These postdictions are not predictions, I challenge you actually pose a testable prediction/hypothesis for this pseudo-science or provide real reliable examples. “Just so” stories is an excellent category for this “science”.
What about the prediction that people would (statistically) sacrifice themselves for three brothers but not one, or for nine cousins but not three? Would this qualify, provided that these specific numbers were empirically observed? After all, no competing theory makes such precise numerical predictions, to my knowledge. So, if observations were to bear out these numbers, then that would provide strong Bayesian evidence for the evolutionary origins of this kind of altruism.
Also, some of the things that you’re calling “postdictions” are not universally acknowledged to be facts — e.g., claims about psychological differences between men and women. So, to the extent that convincing empirical evidence for these differences ultimately arises, wouldn’t that qualify as an honest prediction of evolutionary psychology?
Could you find examples of societies who act differently? Yes. Can culture twist/avoid Kin altruism? If so, I can also invent an evolutionary story to fit that culture just as easily.
Does EP explain all of these different cultures via natural selection? I did not find any so far.
Evolutionary biology always seems to “explain” a narrow provincial behavior and always in postdictions.
What is satisfying? Something accurate enough avoiding ambiguity, taking in account of all of the facts & and provoing an accurate account of the actual cause of behavior (different cultures, sociology, different possible causes—i.e. actually proving it’s bio-psy-evolutionary roots that drive such -detailed behaviors- & not local cultures).
I could understand wider impulses as relatively probable & testable, but “sacrifice themselves for three brothers but not one”, that is one huge kind of a detalied leap. Since when by the way observation is enough? You need to determine the actual cause from all the other possible ones.
I find this to be a major obstacle for the success of this enterprise as a science.
My guess is that simulations of group evolution have already been done with agent based models with simple agents (relative to the both us and the robots they were modeling in the article the grandparent linked). I would guess the answer to your question is yes, and if someone has already done something with these simpler agent based models and made group evolution appear, then I would put a fairly high probability on that ‘yes.’
And if no one’s done it, that just makes me want to do it...
The eating each other’s babies thing is what you get when group selection doesn’t appear.
But yes, simulations of primitive agent models in this sort of context has been done to death. Even before they did experiments with actual living creatures!
I’ve done this kind of experiment myself—more or less as play or tinkering with my research tools. And not so much baby eating as replication suppression via within-group sabotage. (My Masters work was in nature inspired collective intelligence. So not trying to be simulation of reality but relevant in the opposite direction.)
What I found curious (or amusing) was that they were doing it with actual robots. Something about it just seems so cute. :)
The eating each other’s babies thing is what you get when group selection doesn’t appear.
Is that right? I may be misremembering, but I thought that the eating babies thing was a result of group selection. Eating babies arose when a group was under group-selection pressure to have a smaller population. The group complied with the pressure by having individuals within the group consume babies within the group. It was just that the individuals ate the babies of other individuals, thereby acting in accord with both the group-selection pressure and with an individual-selection pressure.
The eating each other’s babies thing is what you get when group selection doesn’t appear.
Ahh, that’s right.
What I found curious (or amusing) was that they were doing it with actual robots. Something about it just seems so cute. :)
I was disappointed when I read the article because it said they were doing it with models of the robots in a simulation. Real robots would have been infinitely cooler (and probably too impractical).
That’s closer, but unfortunately, that does nothing for my “coolness” sensors. I’m imagining a room filed with robots constantly reproducing and dieing, while interacting in interesting ways. I kind of want a room where this is going on in my house...
You assume implicitly a strict connection between cognition and evolution
That assumption follows pretty straightforwardly from evolution and some trivial observations. If I meet a human, I expect with high likelihood that they will have cognitive capabilities in the ballpark with mine. If I meet a lobster, tree, or rock, then I expect with high likelihood that they will not have such cognitive capabilities. I assume that relationship is based on the way the thing is constructed.
You’re assuming evolution (and natural selection) as a basis for shaping cognition with the robots, and then produce altruism.
Why not assume culture and program robots like that and call it evidence for cultural shaping which has nopthing to do with evolution?
My point is, you’re trying to prove cognitive evolution via...cognitive evolution while assuming evolution shapes cognition.
And to remind you, altruism is culturally shaped. See for example capitalism VS communism or the cult named radiofreedomain which advocates acting towards family as any other human being.
My point is, you’re trying to prove cognitive evolution via...cognitive evolution while assuming evolution shapes cognition
No. He’s saying “if evolution shapes cognition this is what we would expect. Oh, hey, look! We see that.”
Incidentally, I suspect that even you don’t think that evolution and cognition are completely unrelated. Different species are in different niches with different degrees of intelligence, and there’s a heavy correlation between what sort of niche a species is in and how intelligent it is. Thus for example, omnivores are generally more intelligent than other species of similar size.
I could understand wider impulses as relatively probable & testable, but “sacrifice themselves for three brothers but not one”, that is one huge kind of a detalied leap. Since when by the way observation is enough? You need to determine the actual cause from all the other possible ones.
You are right that the “three brothers but not one” bit is detailed. That is why observing such specific numbers would provide strong support for the theory, even if you didn’t “determine the actual cause from all the other possible ones”. Mere observation is enough. That is the essence of Bayesian epistemology.
In general, suppose that a theory T says that a highly-specific (and hence a priori improbable) observation E is likely, and then E is actually observed. Then that observation makes the probability of T increase by a very large factor. And the probability of T increases more, the more specific E is. In symbols, if p(E) is small, but p(E|T) is large, then the ratio p(T|E) / p(T) is very large. This is a direct corollary of Bayes’s theorem: p(T|E) = p(T) * p(E|T) / p(E).
Note that this applies even if you merely observed E, but didn’t determine what caused E to happen. (However, if you subsequently did determine what caused E, and that cause differed from what T said it would be, then T would lose whatever favored status it had gained.)
Simplified & short; If P, then Q.
Q.
Therefore, P.
The question remains, postdictions or predictions? I observe a certain group of people in a culture doing something, then I postdict it with EvPsy or alien control.
I observe many people dying around age 80. My theory is that if alines exist, they kill people around age 80.
A postdiction with observation, is utterly worthless. It is “just so” storytelling.
Observation is not enough in our case, take a walk to the the faculty of sociology.
And yet, establishing casual links & correlations isn’t important?
EvPsy have no apparent correlation to these behaviors, or cultures who motivate/shape them.
Can you think of any falsifiable test for this explanation? Is this science?
There have been others explaining why “If P, then Q. Q. Therefore, P.” isn’t what is going on here, and why postdictions are not so bad.
But I’d like to address another issue: there are a lot of historical examples where all the major evidence for a hypothesis was a postdiction, but the hypothesis was so simple and fit the data so well that it became accepted mainly on the power of the postdictions. The most famous example was Kepler’s model of the solar system using ellipses. Based to a large extent on their postdictive power, the basic elliptical model was largely accepted before Newton gave an explanation for why it worked. This example isn’t perfect because the model did provide further confirmation by later astronomical observations.
Simplified & short;
If P, then Q. Q. Therefore, P.
While propositional logic may be a special case of Bayesian reasoning, the Bayes’s theorem formalization of the scientific method cannot be usefully reduced to propositional logic.
Simplified & short; If P, then Q. Q. Therefore, P.
That would be a logical fallacy. But, importantly, its probabilistic analogue is not a fallacy. It really is a mathematical fact that,
If P makes Q more likely, then Q makes P more likely.
In other words, it is a mathematical fact that p(Q|P) > p(Q) implies p(P|Q) > p(P).
I agree with you that predictions are better than postdictions in practice, but postdictions ain’t nothing.
I observe a certain group of people in a culture doing something, then I postdict it with EvPsy or alien control. I observe many people dying around age 80. My theory is that if alines exist, they kill people around age 80.
First, see my final parenthetical remark in my previous comment. We already have causal accounts of why people die around 80. Alternative causal accounts (such as aliens) don’t get much of a probability boost from explaining what we can already explain. In contrast, no competing theory predicts the “three brothers but not one” numbers specifically. If observations bore this out, the EvPysch explanation would not be competing with any alternative explanations.
Second, recall that I said that, when a theory says that an observation is likely, and the observation actually happens, then
that observation makes the probability of T increase by a very large factor.
That is true. Nonetheless, if T started out as very, very improbable, then even an increase by a “very large factor” will still leave T with a small probability. If epsilon is sufficiently small, then epsilon x 10^100 is still very small. Now, “Aliens kill people around age 80″, starts out with a very low prior probability. So it will have to predict/postdict some very improbable observations to rise above a negligible probability.
Third, and most importantly, simply adding an improbable observation to a theory lowers the prior probability of that theory. Take the theory “Aliens kill people”. Now augment the theory by adding the “around age 80″ part to get “Aliens kill people around age 80”. This addition lowers the probability of the theory. (Under the original theory, the aliens could be killing people at any age. Thus, the original theory would be true under a wider variety of circumstances, so it is more probable.)
In fact you can prove that the addition of “around age 80” exactly counteracts the boost that the augmented theory gets for successfully post-dicting that people die around age 80. You don’t gain any probability for your theory by explicitly building post-dictions into it. In symbols, let T be the original theory, and let T&E be the augmented theory. It follows that, if p(E|T) = p(E), then p(T&E | E) = p(T). That is, if the original theory didn’t make E any more likely, then observing E doesn’t make the augmented theory any more probable than the original theory was prior to the observation of E. And “Aliens kill people” started out pretty improbable!
In contrast, the “three brothers but not one” numbers are not just added to EvPsych. They are deduced from simpler premises. So, if observations bear these numbers out, then that really is a big boost to EvPsych.
In fact you can prove that the addition of “around age 80” exactly counteracts the boost that the augmented theory gets for successfully post-dicting that people die around age 80.
Somehow I missed that as a relevant fact when recently trying to explain this stuff to a Popperian. Thanks!
If P, then Q is plausible. Q. Therefore P is plausible.
And it’s a valid argument in probability theory as extended logic; see the first chapter of Probability Theory: The Logic of Science, which is available on the linked webpage.
What about the prediction that people would (statistically) sacrifice themselves for three brothers but not one, or for nine cousins but not three? Would this qualify, provided that these specific numbers were empirically observed? After all, no competing theory makes such precise numerical predictions, to my knowledge. So, if observations were to bear out these numbers, then that would provide strong Bayesian evidence for the evolutionary origins of this kind of altruism.
Also, some of the things that you’re calling “postdictions” are not universally acknowledged to be facts — e.g., claims about psychological differences between men and women. So, to the extent that convincing empirical evidence for these differences ultimately arises, wouldn’t that qualify as an honest prediction of evolutionary psychology?
Could you find examples of societies who act differently? Yes. Can culture twist/avoid Kin altruism? If so, I can also invent an evolutionary story to fit that culture just as easily. Does EP explain all of these different cultures via natural selection? I did not find any so far. Evolutionary biology always seems to “explain” a narrow provincial behavior and always in postdictions.
What is satisfying? Something accurate enough avoiding ambiguity, taking in account of all of the facts & and provoing an accurate account of the actual cause of behavior (different cultures, sociology, different possible causes—i.e. actually proving it’s bio-psy-evolutionary roots that drive such -detailed behaviors- & not local cultures). I could understand wider impulses as relatively probable & testable, but “sacrifice themselves for three brothers but not one”, that is one huge kind of a detalied leap. Since when by the way observation is enough? You need to determine the actual cause from all the other possible ones.
I find this to be a major obstacle for the success of this enterprise as a science.
Would it help to know you could generate altruism in robots just by putting them in a simulation of evolution?
If I put them in a simulation of group evolution do you think I could make them eat each other’s robot babies?
My guess is that simulations of group evolution have already been done with agent based models with simple agents (relative to the both us and the robots they were modeling in the article the grandparent linked). I would guess the answer to your question is yes, and if someone has already done something with these simpler agent based models and made group evolution appear, then I would put a fairly high probability on that ‘yes.’
And if no one’s done it, that just makes me want to do it...
The eating each other’s babies thing is what you get when group selection doesn’t appear.
But yes, simulations of primitive agent models in this sort of context has been done to death. Even before they did experiments with actual living creatures!
I’ve done this kind of experiment myself—more or less as play or tinkering with my research tools. And not so much baby eating as replication suppression via within-group sabotage. (My Masters work was in nature inspired collective intelligence. So not trying to be simulation of reality but relevant in the opposite direction.)
What I found curious (or amusing) was that they were doing it with actual robots. Something about it just seems so cute. :)
Is that right? I may be misremembering, but I thought that the eating babies thing was a result of group selection. Eating babies arose when a group was under group-selection pressure to have a smaller population. The group complied with the pressure by having individuals within the group consume babies within the group. It was just that the individuals ate the babies of other individuals, thereby acting in accord with both the group-selection pressure and with an individual-selection pressure.
Ahh, that’s right.
I was disappointed when I read the article because it said they were doing it with models of the robots in a simulation. Real robots would have been infinitely cooler (and probably too impractical).
They checked their simulation occasionally by using real robots with the same programs ever so often.
That’s closer, but unfortunately, that does nothing for my “coolness” sensors. I’m imagining a room filed with robots constantly reproducing and dieing, while interacting in interesting ways. I kind of want a room where this is going on in my house...
Now I’m imagining a residence rigged up with a robot-friendly Habitrail network.
YES! That’s definitely the sort of thing I’m thinking of!
You assume implicitly a strict connection between cognition and evolution, while that is what we are trying to prove.
That assumption follows pretty straightforwardly from evolution and some trivial observations. If I meet a human, I expect with high likelihood that they will have cognitive capabilities in the ballpark with mine. If I meet a lobster, tree, or rock, then I expect with high likelihood that they will not have such cognitive capabilities. I assume that relationship is based on the way the thing is constructed.
Or am I misunderstanding you?
You’re assuming evolution (and natural selection) as a basis for shaping cognition with the robots, and then produce altruism. Why not assume culture and program robots like that and call it evidence for cultural shaping which has nopthing to do with evolution?
My point is, you’re trying to prove cognitive evolution via...cognitive evolution while assuming evolution shapes cognition. And to remind you, altruism is culturally shaped. See for example capitalism VS communism or the cult named radiofreedomain which advocates acting towards family as any other human being.
No. He’s saying “if evolution shapes cognition this is what we would expect. Oh, hey, look! We see that.”
Incidentally, I suspect that even you don’t think that evolution and cognition are completely unrelated. Different species are in different niches with different degrees of intelligence, and there’s a heavy correlation between what sort of niche a species is in and how intelligent it is. Thus for example, omnivores are generally more intelligent than other species of similar size.
You are right that the “three brothers but not one” bit is detailed. That is why observing such specific numbers would provide strong support for the theory, even if you didn’t “determine the actual cause from all the other possible ones”. Mere observation is enough. That is the essence of Bayesian epistemology.
In general, suppose that a theory T says that a highly-specific (and hence a priori improbable) observation E is likely, and then E is actually observed. Then that observation makes the probability of T increase by a very large factor. And the probability of T increases more, the more specific E is. In symbols, if p(E) is small, but p(E|T) is large, then the ratio p(T|E) / p(T) is very large. This is a direct corollary of Bayes’s theorem: p(T|E) = p(T) * p(E|T) / p(E).
Note that this applies even if you merely observed E, but didn’t determine what caused E to happen. (However, if you subsequently did determine what caused E, and that cause differed from what T said it would be, then T would lose whatever favored status it had gained.)
Simplified & short;
If P, then Q. Q. Therefore, P.
The question remains, postdictions or predictions? I observe a certain group of people in a culture doing something, then I postdict it with EvPsy or alien control. I observe many people dying around age 80. My theory is that if alines exist, they kill people around age 80. A postdiction with observation, is utterly worthless. It is “just so” storytelling. Observation is not enough in our case, take a walk to the the faculty of sociology. And yet, establishing casual links & correlations isn’t important?
EvPsy have no apparent correlation to these behaviors, or cultures who motivate/shape them. Can you think of any falsifiable test for this explanation? Is this science?
There have been others explaining why “If P, then Q. Q. Therefore, P.” isn’t what is going on here, and why postdictions are not so bad.
But I’d like to address another issue: there are a lot of historical examples where all the major evidence for a hypothesis was a postdiction, but the hypothesis was so simple and fit the data so well that it became accepted mainly on the power of the postdictions. The most famous example was Kepler’s model of the solar system using ellipses. Based to a large extent on their postdictive power, the basic elliptical model was largely accepted before Newton gave an explanation for why it worked. This example isn’t perfect because the model did provide further confirmation by later astronomical observations.
While propositional logic may be a special case of Bayesian reasoning, the Bayes’s theorem formalization of the scientific method cannot be usefully reduced to propositional logic.
Also, welcome to Less Wrong!. It sounds like you may want to check out Bayes’ Theorem and/or Technical Explanation.
Thank you for the kind welcome. Will read.
That would be a logical fallacy. But, importantly, its probabilistic analogue is not a fallacy. It really is a mathematical fact that,
In other words, it is a mathematical fact that p(Q|P) > p(Q) implies p(P|Q) > p(P).
I agree with you that predictions are better than postdictions in practice, but postdictions ain’t nothing.
First, see my final parenthetical remark in my previous comment. We already have causal accounts of why people die around 80. Alternative causal accounts (such as aliens) don’t get much of a probability boost from explaining what we can already explain. In contrast, no competing theory predicts the “three brothers but not one” numbers specifically. If observations bore this out, the EvPysch explanation would not be competing with any alternative explanations.
Second, recall that I said that, when a theory says that an observation is likely, and the observation actually happens, then
That is true. Nonetheless, if T started out as very, very improbable, then even an increase by a “very large factor” will still leave T with a small probability. If epsilon is sufficiently small, then epsilon x 10^100 is still very small. Now, “Aliens kill people around age 80″, starts out with a very low prior probability. So it will have to predict/postdict some very improbable observations to rise above a negligible probability.
Third, and most importantly, simply adding an improbable observation to a theory lowers the prior probability of that theory. Take the theory “Aliens kill people”. Now augment the theory by adding the “around age 80″ part to get “Aliens kill people around age 80”. This addition lowers the probability of the theory. (Under the original theory, the aliens could be killing people at any age. Thus, the original theory would be true under a wider variety of circumstances, so it is more probable.)
In fact you can prove that the addition of “around age 80” exactly counteracts the boost that the augmented theory gets for successfully post-dicting that people die around age 80. You don’t gain any probability for your theory by explicitly building post-dictions into it. In symbols, let T be the original theory, and let T&E be the augmented theory. It follows that, if p(E|T) = p(E), then p(T&E | E) = p(T). That is, if the original theory didn’t make E any more likely, then observing E doesn’t make the augmented theory any more probable than the original theory was prior to the observation of E. And “Aliens kill people” started out pretty improbable!
In contrast, the “three brothers but not one” numbers are not just added to EvPsych. They are deduced from simpler premises. So, if observations bear these numbers out, then that really is a big boost to EvPsych.
Somehow I missed that as a relevant fact when recently trying to explain this stuff to a Popperian. Thanks!
Actually, it’s:
If P, then Q is plausible. Q. Therefore P is plausible.
And it’s a valid argument in probability theory as extended logic; see the first chapter of Probability Theory: The Logic of Science, which is available on the linked webpage.