If your GoldenHammer is so great, why didn't evolution use it?
Because evolution is the SurvivalOfTheSurvivors and we haven't had that many generations yet to let statistics kick in.
Some have implied that relational is "objectively better" (paraphrased) than other approaches as claimed in [insert links]. Although I personally prefer relational in general, I will not commit to claiming it objectively better. It just may be a personal preference on my part.
But if relational is universally superior, then
why didn't brains evolve to use it? Wouldn't evolution find the near-optimal solution eventually and dispense with the navigational-like structure it appears to have? The brain is the most sophisticated database known. This applies to any other universally claimed
GoldenHammer or
BigIdea, not just relational, by the way. Relational is merely used as a point of reference.
- Evolutionary processes tend to select for local maxima rather than globally ideal solutions. In other words, the approach brains use (whatever that might be -- we don't know) is good enough for the purpose, and good enough to be selected and preserved across generations. It doesn't have to be optimal.
- As mentioned below, if there is a DiscontinuitySpike hump preventing from it moving up to the next organization level, it has not been identified. And evolution often can jump over such humps by borrowing something used for another purpose. For example, the high metabolism mammals need to support a larger brain is thought to originally have been used to "run" the hyper-active nocturnal life-style of the first mammals. Once the constant body temperature mechanism was perfected for the nocturnal life-style, it spread to daytime mammals also. It took the nocturnal life-style to produce the mechanisms needed for the next step.
- Evolutionary processes need not exhibit DiscontinuitySpikes in order to select a local maximum when a global one is theoretically available. If an evolutionary process happens to randomly select an effective (but not ideal) solution, it will be preserved unless environmental pressures force it to do otherwise, or an even better solution is (eventually) randomly selected.
One possible claim is that the brain is an analog device and that relational cannot be analog. However, digital can (must?) be emulated with analog. One could argue that there is a big
DiscontinuitySpike that evolution must jump over to get a sufficient advantage from digital. But this impediment itself has not been identified so far.
One may also argue that biology requires a high degree of tolerances ("fuzziness") such that the exactness needed for relational to be advantageous cannot exist in biology. But this alone may say something about relational's ability to adapt to a fast-changing or poorly-understood domain. Something that works well in a pristine environment may not work well in churning environment. Some seem to have the opinion that if one is "smart enough", they can precisely encode the rules of a domain such that safe invariants (base idioms) can be found on which to build a reliable model on top of. While this may be mostly true of chemistry and physics, it appears to not be for domains based on human-generated rules (law, marketing, etc.). Thus, the RelationalEvolutionPuzzle may tell us something about the limits of relational.
--top
And, top, is there any more sheer speculation you feel is worthy of a distinct page?
I'd note that it is very well known (via experiment with a variety of evolutionary and genetics-based algorithms) that evolution-based approaches lead only to local maxima in the fitness function. It is not feasible to compete with the local maxima in any slow evolution towards a more global maxima - that would take a revolutionary change (e.g. a mutation) that survives.
- Please clarify.
- Which part did you fail to comprehend? That something with lower fitness won't compete with something with higher fitness? That local maxima literally means you have higher fitness than any evolutionary path away from the maxima? That mutations must not only occur but also survive to be passed on? Please clarify what you want me to clarify.
- You seem to be ruling out mutations. And as described nearby, often evolution "barrows" from other features or adaptations. The warm-blooded lifestyle of a mammal is an example. It took a nocturnal lifestyle to "push" animals out of the local maximum of the more frugal "cold blooded" design. (Although birds also developed warm-bloodedness eventually also.)
- What sort of hand-waving mutation would put the brain on a path towards relational? I mentioned below this seems unlikely at best. I'm not ruling out mutations (those ARE part of the same genetics-based algorithms mentioned above), but mutations that take you near to another local fitness maxima are extremely rare (they usually don't happen with one small tweak... i.e. you're talking about multiple-mutations coming together at the same time). Most often they just kill you or diminish your fitness (e.g. make you less likely to survive or unable to breed). Even if your fitness remains good, you're once again subject to the same evolutionary pressures and competition from those in the local maxima - the only ones you're able or likely to breed with - for generations. This tends to suppress the mutation and drag descendents back towards the local maxima. Such mutations, suppressed within a population, are most likely to express themselves only in isolated populations (no 'new blood', lots of interbreeding) or where the environment is changing (fitness function is fluctuating, moving the minima and maxima so different people survive or die).
Unless one can demonstrate that there is a clear evolutionary path (one not starting in or crossing a local maxima) between neural-networked associative memory and relational, there is simply no reason to believe evolution is even capable of such a change; the issue of whether relational is better or worse is quite irrelevant if there is no evolutionary path from one to the other. Thus, Top's whole line of speculation is moot... culled by Occam's razor, as it were.
- I realize that about "path blockage" and accept it as a decent explanation to the puzzle. However, the bottleneck itself has yet to be identified and it is interesting to speculate on what it is. Learning about this alleged hard-to-cross ravine(s) may tell us something useful about computation and memory.
- What you fail to realize is that every local maxima is a discontinuity-spike or "path blockage" for evolution. A local maxima is a cozy little place of "I'm better than everyone else around me." In the evolutionary sense, this means that the local maxima out-competes others when it comes to breeding and passing forward one's genetic code, and thus (over a few thousand years) usually takes over - even a 2% breeding & survival advantage becomes a million-to-one population advantage after 700 generations. Minor evolutionary variations that compete for the same resources, being less competitive than the local maxima, will simply shrink in relative population to nearly complete obscurity, and be far more subject to extinction. In any case, such local maxima are likely to be extremely common in any complex fitness function... so common that it becomes far more reasonable to assume there is one and demand that you demonstrate (if you wish to call it "alleged" and such) that there isn't. If I were to speculate, I believe that neural-network based associative memory with a few hormone-based modes is almost certainly already a local maxima. Most evolutionary paths will focus on adding more neurons, improving or reducing their strength, etc. Most revolutionary paths will add and remove hormones and receptors, or restructure the initial brain construction. I can't even see any paths, revolutionary or evolutionary, that lead towards relational.
As to whether one is better than another: relational clearly offers much better recall; neural-networked associative is based entirely on recognition-based triggers - i.e. you cannot, on demand, provide a list of all the phone-numbers you remember. Even in simulated neural-network associative memory this is a serious problem: you cannot ask a neural network to tell you what it knows. You can only ask it questions and receive answers possessing associative triggers. If the intent is to support arbitrary recall, brains make truly awful databases.
- [Brains, and their underlying memory representations, are well-suited to enabling a creature to survive. Iterating all known telephone numbers is rarely a life-or-death matter. Obtaining all relevant knowledge applicable given a set of sensory inputs, however, might be. When you see a fire, you want to immediately remember that it's hot, and can hurt you, and can spread and be put out with water, and if the wind comes from that direction the fire might go that way, and so on. This sort of retrieval -- obtain all the various and sundry knowledge linked to a stimulus, possibly ranked by relevance -- is something that relational databases handle poorly. For example, given an existing RelationalLanguage of your choice, create a universal query that will retrieve all tuples that contain the word "fire" from all RelVars. To start you exploring the difficulty: assuming you could issue such a query, what would the ResultSet look like, and what implications does it have?]
- Pattern and relevance-indexed relational isn't too far beyond state-of-the-art; it could happen today if someone wished it (using technologies from search-engines). But the issue of collecting data from many different relvars is a real one. For AI and data mining, my own research leads me in the direction of a more generic DataSpace.
- [My research leads me to believe that the outcome of your research is a common one.]
- If it stinks so bad, that just reinforces the question on why it didn't shift up to relational.
- And the issue of local maxima is a sufficient and complete answer to that question.
Apparently someone else was providing the same answer above and stomped my edit, but I'll leave it here anyway.
It's worth noting that higher intelligence, greater recall, etc. among humans has not been demonstrated to be an evolutionary advantage - i.e. not when it comes to procreation and survival. Indeed, if anything, the opposite seems to be true: persons of lower education (which is only a very loose metric for intelligence) tend to have more children and start having children earlier. But that isn't really sufficient to make good conclusions. As one song goes: "been around the world and found that only stupid people are breeding, cretins groaning and feeding, and I don't even own a TV" - Flagpole Sitta by Harvey Danger.
An IQ test is not a very good test of all possible kinds of "intelligence". Social intelligence and athletic intelligence may require powerful "brain-math", but its not something normally measured in intelligent tests. Sales experts are some of the most highly paid people there are. Thus, our economic system values them. And, the fact that many geeks do poorly at social intelligence suggests that one cannot master all kinds of intelligence at the same time and there is only so much CPU cycles to go around.
... Now feel free to relate this currently irrelevant passage to the RelationalEvolutionPuzzle.
PageAnchor: R-History
The RelationalModel has only been claimed to be "objectively better" in certain contexts -- such as compared to certain hierarchical models for general-purpose database management -- and even then there are exceptions. I know of no claim that the RelationalModel is universally "better", for any definitions of "universal" or "better".
I'd be interested to know what these "certain contexts" were. The claims were beyond hierarchical DB comparisons.
The notable contexts that spurred the development of the relational model were, in particular, the database systems of the late 1960s and early 1970s that were based on network and hierarchical models, and that were particularly tied to specific system architectures. The relational model was (eventually) considered superior due to its simplicity, theoretical rigour (which permitted the creation of provably-correct automated optimisations), and system independence.
It is possible to build an architecture-independent network DB also, but these didn't do so well in the market-place. And I'd use relational even if it didn't have "integrity" features such as ACID and cascading deletes, for it would be no worse than the alternatives.
Who's to say our brains don't use relational techniques somewhere? What I find extremely odd, is that some people assume our brains are using some sort of functional code in them. But if you analyze a cell, which is what our bodies and brains are composed of, it is actually an abstract structure or object. A cell has hidden parts inside it, such as mitochondria structures, walls that let things in and out. It's very structural and modular.
There is a communication system that runs through the body and sends physical and electrical items into these cells - which is what the object oriented paradigm was actually originally trying to get at with it's message passing cellular analogy. A cell also has a number of procedures that it executes based on input and output. A cell's walls fit well with Wirth's definition of defining modules to create walls.
Who's to say that cells and/or our brain do not also utilize relational techniques? Who's to say that when we think, in our brains there are not some relational searches or queries happening? Is there proof that relational is not used in our brains or was that just someone's best guess?
First, argumentum ad ignorantiam is worth even less than reasoned speculation. Second, I take it you've never studied human memory or psychology? There is evidence to indicate we have nothing like 'tables' of data (give a list of all the people and phone numbers you remember! right now! - it's impossible for a human) and we certainly don't do anything like joins on them (MRI scans have never revealed anything like cross-brain memory interactions).
Stop thinking in boxy tables (typical for an amateur who hasn't studied relational model) and start thinking about relationships. If you try to remember where your pencil was last seen, you relate it to your desk, your floor, your pencil case, your container, your binder, your organizer, your pocket. It is not a single tree hierarchy with inheritance.
Relational is about mathematical relationships, not mental associations. And these relationships are very 'boxy' - strictly speaking, a relation - even a constructed relation or view - is a set of tuples all of identical arity. And you seem to be basing your argument upon a FalseDichotomy: there are a great number of non-relational data storage approaches that are also not "single tree hierarchies with inheritance". Associative memory, neural networks, and tuple spaces are among them.
Quote:
- "We suspect (Crick and Koch, 1995c) that meaning derives both from the correlated firing described above and from the linkages to related representations. For example, neurons related to a certain face might be connected to ones expressing the name of the person whose face it is, and to others for her voice, memories involving her and so on, in a vast associational network, similar to a dictionary or a relational database." --http://www.klab.caltech.edu/~koch/crick-koch-cc-97.html
Biological beings tend to use FuzzyPredicate for lack of a better description. In other words, the best match within a given amount of time. This differs from the all-or-nothing matches of traditional predicate queries. Somewhere around this wiki is a related topic where SOM's were discussed. --top
JuneZeroEight
Relational Evolution Puzzle