
I think there is something right about what you are saying, which is that, other things being equal, Bayes' Theorem should give more support to simpler theories than to more complex theories. Unfortunately, Bayes' Theorem has no implications one way or the other. It is the prior probabilities to which the theorem is applied that determine whether simpler theories are selectively confirmed over more complex theories or whether more complex theories are selectively favored over simpler ones. So your constraint is a constraint on prior probabilities, not a consequence of Bayes' Theorem.
Drlogic:
I don't know what "generic ID" means. What I did was show one , out of many, valid and successful prediction of an ID hypothesis.
Drlogic:
I think that it's selective value needs to be taken into account (i.e., why would a designer incorporate a design that would be weeded out?), but when selective value was the only thing that was taken into account, or rather when the specificity/apparent want to be as accurate as possible and employing checkpoints was not taken into account, (which natural processes do not really care about, but engineers do) then it lead to a false prediction from Darwinian theory.
Comment by Guts — December 19, 2005 @ 7:04 pm
December 20th, 2005 at 1:34 pm
Hi Guts,
(Good luck with finals, BTW)
If we look at your coin toss example, both h1 and h2 are predictive. But suppose you have h3 = "the next coin flip will be heads or tails with unknown probability". h3 is consistent, but not in any way predictive. In fact, it's guaranteed. It also has the property of looking like the original question, i.e., we ask "is the next flip more likely to be heads or tails?" and h3 comes back with "the next flip will be either heads or tails, but I can't say with what probability it will be heads."
There will be some theories that are not tested by certain experiments because those theories are consistent with any outcome in those particular experiments. However, we should demand that those theories are inconsistent with some possible experimental outcomes somewhere.
By generic ID, I mean an ID theory that says nothing about the abilities or motivations of the designer, nor the utility of artifacts to the designer.
Can you elaborate on this further?
If there is a selective advantage to proofreading, then non-directed evolution "cares" about it. If an engineer cares about something that natural processes do not, then I would expect the feature to have a purpose beyond mere survival. Is that what you are suggesting here?
Comment by doctor(logic) — December 20, 2005 @ 1:34 pm