Must the designer be more complex than the design?

One of Dr. Richard Dawkins' favorite arguments against Intelligent Design's coherence is that ID does not explain complexity because the designer must be even more complex than the design.

Why does Dawkins say this?  For instance, he believes great complexity comes from very simple origins through the process of Darwinistic evolution.  But, for some reason introducing a designer implies greater preceding complexity.

While I am not sure why Dawkins makes this claim, I can address a reason why he may, and the problem with this reason.

First, he may be thinking of the designer as some complex physical entity, such as a factory.  In this case it is quite obvious that to generate even a seemingly simple object such as a pencil we need an enormously complicated number of processes and mechanisms.  So, it is quite easy for me to accept that if the designer were like a factory, then it in turn would require even more explanation than the pencil.  In which case, intelligent design theory would not be very helpful.

However, intelligent design theory does not say the designer is like a factory.  In fact, it precludes the designer from being like a factory.  To see this, we must examine the core concept of ID, which is complex, specified information.

Complex, specified information (CSI) is a mathematical quantification of an entity.  The two criteria for an entity to possess CSI is that it must be highly unlikely (complex) given the environment in which it came to exist, while also precisely and concisely described by a specification that is independent from its environment.

For a causal agent to be the intelligent designer responsible for the CSI in the entity, the entity must abide by the two criteria in the context of being generated by a particular agent.  Take the pencil factory as an example, since pencils are quite evidently designed, are relatively complex and can be described quite simply as an erasing and writing instrument.  Can we say the pencil factory is the designer of the pencil?

Well, let's look at the criteria for CSI.  Would we consider it highly unlikely for a pencil factory to generate pencils?  Probably not, unless it is a particularly bad pencil factory.  Next, do we consider the pencil's description to be independent from the factory?  In other words, does the pencil factory produce something that is better described as totally unlike a pencil?  Again, probably not, unless it is a particularly bad pencil factory.  As such, the pencil factory cannot be said to be the designer of the pencil.

In this way we see that if the supposed intelligent designer is to a designed entity like a pencil factory is to a pencil, then ID states the supposed intelligent designer is not the real designer.  Rather than disagreeing with Dawkins, ID actually agrees and says the designer cannot be like a pencil factory in relation to its design, ever growing in complexity.

Instead, the designer must be quite independent from its design.  This means that a design implies nothing about the complexity of the designer.  While it may well be the case that the designer is more complex, this is not necessitated by ID and the designer may be much, much simpler than the design.

In fact, the foregoing argument logically entails that there is more to intelligent design than just complex specified information:
http://appliedintelligentdesign.blogspot.com/2012/07/intelligent-design-goes-beyond.html

Background, experiment and results of CSI Collecting (CSIC)

The essential idea of CSI Collecting is to incorporate human pattern detection to improve the capabilities of search and optimization algorithm.

The question is, do humans have the ability to detect patterns any better than computers?  If not, then any advantage human interaction brings is equivalent to some computer algorithm.  Hence, there is nothing especially unique about human involvement, and there is no guarantee it will bring an advantage to a particular problem.

On the other hand, if humans do have a pattern recognition capability that is non-computational then we can expect human involvement to bring an advantage to all problems where an advantage is possible.  There are problems where there is no other way to solve the problem than by trying all combinations, and in such problems human interaction will bring no advantage whether such interaction is non-computational or not.

Why would we think a non-computational ability exists in humans?  This question is a bit more involved to answer.  The answer is a hypothesis, and is based on a couple assumptions which are potentially testable.

The starting point is Dr. Dembski's paper "A search for a search".  This paper states two important points: 1) a search process cannot produce more information than is originally placed in the search algorithm, and 2) a search for search also cannot produce more information than is place in the initial search algorithm.  The term information in these papers refers to a metric of how much more likely a search algorithm will find a search target than a purely random sample based search.

According to these two points, as long as the only means we have of inserting information into a search process is another search process, then we can never insert information into a search process.  This means that for all problem domains any search will perform no better than a random sample search.  The only way to get information into a search process is by a non-search.  However, all algorithms that are capable of inserting information can be characterized as searches.  Thus, the source of information must ultimately be non-algorithmic.  Since all computers are algorithmic, this means that the source of information must also be non-computational.

The one known source of non-algorithmic information is intelligence.  According to intelligent design, intelligence can create information, in particular, complex specified information.  However, is complex specified information the same kind of information that a search process requires?

The mathematical definition of complex specified information is approximately the amount of specificational resources multiplied by the number of specifications we are interested in divided by the number of possible specifications.  This characterization can be fit to the search process.  The range of potential solutions that a search will examine is equivalent to the number of possible specifications.  The target set of solutions is equivalent to the specifications we are interested in.  The specificational resources are the number of solutions sampled by the search process.  If the search process selects the target specifications more often than we would expect given their proportion to the number of potential specifications, then when we calculate the CSI formula on the search process, it will contain information.

Now if we invert the argument that CSI implies intelligence, then we can say intelligence can potentially, but not necessarily, impart information into a search process.  As we saw when comparing the definition of CSI to information in a search process, if information is imparted into a search process, it will find its target more often than without the information.  So, if an intelligent being, such as a person, is involved in a search process, then the intelligent being can cause any given search algorithm to perform better than mathematically possible.

This is the theoretical basis behind the CSI Collecting (CSIC) method.  Since humans have some sort of ability to insert information into any search algorithm, then there must be a general methodology for allowing humans to interact with search algorithms.  Does such a methodology exist in modern computer science?  Currently, there are a number of researchers are attempting to incorporate human interaction into search processes, the most well known being the Google search engine, which uses our search patterns and web page linking to improve our searches results.  However, none of these methods attempt to establish a generalized approach.  The general field is known as human computation, and there is currently no theoretical foundation for the field that is compatible with intelligent design's dual claims that algorithms cannot create information and that intelligent agents can.

Now, you may be asking, given that CSIC seems to be a unique approach to human computation, how do I plan to test my hypothesis, and what are my results?  First, I must admit there are practical problems to testing the efficacy of human computation.  The most glaring problem is distinguishing between A) a general non-algorithmic ability to improve a search algorithm and B) happenstance that an algorithmic ability improves a search algorithm.  In theory the two cases are quite distinct: there will always be instances of search algorithms where A improves the search and B does not, for any type of algorithmic ability in B.  In practice it is not possible to test an ability over all possible problems, we can only test a very small subset of all problems.  However, the good thing is that there are many finite problem subsets that have approximately the same characteristic as the set of all problems.  So, if I pick such a subset, while I cannot categorically state that the tested ability falls in A instead of B, I can make such a claim with a strong degree of confidence.

Which subset of problems have the necessary characteristic?  It is the subset of problems that the No Free Lunch Theorem (NFLT) applies, or almost applies (ANFLT).  Unfortunately, the set of NFLT problems is not very large.  Fortunately, the nest of ANFLT problems is very large.  In fact, most problems of interest quite likely fall into this set because they are NP complete or harder.  So, as long as I pick a problem that is ANFLT, I can at least give a confidence rating to my hypothesis, as well as potentially generate useful results.

The second practical issue is determining whether humans categorically contributed to the search process, or I just picked a bad search algorithm.  This case too I cannot say definitively whether my algorithm is to blame or not.  The best I can do is give a degree of confidence.  And in my case, I cannot even mathematically quantify this confidence since my current work is merely a first pass to see if the idea shows promise.

The general technique I use is as follows.  I use a standard multi-objective genetic algorithm.  A genetic algorithm takes a set of solutions, measures how good each solution is according to some fitness valuation function, varies the solutions to generate a new set using variation operators such as mutation and crossover, and then from both sets of solutions selects a final set according to some critera.  The algorithm then reiterates this process on each subsequent final set until a stopping criteria is reached.  The solutions themselves are represented to the algorithm as fixed length bit strings.  The one innovation I add to create CSIC is to allow a person to provide input for the selection phase.  The person can only see the bit string solution and its fitness valuation.

The metric of comparison between the human and the genetic algorithm is the number of unique solution valuations needed to improve on the best found solution to date.  The search (algorithmic or human manual) process requiring fewer unique solution valuations to find a better solution is considered the superior search process.

My first problem domain consisted of scheduling member roles for multiple club meetings.  This is a form of the well known scheduling problem, and as such is harder than NPC.  My method was to run the genetic algorithm until it stopped finding better solutions for a large number of iterations.  Then I manually attempted to find better solutions.  In about 0.05% of the fitness valuations the algorithmic search used and was still unable to find better solutions, I was able to find three superior solutions.

My second problem domain of choice is that of coming up with the primes that generate an RSA public private key pair.  This domain is much harder than NPC, and as such is quite likely an ANFLT domain.  I ran the algorithmic search algorithm in tandem with 500 attempts by humans using the Amazon Mechanical Turk service.  The humans contributed 2.4% of the overall set of superior solutions.

These results are still preliminary, and there is much hard mathematical work to be done to quantify specifically the degree to which humans contributed, and whether this contribution is statistically significant enough to validate the ID hypothesis.  However, at the very least, it is clear that humans can contribute to algorithmic optimization when given only the exact same data available to the algorithm.  So, the initial results show promise, however so slight.

Is a programmer a computer?

Does the programmer's ability to formulate halting solutions to arbitrarily posed problems signify a non-computational capability in the programmer?  In other words, is the programmer doing something that no computer, nor any other object driven by the laws of physics, can ever do?


First of all, if the solutions to these arbitrary problems are halting solutions in a Turing complete environment, it is trivial to computationally generate halting programs if the generator is cognizant of Turing machines.  In which case, the question still remains whether it is trivial enough to generate a halting solution out of these trivial halting programs.


Granting that the computational generator can trivially generate halting programs, there are a still a number of programs it cannot determine whether they halt or not.  So, the generator has to deal with two sets of programs: P1) those it already knows halt and P2) those it does not know halt.


1.  If P1 is sufficient to generate a halting solution (HS), and P1 is trivial to generate, then it is trivial to solve most if not all programming problems we encounter in the real world.  This is highly unlikely.


2.  If P1 is non-trivial to generate, then it is another set of HS2 programs to find, and is thus a variant of the same initial issue of generating a HS for an arbitrary problem.  In this case, HS2 must already have been found in order to use it to find HS.  What this means is that all the solvable posed problems have already been solved by the programmers before they are even posed.  Pretty neat, but also seems highly unlikely.


Perhaps a non-trivial P1 set can be built up by a previous trivial set, if uses enough trivial steps in between.  However, this is merely a variant of #1.


3.  If the generator also requires P2, to any degree, then it runs into the halting problem, and won't work.  So this option is out. 


Therefore, if the programmer's ability to generate halting solutions to arbitrary problems is computational in nature, then either 1 or 2 must be true.  Since it seems very unlikely either is true, then it is also very unlikely programming is a solely computational activity.

How Dembski's "The Search for a Search" is related to ID

Dr. Dembski and Dr. Marks published a peer reviewed article in an IEEE journal:

http://evoinfo.org/papers/2010_TheSearchForASearch.pdf

First a bit of background.  The evolutionary process has been characterized by Dembski as a form of search for a target.  The question is, can evolution search for a particular target (such as a functioning eye) effectively?  Effectiveness is defined as finding the target more quickly on average than a random search process.

Intelligent Design Theory (IDT) claims that evolution is not inherently better at finding a target than a random search.  In order for evolution to be effective, IDT claims that information about where to find a target must be placed in the search process by something else.

Dembski's paper mathematically addresses whether this "something else" could be yet another search process.  It turns out the answer is: only if the second search process in turn has received information about the target from yet another source.  If we try to account for such information using only searches we become stuck in an infinite regress.

Dembski's paper shows evolution is not an adequate explanation for the kinds of targets biologists claim it finds.  Neither can biologists resort to another search process because they will need yet another search process, and so on, leading to an infinite regress.

The useful conclusion is that something other than a search process is necessary for evolution to find its target, and therefore evolution is not an adequate explanation for the complex functionality we find in biology.

The core claim of ID is that only intelligence can produce the sort of complex functionality we find in biology.  Dembski's paper attacks half of this problem by showing evolution is a insufficient explanation for complex functionality.  Consequently, it is a useful, peer reviewed paper derived from the core claim of ID.

Is ID essential for secularism?

Secularism is defined as a form of government independent from solely religious sources.  It is not a priori biased in favor or against any religious position.  It can also take into account religious truth claims if they are also based on natural arguments.

However, today secularism is considered to be almost synonymous with atheism.  Such equivocation is incorrect, since atheism itself is a religious claim, and any government based on such is consequently religious and not secular.

Why does this equivocation between secularism and atheism exist today?  Well, it is because modern intellectuals have persuaded many that natural arguments for God are meaningless, since science supposedly excludes the supernatural.  But, once the supernatural is excluded, there is no ground for any claims based on theism.  The only claims valid in such a worldview are atheistic claims.  Thus, secularism disappears, since the system is now a priori completely biased in favor of atheism, and religion and state are combined.

ID breaks the stranglehold of atheism on the modern state and opens the door for secularism again.  ID shows that it is possible for natural arguments for God to exist.  Consequently, it is now rationally valid for the state to consider theistic claims.  Atheism no longer monopolizes the state and ID has saved the day.

How ID sheds light on the classic free will dilemma

Copied from Uncommon Descent:

http://www.uncommondescent.com/philosophy/how-id-sheds-light-on-the-classic-free-will-dilemma/

The standard argument against free will is that it is incoherent.  It claims that a free agent must either be determined or non-determined.  If the free agent is determined, then it cannot be responsible for its choices.  On the other hand, if it is non-determined, then its choices are random and uncontrolled.  Neither case preserves the notion of responsibility that proponents of free will wish to maintain.  Thus, since there is no sensible way to define free will, it is incoherent. [1]

Note that this is not really an argument against free will, but merely an argument that we cannot talk about free will.  So, if someone were to produce another way of talking about free will the argument is satisfied.

Does ID help us in this case?  It appears so.  If we relabel “determinism” and “non-determinism” as “necessity” and “chance”, ID shows us that there is a third way we might talk about free will.

In the universe of ID there are more causal agents than the duo of necessity and chance.  There is also intelligent causality.  Dr. Dembski demonstrates this through his notion of the explanatory filter.  While the tractability of the explanatory filter may be up for debate, it is clear that the filter is a coherent concept.  The very fact that there is debate over whether it can be applied in a tractable manner means the filter is well defined enough to be debated.

The explanatory filter consists of a three stage process to detect design in an event.  First, necessity must be eliminated as a causal explanation.  This means the event cannot have been the precisely determined outcome of a prior state.  Second, chance must be eliminated.  As such, the event must be very unlikely to have occurred, such that it isn’t possible to have queried half or more of the event space with the number of queries available.

At this point, it may appear we’ve arrived at our needed third way, and quite easily at that.  We merely must deny that an event is caused by chance or necessity.  However, things are not so simple.  The problem is that these criteria do not specify an event.  If an event does meet these criteria, then the unfortunate implication is so does every other event in the event space.  In the end the criteria become a distinction without a difference, and we are thrust right back into the original dilemma.  Removing chance and necessity merely gives us improbability (P < 0.5), also called “complexity” in ID parlance.

What we need is a third criteria, called specificity.  This criteria can be thought of as a sort of compression, it describes the event in simpler terms.  One example is a STOP sign.  The basic material of the sign is a set of particles in a configuration.  To describe the sign in terms of the configuration is a very arduous and lengthy task, essentially a list of each particle’s type and position.  However, we can describe the sign in a much simpler manner by providing a computer, which knows how to compose particles into a sign according to a pattern language, with the instructions to write the word STOP on a sign.

According to a concept called Kolmogrov Complexity [2], such machines and instructions form a compression of the event, and thus specify a subset of the event space in an objective manner.  This solves the previous problem where no events were specified.  Now, only a small set of events are specified.  While KC is not a necessary component of Dr. Dembski’s explanatory filter, it can be considered a sufficient criteria for specificity.

With this third criteria of specificity, we now have a distinction that makes a difference.  Namely, it shows we still have something even after removing chance and necessity: we have complex specified information (CSI).  CSI has two properties that make it useful for the free will debate.  First, it is a definition of an event that is neither caused by necessity or chance.  As such, it is not susceptible to the original dilemma.  Furthermore, it provides a subtle and helpful distinction for the argument.  CSI does not avoid the distinction between determinism and non-determinism.  It still falls within the non-determinism branch.  However, CSI shows that randomness is not an exhaustive description of non-determinism.  Instead, the non-determinism branch further splits into a randomness branch and a CSI branch.

The second advantage of CSI is that it is a coherent concept defined with mathematical precision.  And, with a coherently definition, the original argument vanishes.  As pointed out in the beginning of the article, the classic argument against free will is not an argument against something.  It is merely an argument that we cannot talk about something because we do not possess sufficient language.  Properly understood, the classical argument is more of a question, asking what is the correct terminology.  But, with the advent of CSI we now have at least one answer to the classical question about free will.

So, how can we coherently talk about a responsible free will if we can only say it is either determined and necessary, or non-determined and potentially random?  One precise answer is that CSI describes an entity that is both non-determined while at the same time non-random.

——————-

[1] A rundown of many different forms of this argument is located here:

http://www.informationphilosopher.com/freedom/standard_argument.html

[2] http://en.wikipedia.org/wiki/Kolmogorov_complexity

Infinite Probabilistic Resources Makes ID Detection Easier (Part 2)

Copied from Uncommon Descent:

http://www.uncommondescent.com/philosophy/the-effect-of-infinite-probabilistic-resources-on-id-and-science-part-2/

Previously [1], I argued that not only may a universe with infinite probabilistic resources undermine ID, it will definitely undermines science. Science operates by fitting models to data using statistical hypothesis testing with an assumption of regularity between the past, present, and future. However, given the possible permutations of physical histories, the majority are mostly random. Thus, a priori, the most rational position is that all detection of order cannot imply anything beyond the bare detection, and most certainly implies nothing about continued order in the future or that order existed in the past.

Furthermore, since such detections of order encompass any observations we may make, we have no other means of determining a posteriori whether science’s assumption of regularity is valid to any degree whatsoever. And, as the probabilistic resources increase the problem only gets worse. This is the mathematical basis for Hume’s problem of induction. Fortunately, ID provides a way out of this conundrum.Not only does intelligent design detection become more effective as the probabilistic resources increase, but it also provides a basis (though not a complete answer) for the assumption of regularity in science.

In [1], I point out that as the resources approach infinity the proportion of coherent configurations in the space of possible configurations approach zero. This is important because Intelligent Design is really about hypothesis testing and specifying rejection regions [2], and coherent configurations allow us to form a rejection region. In hypothesis testing, the experimenter proposes a hypothesis and a probability distribution over potential evidence, signifying what results the hypothesis predicts. If the experiments produce results outside the predicted range to a great enough degree, then the result fall within the rejection region and the hypothesis is considered statistically unlikely and consequently rejected. Note that in this case it is actually better to have more result samples rather than fewer samples. With a few samples the variance is large enough that the results don’t render the hypothesis statistically unlikely. But, with enough samples the variance is reduced to where the hypothesis can be rejected. With an infinite number of samples we can see almost exactly whether the true distribution matches the predicted distribution.

Such is the case with an infinite universe and infinite samples. The infinite universe is composed of all possible configurations, which create a probability distribution over how ordered an arbitrary sample is expected to be. With an assumption of infinite samples (i.e. a conscious observer in every configuration), we can say in what proportion of the configurations intelligent design detection will be successful, which is the inverse of the proportion of ordered configurations. Unfortunately, the number of unsuccessful detections never actually reaches zero, since there will always be coherent configurations as long as they are possible. If I happen to find myself in a coherent configuration I may just be extremely lucky. But, in the majority of configurations the chance and necessity hypothesis will be validly rejected in favor of the intelligent design hypothesis.

At this point it may seem suspicious that I write we can reject a hypothesis in favor of another. Why should rejecting one hypothesis favor another hypothesis? This begins to sound like a god-of-the-gaps argument; just because we’ve dismissed chance and necessity doesn’t necessarily imply we can accept design. There may be yet another alternative we’ve yet to think of. While a good caution, science does not deal with unknown hypotheses. Science deals with discriminating between known hypotheses to select the best description of the data. But, what is the probability distribution over the evidence that ID provides? Well, the specific prediction of ID is that orderly configurations will be much more common than statistically expected. For example, we can see this in that Kolmogrov complexity provides an objective specification for CSI calculations [2]. Kolmogrov complexity is a universal measure of compression, and orderliness is a form of compression. [3] So, when I end up in a configuration that is orderly I have a higher probability of being in a configuration that is the result of ID than in a configuration that is the result of chance and necessity. Hence, an orderly configuration allows me to discriminate between the chance and necessity hypothesis and the ID hypothesis, in favor of the latter. Additionally, since orderly configurations drop off so quickly as our space of configurations approach infinity, then this shows that infinite resources actually make it extremely easy to discriminate in favor of ID when faced with an orderly configuration. Thus, intelligent design detection becomes more effective as the probabilistic resources increase.

Now that I’ve addressed the question of whether infinite probabilistic resources makes ID detection impossible or much, much easier, let’s see whether ID can in turn do science a favor and provide a basis for its regularity assumption. I will attempt to do this in part 3 of my series.

[1] http://www.uncommondescent.com/intelligent-design/the-effect-of-infinite-probabilistic-resources-on-id/
[2] http://www.designinference.com/documents/2005.06.Specification.pdf
[3] Interestingly, Kolmogrov complexity is uncomputable in the general case due to the halting problem. This means that in general no algorithm can generate orderliness more often than is statistically expected to show up by chance. Hence, if some entity is capable of generating orderliness more often than statistically predicted, it must be capabable, at least to some extent, of solving the halting problem.