|
Author
|
Topic: Simulating Self Assembly
|
RBH
Member
Member # 380
|
posted 10. February 2003 10:53
warren offered quote:
Examples of the types of variation-selection processes described above can be 1)created artificially, and 2)can be identified in biological systems. It should not be difficult to think of illustrations of these properties of variation-selection processes. I will be glad to provide illustrations if you can't think of them.
Please do offer illustrations in biological systems. Part of the problem in understanding your ideas is the difficulty of connecting them with concrete examples, in discerning how your model maps into biological phenomena.
warren asked quote:
TA systems involve multiple variation-selection processes occurring simultaneously. This is possible because TA systems involve multiple selection variables being measured or recorded at the same time. Is this or is this not a departure from your concept of GA models?
It is not a departure from my "concept" of GA models. In the GAs my company builds, the artificial critters comprising the population each sample up to 16 environmental variables simultaneously at each of a sequence of time intervals (each environmental variable having 4 16-bit 'genes' to represent its value and some other properties). The short-term behavior (time sample to time sample) of each critter depends on the configuration of those values at any given sample time. Long-term evolution in the population of critters depends on the adaptive (or non-adaptive) character of their sample-to-sample behavior through generations, which determines differential probabilities of reproduction across critters.
RBH
[ 10. February 2003, 11:01: Message edited by: RBH ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 10. February 2003 12:08
I think that some of the concepts that Warren discusses relate to the “meme” concept by Richard Dawkins. (See Richard Dawkins: Biography and Background for links to his materials.)
Some notes on “memes” on that page:
quote:
Memes are to cultural inheritance what genes are to biological heredity. A meme for, say, astrology, could parasitize a mind just as surely as a hookworm could infest someone's bowels. Ideas - like genes - could compete and cooperate, mutate and conserve. They, too, are operated on by natural selection. Human evolution, Dawkins postulates, is a function of a co-evolution between genes and memes.
and
quote:
This was apostasy to Darwinian evolutionists, who took it as dogma that the dynamics of natural selection cared only for the fitness of individual organisms and absolutely nothing else. But here was Dawkins saying that what really counted in "nature tooth and claw" was the replicating code beneath the organism. Evolution is really the story of replicators Ÿber alles.
Dawkins aggressively evolved this replicator concept. He noted that discussing the evolution of birds without looking hard at the evolution of their nests, or at beavers without considering the evolution of their dams would be prima facie ridiculous. Each is essential to the survival of the other. It is the combination of bird and nest, the combination of beaver and dam, that gives a competitive edge to the animals who build them. Not only does the body of an organism march to the orders of its genes, but so do the artifacts the organism builds or uses. In this sense, the egg uses both a chicken and a nest to make another egg, and so the nest, too, is an evolutionary extension of the egg.
Now what I think is relevant here is that the details of the individual short-term activities of organisms has not been ignored in the study of evolution. This is just an example from a well known scientist.
Warren may be wanting to model such detailed activity, and makes claims that such details are needed to more fully understand evolutionary progress. I don’t disagree with this (though this is outside of my area of expertise, and I make no comment on the degree of acceptance of Dawkins’ concepts for example.)
The problem is that Warren doesn’t just claim that such detailed modeling will be useful. He claims certain conclusions from the speculation on what such modeling would produce. But without doing the details, and presenting the results, and showing how the details are consistent with observation, we have nothing that can be judged, and no way to examine Warren’s conclusions. (Except that we can see that certain aspects may already be in conflict with observation in some instances, and that Warren’s statements about results from other biological studies may be in error -- once again I leave that to experts in those fields.)
So Warren may have an interesting idea. But without detail, we can’t know of its value, and we can know that his arguments against standing science are without any basis of detail. Simple claims that more detailed modeling would result in greater degree of predictivity are not new concepts nor are they arguments of error in the evolutionary models that abstract such details.
To an extent Warren is making claims that the results of this speculative possibility of modeling can be known. This is where Warren’s points appear to falter, because they are based on claims about the details of present science that don’t seem to hold up to thorough examination.
I suggest that Warren should stop making claims without sufficient detailed support, and start working toward details of his concept so that it can be developed and analyzed. The concept of dealing with short-term details of the life cycles of individuals in populations is not foreign to evolutionary modeling. The problems of the great amount of detail to be integrated into such a model makes the problem daunting -- and thus very unlikely to have well-characterized mathematical results from simplistic claims.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 10. February 2003 15:00
Gedanken,
Quote: Such modeling would, like the modeling of pressure or weather, possibly give greater detail in understanding, but would similarly be expected to fail to provide long term accurate predictions of behavior due to supersensitivity. (And indeed even quantum indeterminacy again would make these predictions inconclusive in principle, not simply in practical terms.)
As a starting point, it is important to note that there is no a priori level of detail that is most appropriate or most productive. Some types of analysis are most effective at one level and other forms of analysis are more effective at other levels. Some naïve interpretations of the ‘laws of nature’ suggested that forces such as gravity which operate on planets should operate in exactly the same ‘lawful’ manner down to the level of the mathematical point. As I understand it, the study of very small particles has demonstrated that the naïve interpretation of laws of nature was wrong. There is no scientific principle that says that predictions must be accurate at one level and inaccurate at another.
Second, it must again be repeated that ‘whole organisms level’ models based on Darwinian and neo-Darwinian concepts do not produce reliable predictive models. It must also be pointed out the descriptive models of developmental processes have not led to the development of predictive ‘causal models’.
It is recognized that the transformation from genes to phenotypes or from egg cell to fully functioning organism can be expressed as a developmental tree with billions of branches/ developmental pathways involving very large numbers of ‘survival essential phenotypes’. As has been discussed, based on known processes it is at least in theory possible to represent these developmental pathways or transformations by complete or partial sets of ‘automated assembly instructions’.
There is no a priori reason that it should or should not be possible to develop ‘predictive scientific theories’ at this level. It is, however, worth repeating that to date no one has succeeded in developing such theories at any other level.
Quote: Asking for the detail does not constitute a rejection. But we may fairly reject that you have any such concept -- and certainly that we should accept any particular conclusions that you may wish to speak of, if you do not actually present details that we can observe.
If you are asking for details then I guess we have progressed from calling standard reductionist techniques speculative. I only ask that in requesting details you 1)allow me to present those details in my format, that 2)you be willing to consider concepts and techniques from other fields, particularly applied science fields, and 3)if you disagree with an approach or technique be willing and able to provide the logic justifying the alternative you support.
I provide my opinion or prediction on the conclusions of certain types of analysis because, IMO, 1)once you understand the techniques being used, the conclusions offered appear obvious, and 2)the conclusions listed provide the reader with ‘hints’ on the direction the discussion is taking.
From what I know of it, Dawkin’s concept of ‘meme’ has little or no relevance to what is being discussed here. My approach asserts that all biological information processing, and thus the assembly, operation, and evolution of biological systems, whether in bacteria or in Shakespeare- can be modeled by sets of ‘dynamic and teleological causal relationships’. An assembly instruction is an example of such a causal relationship.
Quote: The problems of the great amount of detail to be integrated into such a model makes the problem daunting --
But not impossible. Models today can address vast volumes of processes. It is also important to note that you can analyze some features of the behavior of very complex systems without actually being able to build such a system. This type of analysis is the essence of engineering analysis. One can analyze in great detail the voyage to another solar systems despite not being able to actually building and launching such as space vehicle.
RBH,
I listed four properties or features of ‘variation-selection’ processes. I you would like to see an illustration, please select which feature you would like to see illustrated first.
Quote: It is not a departure from my "concept" of GA models. In the GAs my company builds, the artificial critters comprising the population each sample up to 16 environmental variables
This would suggest that the difference between GA’s and TA’s is a matter of degree. In a TA it is, in theory, possible to model or simulate any ‘evolutionary’ change or any ‘search for an adaptive solution’ within the lifetime of a single critter. [A completely different subject, but my approach suggests that the function in nature of the ‘high levels of death and reproduction’ is ‘preventing evolutionary divergence’ and this death and reproduction plays a rather trivial role in evolutionary/adaptive change. ] Once you incorporate the concept of multiple ‘variation-selection processes’ and multiple ‘selection variables, then you can model or simulate adaptive or evolutionary change in a single critter. Right?
[ 10. February 2003, 15:39: Message edited by: warren_bergerson ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 10. February 2003 16:20
Warren, let’s examine another case -- that of modeling river erosion and the development of the complex boundary of a river or stream.
This could possibly benefit from a much greater deal of detail dynamic modeling of the individual actions of the erosion process. One could, in principle, model the actions of each grain of sand or dirt in the river bank. One could model the striking of rocks against the bank and against each other in a dynamic and time varying manner, discovering the details of each rock impact that causes a fracture and division. One could examine the detailed surface characteristics at a moment in time of the river bank, and the contents of the river flow, and dynamically model the flow turbulence in great detail, even including the masses of particles that change the dynamics of the water’s uniformity, and modeling the water flow in a grid of sufficient detail to accurately follow the turbulence effects.
However we have problems with this approach. Turbulence is not well understood, and simulations of the fine details to some degree just show continued complexity and do not give larger principles that can be easily understood or extracted. It is well known that such modeling as is done in weather and climate shows the effects of supersensitivity makes predictions only accurate short term, and only accurate when a tremendous amount of simultaneous state information was gathered highly accurately. So even with that high degree of state information, long term predictions do not ensue.
Now what is very important here is that there is nothing new in the essential components of the river flow and erosion modeling problem as described. Nor is there anything new in detailed knowledge of physics in the detailed weather modeling, what is new is methods of approximations rather than of working with the extremely refined detail.
Now there is another approach in the river modeling problem. One could deal in ensembles of such details, rather than with the individual details. One could learn and test for aggregate (e.g. abstracted or “reductionist”) pattern descriptions which produce a degree of predictivity but not highly accurate long term forecasts.
A very important question is whether such detailed high information models (which diverge quickly with supersensitivity) actually provide greater understanding, and to what degree. Specifically, do they invalidate the lower information models, or do they simply confirm those models? I assert that they for the most part have been found to confirm the lower information models as correct -- to the degree of predictivity that is expected from such lower information models.
Remember that erosion models will state that there are bounds of erosion that will most likely be obeyed -- a relationship is specified, not an exact prediction. Also note that evolution models predict that a relationship will be held to over time, and also do not provide a precise prediction of what will happen at a given future time. Similarly for weather modeling, wherein there is a greater detail of short-term accuracy in forecast, but this is only a matter of degree and is still a matter of the prediction giving a relationship that is expected to be obeyed (the prediction of the model is that the future will fall in the bounds of the relationships described in the theory.)
I see exactly the same sort of problem in what appears to be Warren’s suggestion. Warren talks about how all this detail could be analyzed for details of information -- just as the detail of information in the river flow or weather modeling system could be used. I don’t (once again) see sufficient detail being presented yet to make any analysis of Warren’s “model”.
quote:
From what I know of it, Dawkin’s concept of ‘meme’ has little or no relevance to what is being discussed here. My approach asserts that all biological information processing, and thus the assembly, operation, and evolution of biological systems, whether in bacteria or in Shakespeare- can be modeled by sets of ‘dynamic and teleological causal relationships’. An assembly instruction is an example of such a causal relationship.
Where here is the “teleological” model any different from the details already understood by biologists? Claiming that they can be assembled into a grand model rather than dealt with reductionistically as separate aspects of evolution does not change the importance of the abstractions dealt with by biologists, nor their accuracy. I see Warren discussing stimulus-response maps -- certainly nothing strange to a biology analysis, so where is the difference? Is the only difference the aggregation of these details into a grand overall modeling process? If so, where are the details? I only see Warren producing his own different sort of simplified abstraction, without support from observation of nature and without detail.
quote:
… Models today can address vast volumes of processes. It is also important to note that you can analyze some features of the behavior of very complex systems without actually being able to build such a system. This type of analysis is the essence of engineering analysis. One can analyze in great detail the voyage to another solar systems despite not being able to actually building and launching such as space vehicle.
A model of a single flight to a distant location is highly different from the model of the tremendous complexity of ensembles of complex systems. That one could model a single system in some degree of accuracy is not a statement about aggregate assemblies of tremendous numbers of interacting systems. This is what we learn from the weather modeling case, that the complexity problems are very significant, and can’t be compared to modeling a single small component. And weather is extremely simple as compared to modeling progress of evolution of extremely large numbers of interacting individuals in a dynamic manner.
quote:
Quote: I suggest that Warren should stop making claims without sufficient detailed support, and start working toward details
I assume you would also like to see that suggestion applied to supporters of Darinian and neo-Darwinian concepts.
But as I pointed out above, the detailed “high information” models that have been achieved in some domains (like weather) are not known for overturning the more statistical or “low information” models -- rather they simply refine them in some circumstances to the degree that supersensitivity does not limit such refinement. In no way is there a history of detailed models invalidating such reductionistic or abstractionist models.
From early in post:
quote:
As I understand it, the study of very small particles has demonstrated that the naïve interpretation of laws of nature was wrong. There is no scientific principle that says that predictions must be accurate at one level and inaccurate at another.
No, the study of very small particles has demonstrated that the naïve interpretation of laws of nature should not be expected to apply to all domains. In fact “naïve” interpretations of laws of nature can always cause problems because there is a great deal of learning and expertise in applying them consistently and with understanding. Newtonian physics is not “wrong”, it simply does not apply in all domains.
Likewise Darwinian concepts are not “wrong” when we learn of details that refine the theory. There needs to be correspondence to the observations that show that Darwinian relationships are virtually always consistent with observation in the larger longer term picture. (An aggregate of a large number of detailed quantum events still obeys or corresponds to a Newtonian result in the common domains in which it is known to be accurate.)
I see no need for any retraction of claims that Darwinian models have been very reliable in conforming to observation. However I have seen no indication that “Warrinian” models have even been compared with observation of nature. Sure, take your time and present your ideas as you see fit. My objection is not to your claim that you could potentially explain some more advanced model. My objection is to claims about well supported science without giving adequate demonstration of errors in that well supported science. I don’t need to be an expert to see the lack of support in such claims. And as my “alternative” I propose the existing well-supported science.
It is the responsibility of the person making a new claim to support that claim, it not my responsibility to come up with alternatives that would work at the same level of detail. I have no knowledge whether such models at that level of detail would be fruitful. I know that to be fruitful, they need to be presented in sufficient detail to be analyzed and understood.
--
[Edit added:]
quote:
Once you incorporate the concept of multiple ‘variation-selection processes’ and multiple ‘selection variables, then you can model or simulate adaptive or evolutionary change in a single critter. Right?
Not to “pre”-disagree with RBH in the next post (and sorry for flow problems here but I can’t really add another post today):
Would not Dawkins’ “memes” be an example of something that is incorporated in an individual organism. And it would be one that would have a “within-lifetime” adaptive effect (being shared by learning within the lifetimes of the organisms)? And possibly have a character similar to evolution that if the “meme” gave apparent advantage, the organism would itself tend to keep the behavior but if not advantageous than shedding the behavior in a vaguely evolutionary processes of behavioral feedback (e.g. learning)?
But of course Warren says that Dawkins’ memes don’t have anything to do with what he is talking about so I guess not. But of course if it does, it shows that Warrens’ claim that scientists do not try to account for such details in evolutionary models are in error. In either case there is no invalidating of basic principles of Darwinian models even in their being refined.
[ 10. February 2003, 18:57: Message edited by: gedanken ]
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 10. February 2003 17:15
warren wrote quote:
I listed four properties or features of 'variation-selection' processes. I you would like to see an illustration, please select which feature you would like to see illustrated first.
Since the goal of giving examples is to clarify similarities and differences, how about one example of each?
Quoting from me, warren wrote quote:
quote:
It is not a departure from my "concept" of GA models. In the GAs my company builds, the artificial critters comprising the population each sample up to 16 environmental variables
(truncated as in warren's post)
This would suggest that the difference between GA's and TA's is a matter of degree. In a TA it is, in theory, possible to model or simulate any 'evolutionary' change or any 'search for an adaptive solution' within the lifetime of a single critter.
...
Once you incorporate the concept of multiple 'variation-selection processes' and multiple 'selection variables, then you can model or simulate adaptive or evolutionary change in a single critter. Right?
What degree? In principle, our GAs can be expanded to any number of input variables subject only to the limits of computer memory, CPU cycles, and my patience.
When you say TA is designed to "model or simulate adaptive or evolutionary change in a single critter" I lose you. "Adaptive" and "evolutionary" are not synonyms. I don't know what "evolutionary" change in a single critter means. To be sure, single critters display adaptation during their lifetimes, in biology at least, but we call the various kinds of adaptation during critters' lifetimes "learning," "habituation," "sensitization," and the like, not "evolution." Some of the within-critter adaptive capabilities are themselves evolved over generations, but not during the lifetime of a single critter. "Evolution" is generally reserved to denote changes in populations over generations. As a pure pedagogical note, if you use a well-established term in a novel and nonstandard way, people are going to misunderstand you. Witness Dembski's use of "complex" to mean "improbable." A good deal of confusion has been engendered by the connotational load "complex" carries that does not appropriately apply to "improbable." Similarly, to use "evolution" in this novel way is already confusing, and it will become more so as you elaborate your model.
RBH
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 10. February 2003 21:10
quote:
TA systems involve multiple variation-selection processes occurring simultaneously. This is possible because TA systems involve multiple selection variables being measured or recorded at the same time. Is this or is this not a departure from your concept of GA models?
Multiple variation-selection processes are equivalent to a single more complex variation-selection process, if you are going to end up with an answer at the end. For ease of implementation, GAs typically use a single process.
There are departures elsewhere, however.
quote:
A TA model assumes that the values the recorded environmental values reflect the values at the time the values are recorded. The value V1 for option 1 can be very different at time t=1 than at t=0. Using TA models, adaptive solutions can change rapidly.
This seems to indicate that instead of building a system that responds to changing conditions on its own, the TA actually changes the system along the way--potentially using some high-level manipulations not avavailable to the system itself.
This is definitely different from genetic algorithms, because the timescale of selection is the generation, which could be viewed as some sort of cumulative product of survival values over the lifespan of each organism. Any variation from this--for instance, by modifying the organism in response to the environment in a way that is under the control of the algorithm rather than the organism--is a deviation from a standard definition of a genetic or evolutionary algorithm.
quote:
A third important feature of the TA definition of selection is that ?variation-selection processes can operate on variation-selection processes?. Consider the pseudo code defined above. A variation- selection operation could operate to change the method of recording environmental values or it could operate to change the logic used in selection.
Fine--but you still haven't given us an example of logic used in selection! The closest you have come is "based on a comparison of values V1 and V2". That is too general to be of any use.
And I hope that the programs don't get to define their own survival value. Otherwise, you just set that to "perfect", and quit, and have an utterly useless "perfect survivor".
Without actually seeing a much more concrete example, I also can't tell if your method is really a form of exhaustive search. This is a very important point, because GAs and simulated annealing and other branches of optimization/maximization theory exist precisely because exhaustive search is too expensive for many problems. You cannot exhaustively search a very large space with any degree of success. Instead, iterative methods are necessary, assuming that the space is locally nonrandom, so that if you start out in a "good spot", nearby points in the space will also be "good".
So far, I do not see how in detail TAs engage in this iterative process. They seem to be trying out giant lists of possibilities, which are generated in a method that you still haven't fully explained, and are compared by a method that you haven't explained, that causes some outcome that you haven't fully explained. You have plenty of overviews--just not enough details to work with.
So, to try to get more specifics:
quote:
Survival values need not, however, be a variable or set of values. We can define, identify, large hierarchies of indirect survival variables.
Please give an example! Preferably one that does not fit within the normal GA/Darwinian framework!
I'll give an example: a coevolving system of wolves and rabbits. Wolves have a speed, bite strength, keenness of sight, and brood size, each of which takes energy. Rabbits have speed, toughness, camouflage, and brood size, each of which takes energy.
We have a scoring function that looks something like 1/(1+e^(x-y)) for the probability that a rabbit with quality x will be eaten on encountering a wolf with corresponding quality y; then the survival value of rabbit r_i is the probability that it escapes each wolf, times its brood size.
This type of scheme wouldn't be terribly uncommon for certain types of genetic algorithms used to study predator-prey relationships and tradeoffs. Do you envision a TA doing anything radically different? Please be specific.
After survival, the progeny of rabbit r_i would be initialized with speed, toughness, camouflage, and brood size values that were clustered about those of its parent. (We'll assume that rabbits reproduce by asexual budding for this example....) Do you envision a TA doing anything radically different? Please be specific.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 11. February 2003 10:47
Rex,
Quote: Multiple variation-selection processes are equivalent to a single more complex variation-selection process, if you are going to end up with an answer at the end. For ease of implementation, GAs typically use a single process.
This is true, but note that the complexity of the combined process increases exponentially. If you have 5 variation-selection processes each with 10 possible options for each adaptive option, then finding the correct options involves testing 5*10 or 50 options. If you combine the options into a single variation-selection process you end up having to evaluate 10^5 or 100,000options.
To illustrate, consider a combination lock with a 5 codes each with 10 options. If you can search for the solution one code at a time it will take a maximum of 50 trials (an average of 25 trials) to break the code. If you have to solve for the aggregate code you need a maximum of 100,000 trials (an average of 50,000 trials) to find a solution.
The relevance of multiple variation-selection processes to assembly instructions should not be difficult to visualize. When the transformation from A to B is represented by a set of assembly instructions C containing instructions c1,c2, …,cn, then a TA with n variation-selection operations can model finding or evolving the A to B transformation by finding or evolving each of the n assembly instructions.
Since complex assemblies could involve well in excess of billions of assembly instructions, it may not be practical to create a full scale simulation of the A to B transformation. We would instead start by analyzing in detail some assembly instruction cx and the processes responsible for forming and transforming cx. Based on what was learned from the analysis of a small number of assembly instructions, it is suggested or predicted, we can formulate ‘theories’ which attempt to explain how the overall process works.
A TA is a mathematical model, a mathematical abstraction. Like engineering models, a model of a bridge for example, a TA is primarily designed to test how detail knowledge of components fits in or is logically consistent with an overall form.
Quote WB: A TA model assumes that the values the recorded environmental values reflect the values at the time the values are recorded. The value V1 for option 1 can be very different at time t=1 than at t=0. Using TA models, adaptive solutions can change rapidly.
Quote R: This seems to indicate that instead of building a system that responds to changing conditions on its own, the TA actually changes the system along the way--potentially using some high-level manipulations not avavailable to the system itself.
The conclusion is partially correct. TA systems change along the way. They can evolve or grow. But I don’t quite see how you get that from my statement to your conclusion. The mechanism I described is fairly simple and mechanical
Consider as simple example a system with 1 assembly instruction (s, r) involving a single stimulus or trigger s and two possible responses r1 and r2. (we could consider systems with multiple instructions, multiple stimuli and multiple responses). The system has an ‘input device’ that records some environmental condition which is ‘selection criteria associated with the variation-selection process controlling our assembly instruction. As defined yesterday(with a modification), our selection program has the form :
CODE FOR A SELECTION PROCESS
1. LN1: record environmental value V1 for options 1
2. record environmental value V2 for option 2
select option 1 or option 2 (and reject the other option) based on a comparison of values V1 and V2
3. GO TO LN1
If assembly instruction (s, r1) is option 1 and (s, r2) is option 2, then assume that a time t=0 the recorded environmental value or survival value for option 1 is 5 and the value for option 2 is 8. If the selection logic picks the option with the higher selection or survival value then at time t=0 the assembly instruction (s, r2) will be selected.
The selection process as defined operates continuously. If at time t=3 the survival value of (s, r1) is recorded as 7 and the survival value of (s, r2) is 4, then the selection logic will change the assembly instruction from (s, r2) to (s, r1).
It is not difficult to design assembly process programs with multiple levels of selection processes operating on selection processes … operating on an assembly instruction. Take a simple example of an oven with a timer. The initial instruction is ‘if temp less then X turn on heating element’ . A higher level instruction says turn on ‘if time greater than Y modify initial instruction.
One of the practical problems with analyzing complex biological systems is complex structure of operations operating on operations which operate on operations etc. There are effective mathematical techniques for dealing with these complex logical structures.
Quote: This is definitely different from genetic algorithms, because the timescale of selection is the generation, which could be viewed as some sort of cumulative product of survival values over the lifespan of each organism. Any variation from this--for instance, by modifying the organism in response to the environment in a way that is under the control of the algorithm rather than the organism--is a deviation from a standard definition of a genetic or evolutionary algorithm.
This was always my ‘impression of GA’s’ but many people seem reluctant to admit the existence of this constraint. Note that a TA system can model both ‘between generations’ variation-selection processes and within generation variation-selection processes. TA systems can thus be used to test hypothesis which incorporate some form of the ‘only between generations variation-selection constraint’.
Note that GA’s are used in attempts to simulate human problem solving. Since human problem solving involves ‘within lifetime variation-selection’, GA’s are not always be limited to between generation variation-selection.
Quote: Fine--but you still haven't given us an example of logic used in selection! The closest you have come is "based on a comparison of values V1 and V2". That is too general to be of any use. And I hope that the programs don't get to define their own survival value. Otherwise, you just set that to "perfect", and quit, and have an utterly useless "perfect survivor".
In an ‘idealized’, perfectly efficient abstract universe, a system capable of evaluating essentially all possible adaptive solutions would, it might be speculated, evolve the ultimate perfect survivor. Clearly in the real world, evolutionary/adaptive processes, while extremely powerful, are not completely effective/efficient and individuals, species and groups of species on occasion fail. The concept of the ‘idealized perfectly efficient process or machine’ is widely used in scientific analysis. The idealized perfect survivor is not a useless scientific concept.
Analysis of ‘assembly instructions’ and ‘processes creating and modifying assembly instructions’ begins with generalized ‘mathematical concepts of assembly instructions as ‘dynamic and teleological causal relationships’ and generalized ‘variation-selection processes’ to maintain the teleological or adaptive form of the assembly instruction. The ‘generalized’ model is then fit to a set of observed data.
Given a model that fits a set of assembly process data, we then form hypotheses designed to explain the observed change process. By making additional observations, we test the validity of the proposed theory and/or identify additional criteria and or constraints which the model/theory must satisfy.
One of the basic concepts underlying the proposed experimental paradigm is to start with a generalized mathematical model capable, at least in theory, of 1)modeling and simulating the assembly and operation of any biological system and 2)modeling and simulating the evolution/development of any set of assembly or operating instructions. As knowledge is accumulated and reflected in the model, the generalized model becomes progressively more specific.
Quote: You cannot exhaustively search a very large space with any degree of success. Instead, iterative methods are necessary, assuming that the space is locally nonrandom, so that if you start out in a "good spot", nearby points in the space will also be "good".
This, IMO, has always been a key finding from GA models. A basic or elementary random variation-selection based on survival process can not effectively or efficiently search a large space. Since biological systems are very complex or improbable systems, (small solution areas in very large solution spaces) it is obvious that the development or evolution of these processes can not be the product of simple RM&NS processes. GA’s provide convincing evidence/proof that evolutionary change could not be the product of a basic RM&NS process.
But if evolution is not the product of a ‘whole organism random variation-selection process’ what is the alternative. I am suggesting that since between generation variation- selection does not by itself seem to explain evolutionary change we should look at the role of within lifetime variation-selection processes. To do this we start by defining developmental and operating processes in terms of sets of ‘automated assembly or operating instructions’. We then note that we can mathematically model the change or evolution of any assembly instruction with a variation-selection process. If an biological assembly or operation involves n automated instructions or S-R relationships, then the evolution of that systems can be modeled by a set of n variation-selection operations. This is the TA system.
Given a generalized model, the next step is to analyze the processes by which actual assembly instructions change or evolve overtime. If you can accept that ‘changes in any assembly instruction can be modeled by a variation-selection process’ then we are ready to start considering how ‘actual evolutionary/adaptive change processes’ are similar to and/or differ from the basic variation-selection model.
You had additional good questions, but this is more than enough for one day.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 11. February 2003 11:34
Gedanken- It is useful to note that you entered this discussion claiming, without any supporting evidence, that ‘treating ‘transformation from A to B’ as ‘as set C of automated assembly instructions’ was speculative. I claimed and continue to claim that this type of transformation is a standard, legitimate analytical technique. The conclusions that may or may not eventually be reached starting with this technique may unknown, but the technique is not speculative.
RBH,
Time permitting, I will try to get to the examples tomorrow. Just a couple of quick comments on some other issues.
Quote: What degree? In principle, our GAs can be expanded to any number of input variables subject only to the limits of computer memory, CPU cycles, and my patience.
It is useful to distinguish between 1)mathematical systems or logic machines and 2)computer simulations or manifestations. Logic machines do not necessarily have limited memory or CPU cycles. The computer manifestation of a GA may be limited in what it can model or simulate. Specifically, it is generally recognized that computer manifestations of GA’s can only model simplified expressions of evolutionary change. In most applications a TA is an abstract logic machine with memory and CPU requirements far beyond those of existing computers. There are obviously many analytical uses for both abstract logic machines and actual computer manifestations.
Quote: Some of the within-critter adaptive capabilities are themselves evolved over generations, but not during the lifetime of a single critter. "Evolution" is generally reserved to denote changes in populations over generations. As a pure pedagogical note, if you use a well-established term in a novel and nonstandard way, people are going to misunderstand you. Witness Dembski's use of "complex" to mean "improbable." A good deal of confusion has been engendered by the connotational load "complex" carries that does not appropriately apply to "improbable."
I am in complete agreement that the use of terminology is an important point, but I do not fully agree with your conclusion. An individual or group of individuals like biologists does not own a scientific term. If common usage of a term is inconsistent, imprecise, inappropriate or not useful, then alternative definitions can be proposed. Evolution change processes, for example, refers to ‘the processes generating measurable heritable changes in groups of organisms’. There is nothing inappropriate is modify the definition to include ‘processes responsible for heritable changes from set C1 or assembly instructions to set C2 of assembly instructions’. There is nothing inappropriate in introducing definitions of ‘biological information/complexity’ and ‘biological information processing’ that make it possible to quantify both complexity and information generating capacity.
Introducing a new or refined definition may for some be source of confusion, but it can also be a way to eliminate ambiguity and inconsistencies that may exist with ‘generally accepted’ definitions and interpretations.
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 11. February 2003 13:42
Sorry for the quote-response format, but I think that it is important to point out discussion here in context of what was said in this instance.
Warren’s comments to Rex:
quote:
This is true, but note that the complexity of the combined process increases exponentially. If you have 5 variation-selection processes each with 10 possible options for each adaptive option, then finding the correct options involves testing 5*10 or 50 options. If you combine the options into a single variation-selection process you end up having to evaluate 10^5 or 100,000options.
This view of a “search” for a particular code will become relevant in a moment. As has been explained by RBH and others multiple times, and in multiple thread, evolution is not a “search” for a particular result. One can model a vaguely “search” like behavior in that the increased fitness is given priority in evolutionary models, and in that vague form the model is a “search” for better fitness. But it is not a “search” for a specific code. One must include in issues of finding adaptive options how many different codes would produce adaptive options, and just looking at the improbability of the particular pathway taken is not particularly of interest. The particular pathway taken in a sequence of options for a system will always be an exponentially decreasing probability, no matter whether the operation of choice of option had anything to do with fitness or not.
quote:
… Based on what was learned from the analysis of a small number of assembly instructions, it is suggested or predicted, we can formulate ‘theories’ which attempt to explain how the overall process works.
Here we must ask what is the difference between this approach and what is already being done in biological science.
quote:
A TA is a mathematical model, a mathematical abstraction. Like engineering models, a model of a bridge for example, a TA is primarily designed to test how detail knowledge of components fits in or is logically consistent with an overall form.
I am confused here. If one is constructing a “scientific” model, one produces a model that has some predictive value, some predictive outcome at some level. If it has “explanatory power” in terms of understanding of processes, then it can be valuable. We would test that the predictions are in conformance with observation, and we would do that in many different ways to see that it is robust in its conformance with observation. Now an “engineering model” can have a scientific aspect of similarly needing to be in conformance with observation, but also is predictive in the sense of providing useful feedback in a design process -- but its “explanatory power” in terms of a model of now nature works becomes less important and is not normally evaluated. So Warren, if you are proposing that we are trying to discover how nature works, I suggest that we need to concentrate on the explanatory power of the model, and the degree to which it conforms to observation in a robust manner.
quote:
If assembly instruction (s, r1) is option 1 and (s, r2) is option 2, then assume that a time t=0 the recorded environmental value or survival value for option 1 is 5 and the value for option 2 is 8. If the selection logic picks the option with the higher selection or survival value then at time t=0 the assembly instruction (s, r2) will be selected.
The selection process as defined operates continuously. If at time t=3 the survival value of (s, r1) is recorded as 7 and the survival value of (s, r2) is 4, then the selection logic will change the assembly instruction from (s, r2) to (s, r1).
Now I’m really confused.
How can the “survival value” of any aspect be tested? Remember that Warren said “The mechanism I described is fairly simple and mechanical”. “Survival” is tested when one continues to survive, or fails to survive. Actually the reproduction value is the more important to evolution, but that is closely related to “survival” as one must survive past the point of reproduction. What is the basis for determining that “survival value” has different numerical values at these different times? What the GA in more standard form would do, even if it had dynamic time-varying subsystem simulation, is to record whether the individual reproduced or didn’t (along with the individual’s life-process simulation). But the information that relates to the GA’s information of “what works” comes from the reproduction. How does this “TA” do this in any kind of “mechanical” manner, and how can it evaluate survival value dynamically in the way described?
quote:
Note that GA’s are used in attempts to simulate human problem solving. Since human problem solving involves ‘within lifetime variation-selection’, GA’s are not always be limited to between generation variation-selection.
I’m Extremely confused! That GA’s can be used for multiple purposes does not seem to be relevant to how GA’s are used to simulate evolution, or a comparison of GA as evolution simulation to Warren’s “TA”. And if “within lifetime” information transfer mechanisms are to be simulated, any model of learning would be useful and no reason exists to restrict to the GA.
I think that the rest of the post illustrates the confusion about the evolutionary model and how overall life-long “selection” processes increase information.
I fully accept that within-lifetime processes can be modeled, and that information can be carried by other mechanisms from generation to generation by mechanisms other than the genetic information. My example of Dawkins’ “memes” is directly relevant to learning being carried forward, and is an example of scientists already doing precisely that. And processes of development, such as is studied in the area of cloning, are examples of a large number of scientific areas of study of how the biological development processes work. I think that the physical impacts of the environment on the individual organism have most certainly not escaped the notice of biologists -- what is ecology but the study of the relationships of individual environmental effects to population interaction effects?
Warren has not answered Rex Kerr’s question of how the “TA” differs from the GA, in the following senses:
No information has been given on details of how non-genetic information is carried forward. Only vague stimulus response claims have been made. These are speculative statements that Warren can make a meaningful model based on this concept, as none has been shown yet.
The question of how the time series model of the individual simply results in an overall “reproduction” event(s) has not really been addressed. So if a “TA” model models individual lifetime events, and survival and reproduction is an accumulation of those events, where is the TA any more than a detailed simulation of the individual embedded in a GA? Where is any demonstration that the GA’s overall result has been in error in its predictions (using a GA specifically as a model of evolution, and speaking of that specific use of GAs as models of evolution and not of other uses of the GA)?
quote:
This, IMO, has always been a key finding from GA models. A basic or elementary random variation-selection based on survival process can not effectively or efficiently search a large space. Since biological systems are very complex or improbable systems, (small solution areas in very large solution spaces) it is obvious that the development or evolution of these processes can not be the product of simple RM&NS processes. GA’s provide convincing evidence/proof that evolutionary change could not be the product of a basic RM&NS process.
Here you go again! Where is the support for this claim? Here is the type of unsupported claim that I am repeatedly objecting to. First, as RBH and others have pointed out, evolution is not a “search” for the extremely improbable result, rather it is a process that steps from state to state with relatively high probability events. The “search space” of all genomes is not “searched.” We know by observation that there is a high density different genomes producing different organisms that are relatively well adapted to their environments, and there is no presentation given that the density of working organisms in the “space” of possible genomes is not high in relative terms (small in fraction of all possible genomes, yet still a very highly rich space of possibilities.)
quote:
Gedanken- It is useful to note that you entered this discussion claiming, without any supporting evidence, that ‘treating ‘transformation from A to B’ as ‘as set C of automated assembly instructions’ was speculative. I claimed and continue to claim that this type of transformation is a standard, legitimate analytical technique. The conclusions that may or may not eventually be reached starting with this technique may unknown, but the technique is not speculative.
Warren, you misrepresent my statements. I can see how you might want to claim that I was saying that an individual component of what you propose to propose is my focus. I entered this discussion saying that you had not presented mathematical detail so that your proposal can be evaluated. Here, if what you bring up is to be taken as “mathematical detail”, I still stick to that statement. The example of the evaluation of the “survival value” is illustrative. No detail has been presented, and even the terminology is inconsistent with the normal use of such terms. So yes, what you are presenting is still “speculative”! You can’t get out of the speculative nature, that it is specculation that your presentation will eventually contain sufficient detail for analysis, by suggesting that generic mathematical techniques might be valid and therefor what you intend to present or have presented is meaningful.
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 12. February 2003 05:46
quote:
To illustrate, consider a combination lock with a 5 codes each with 10 options. If you can search for the solution one code at a time it will take a maximum of 50 trials (an average of 25 trials) to break the code. If you have to solve for the aggregate code you need a maximum of 100,000 trials (an average of 50,000 trials) to find a solution.
No, no, no. This isn't anything like what GAs do.
GAs do not implement brute force search. Problems which are only amenable to brute force search are much faster solved by brute force than by GAs--finding the combination to a lock is one example. Everything has fitness zero, except the correct combination which has fitness 1.
GAs are good at searching fitness landscapes where there are fairly smooth paths to local maximums. Read Climbing Mount Improbable by Dawkins, or any textbook on genetic algorithms, or pretty much any textbook on evolution for more details.
quote:
This, IMO, has always been a key finding from GA models. A basic or elementary random variation-selection based on survival process can not effectively or efficiently search a large space.
Except that this is not the key finding. There are certain types of incredibly vast spaces that are searched amazingly effectively using random variation-selection schemes.
I'm sorry, but I simply don't have the time to teach an online course in the architecture, capabilities, and limitations of genetic algorithms and evolution-simulation systems. I especially don't have time to do it while simultaneously having to ask for every tiny addition of detail to your named-but-underspecified TA system.
I suggest that if you are serious about developing your ideas, that you first write a paper on the architecture of TAs, including some examples. Anything less than ten pages will probably be so brief as to be useless; I'd aim for about twenty, and I'd include actual working code (not pseudocode) for a toy example. Then get a book or two on genetic algorithms and refine your presentation and the capabilities of your ssytem based on that. (I'd suggest the other way around, but you're less likely to make novel insights after already being exposed to too many details of another system--it encourages focusing on too small of problems.)
With that, I am going to mostly retire from this discussion. If I am going to spend so much time helping to develop a novel high-dimensional optimization scheme, I'd rather spend it working on my own ideas (which seem to me more concrete and promising than TAs, albeit without evolutionary relevance).
I will, however, comment on any substantive, detailed additions to the description of teleological algorithms.
Edited to complete a truncated sentence.
[ 12. February 2003, 05:49: Message edited by: Rex Kerr ]
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 12. February 2003 11:55
Rex,
Quote: Except that this is not the key finding. There are certain types of incredibly vast spaces that are searched amazingly effectively using random variation-selection schemes.
The old miracle of chance argument. A search technique that routinely finds a million to one chance in a single trial is not based on random variation. Far too many people treat GA’s as magic machines rather than mathematical entities. Random variation, in mathematics, means that each option in the selection space has an equal probability of occurring. The speed or expected speed of a ‘whole system-random variation-consider one alternative at a time’ process is .5 times the ratio of the size of the solution space to the size of the set of adaptive solutions. As was stated early on, this is the ‘base’ case used here to quantify the volume of information processing needed to find an adaptive solution.
There are, obviously, many search techniques which, under appropriate conditions, can find adaptive solutions much, much faster, with far fewer cycles or trials than a whole system random search. Common examples of ‘fast’ variation-selection processes would include ‘stored solution techniques’ and the ‘multiple independent variation-selection’ techniques described for the five code lock. These high-speed searches are not random variation approaches.
It is important to note that ‘fast search’ is not necessarily the same as ‘efficient search’. A fast search is ‘efficient’ only if it actually finds an adaptive solution or a solution compatible with survival. You can speed up a search process by 1)increasing the number of cycles per unit of time, 2)increasing the number of alternatives evaluated per cycle, and/or 3)using a non-random process which either looks at a limited number of options or looks at a certain subset of options first. If you use some type of non-random search, then the search will be more efficient only if a solution exists within the limited set of options being considered. This depends on the shape of the fitness landscape. A specific non-random search technique is only more efficient if the search routine is compatible with the specific fitness landscape being searched. Combine a ‘fast non-random search routine’ with an incompatible adaptive landscape and you can get an extremely inefficient search routine or one that will never find an adaptive solution.
There is a seriously flawed argument that the ‘evolution of complex, higher improbable adaptive forms or solutions’ can be produced by very simplistic Darwinian and neo-Darwinian processes because, to quote your phrase, "certain types of incredibly vast spaces that are searched amazingly effectively using random variation-selection schemes". These amazingly fast searches are achieved by non-random, not random search processes and they only work if the fitness environments have been manipulated to fit the search routine. It should be apparent that the ‘simple evolutionary processes’ and a ‘highly manipulated fitness landscape’ is simply a variation of the ‘external designer’ explanation of evolution.
It is again useful to note that TA models recognize that adaptive or evolutionary change processes can be speeded up both by the use of non-random search techniques and by increasing the number of cycles per unit of time. Specifically, including multiple variation-selection sub-processes and including within lifetime variation-selection techniques massively increases search power. Non-random search techniques are recognized and used in analysis with TA’s, but it is recognized that non-random search techniques are only effective if the search routine selected is compatible with the specific fitness landscape being searched.
Quote: With that, I am going to mostly retire from this discussion
I am sorry to see you withdraw as you raised any number of good issues. However, as the above discussion of non-random search points out, there are any number of technical issues that need to be addressed before we can seriously discuss the ‘simple’ experimental paradigm being proposed here.
SUMMARY OF CURRENT STATUS OF DISCUSSION
The discussion to date has provided a reasonable basis for concluding
1. There are ‘standard’ scientific analytical techniques which can reasonably be expected to make it possible to possible to represent or model and developmental transformation or operational transformation ‘A transforms to B’ as a complete set C or a partial set P of automated assembly instructions. C and P can be expressed as finite sets c1,c2, …,cn and p1,p2,… ,pm respectively.
2. Using the set C of automated assembly instruction it would, in theory, be possible to simulate transformations of the type ‘A transforms to B’
3. Using standard mathematical techniques, it is possible to define a system TA with n variation-selection process, one for each cx in C.
4. It therefore follows that for any evolutionary change involving a change in assembly, it is possible to define a set CB (before change) and a set CA (after evolutionary change). We can similarly define two systems TAB and TAA associated with CB and CA.
5. If CB and CA involve different forms of the same set of assembly instructions, then TAB=TAA and the system TAB can model and simulate the evolutionary change from CB to CA.
6. If we create a system TAC with the ability to a)add assembly instructions, b)increase the set S of assembly instruction triggers, and c)increase the set R of assembly instruction actions, then TAC can model and simulate any evolutionary change. [the pseudo code representing these processes has not yet been presented but it is available.]
I have not yet demonstrated that the defined approach is useful or productive. I have, however, I believe, demonstrated that the defined approach makes it possible to produce causal models and simulations of any type of evolutionary change using TA (modified GA) system.
The rational underlying the proposed approach is that in order to understand and explain either the development or operation of biological systems you have to understand the S-R relationships, (also referred to as automated assembly or operating instructions and dynamic and teleological causal relationships). To understand S-R relationships you need to understand the processes responsible for changing S-R relationships. As discussed, these change processes can be modeled by TA systems.
The next step in the discussion is to understand how S-R relationships and processes which model changes in S-R relationships are analyzed. I suggest or propose or predict that this analysis of S-R relationships will lead to ‘predictive models not only of developmental processes, but also predictive models of evolutionary change.
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 12. February 2003 12:10
gedanken,
Quote: This view of a "search" for a particular code will become relevant in a moment. As has been explained by RBH and others multiple times, and in multiple thread, evolution is not a "search" for a particular result.
Some individuals may prefer not to describe as evolution as a search and some may prefer not to describe evolutionary processes as teleological. Such preferences are not relevant to the discussion here where evolution is viewed as a teleological process controlled by ‘laws of nature’ that are ‘dynamic and teleological’.
Quote WB: Based on what was learned from the analysis of a small number of assembly instructions, it is suggested or predicted, we can formulate ‘theories’ which attempt to explain how the overall process works.
Quote g: Here we must ask what is the difference between this approach and what is already being done in biological science.
I have answered this before but it is probably worth repeating. Predictive scientific theories have the general form "Under ideal conditions A always causes B" or "Under ideal conditions F(A)=B". Using the ‘permanent and universal’ interpretation of determinism’, the function F is required to have a permanent and universal format(this would include permanent and universal stochastic relationships’.
As has been discussed, the assembly and operation of biological systems can be expressed as or reduced to assembly or operations ‘instructions’. These instructions are, or can be interpreted as, causal relationships which are ‘dynamic and teleological’. An assembly or operational instruction has the general form S causes R or F(S)=R. However, if F(S)=R is tracked over time, we find that F is changeable, dynamic, or programmable. We also find the changes in F tend to be in the direction of increasing the likelihood of survival. The assembly instructions tend to change so as to assembly useful end products.
If we analyze the operation and assembly of biological systems we find one level of dynamic and teleological process operating on another. So far no one has succeeded in finding a set of non-dynamic permanent and universal causal relationships controlling the sets of dynamic assembly/operations instructions. This, I suggest, explains the lack of reliable, testable, predictive theories in the life sciences.
My proposed solution to this dilemma is to construct predictive theories from dynamic and teleological causal relationships. Such theories have the general form "Under ideal conditions, the relationship F(S)=R takes the form most likely to produce goal G". This type of theory always produces a reliable prediction, but can produce different predictions at different times and locations. My work suggests this type of predictive scientific theory works as well or better in the scientific paradigm as the traditional ‘permanent and universal theory’ (which is a special case of the dynamic and teleological theory).
Quote: How can the "survival value" of any aspect be tested? Remember that Warren said "The mechanism I described is fairly simple and mechanical". "Survival" is tested when one continues to survive, or fails to survive. Actually the reproduction value is the more important to evolution, but that is closely related to "survival" as one must survive past the point of reproduction. What is the basis for determining that "survival value" has different numerical values at these different times?
This gets into the technical aspects of testing a predictive theory involving dynamic and teleological causal relationships. In a dynamic and teleological theory, the goal or purpose G, or more accurately the expected outcome G, is one of the ‘input variables’. Given the range of values for S and R, and given G, you calculate the F with the highest ‘expectation’ of achieving G’. From this information you predict the value of R that would be generated under ideal conditions by the biological system.
Consider as an example theories such as "Under ideal conditions, engineers will design and build the bridge most likely to achieve defined goals G" or "Under ideal conditions, investors will select investments with the most favorable expected return". If you know the range of options available and the criteria or goals to be satisfied, the first theory will normally produce a good prediction of the type of bridge to be built. The second theory would produce reliable predictions except for identifiable inefficiencies in the decision making process.
The concept of dynamic and teleological scientific theories may take a while to appreciate.
Quote: I’m Extremely confused! That GA’s can be used for multiple purposes does not seem to be relevant to how GA’s
My comment was in response to Rex’s view that GA’s were based on ‘between generation’ processing.
IP: Logged
|
|
RBH
Member
Member # 380
|
posted 12. February 2003 13:35
warren wrote quote:
There is a seriously flawed argument that the 'evolution of complex, higher improbable adaptive forms or solutions' can be produced by very simplistic Darwinian and neo-Darwinian processes because, to quote your phrase, "certain types of incredibly vast spaces that are searched amazingly effectively using random variation-selection schemes". These amazingly fast searches are achieved by non-random, not random search processes and they only work if the fitness environments have been manipulated to fit the search routine. It should be apparent that the 'simple evolutionary processes' and a 'highly manipulated fitness landscape' is simply a variation of the 'external designer' explanation of evolution.
But it is not universally true that the "fitness environments have been manipulated to fit the search routine." When GAs are used in applied search techniques, it is true that if the fitness environments (the several fitness landscapes on which the population of a GA resides) all have uncorrelated surfaces then the GA reduces to stochastic search. But many many fitness landscapes are naturally correlated, and therefore a GA is not performing a random stochastic search. See Richard Wein's critique of Dembski's NFL for a development of this point with reference to natural fitness landscapes.
I will note that in the GAs my firm develops and deploys, we have little control of the "fitness environment" - in particular, we cannot "manipulate" it. It is piped into the computers from the outside world - it is an array of a multitude of variables from the outside world - and we couldn't "manipulate" it if we wanted to. All we can do is offer it to the GAs and see what evolves. We offer as much as the input/output pipelines permit and that we have CPU cycles to process. The GAs "decide" (by evolving perceptual representation weightings) what aspects of that flood of information they will pay attention to and what they will ignore, not the human designers.
Further, no one to my knowledge has argued that evolution in either GAs or biology is "random." Gedanken's post, in speaking of "random variation-selection schemes," is explicitly identifying a non-random process: The "selection" step ensures non-randomness. In fact, the combination of random populational variation-producing processes (mutation and recombination being the principal such processes) and selection (differential reproduction as a function of phenotypic fitness) is a non-random process. warren is leaning on a non-existent door in his critique of evolutionary operators.
I await the examples warren promised. Perhaps they will help me understand what TA does that GAs and evolutionary processes in biology cannot. Meanwhile, I will repeat a remark I made in the Evolving Inventions thread about GAs and search and biology: quote:
This will sound like heresy to some, but biological evolution is not a search process. As conceived by evolutionary theory, it is a process in which variation and selection (along with the other evolutionary operators) in co-evolving populations (with the attendant dynamically deforming fitness landscapes) produce the appearance of search for optima, but that appearance is deceptive. To be sure, populations find sufficiently high local optima (or perish), but they are not searching for those optima. They find optima as a by-product of the operation of evolutionary operators, not as a goal. Hence analyzing biological evolution as a search problem leads to false analogies with human uses of search algorithms on both sides of the debate.
warren has fallen into that false analogy, and to the extent that he wishes his model to apply to biological evolution I fear he is expending a good deal of energy on a red herring. warren claims that quote:
Some individuals may prefer not to describe as evolution as a search and some may prefer not to describe evolutionary processes as teleological. Such preferences are not relevant to the discussion here where evolution is viewed as a teleological process controlled by 'laws of nature' that are 'dynamic and teleological'.
However, the question is not, as warren claims, whether one "prefers" to regard evolution as a search process, it is whether one can create a veridical model of biological evolution by so regarding it. Modeling biological evolution as a directed search process is to risk producing a false model, a model that does not represent the phenomena it purports to represent. In persistently misrepresenting both biological evolution and genetic algorithms as being exclusively directed search mechanisms, warren has created what I can only see as a false dilemma. The concrete examples he offered to provide may help resolve this, and I look forward to them.
RBH
[ 12. February 2003, 13:43: Message edited by: RBH ]
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 12. February 2003 14:57
I’ve got to focus on work today, so I won’t write long. (Oops, well so much for that, in EDIT)
Warren said:
quote:
I have answered this before but it is probably worth repeating. Predictive scientific theories have the general form "Under ideal conditions A always causes B" or "Under ideal conditions F(A)=B". Using the ‘permanent and universal’ interpretation of determinism’, the function F is required to have a permanent and universal format(this would include permanent and universal stochastic relationships’.
Warren includes “stochastic” mappings (e.g. that F(A)=B is an incorrect description). Recognition of this is helpful, but then the formalism does not represent what Warren just recognized.
The science of physics has predictions, and in quantum mechanics those predictions are of the form of relationships, not of functional maping of an “always” or deterministic nature. This is also exactly the state in evolutionary study, though much of the relationships in evolutionary studies are descriptive relationships rather than hard mathematical relationships. (There are very much both in biological studies -- don’t mistake my statement as implying that there is not a great deal of mathematical formalism to the relationships studied.) For example (and biologists please correct me if I’m wrong) evolutionary theory predicts that after a few starts of first life, that organisms will have a relationship of lineage. And predictions (of recent evolutionary theory) are that genetic material of descendents will be very close to the ancestors, with only a few minor changes beyond combinational effects. (Don’t here mistake genetic exchange between species that happens by other mechanisms as a contradiction to “lineage” concepts, they are very much physically realizable, local, and causal transfer mechanisms.) I will have to refer to biologists to provide links to more detailed examples. (The combinational effects are in part described by what was observed as old as Mendal, but now modern theory explains those effects in genetic terms).
So evolutionary theory does make predictions that are more or less “always” held in ideal conditions, but those are predictions of relationships not fixed mappings. (In other words they can be described as mappings of one to many and are most distinctly not deterministic.)
One problem is that any discrete stochastic mapping involving description of the relationships must be done in terms of a Markov model, or some other way of describing the various next states that could occur. The simplistic mapping is very inadequate for this. And furthermore many processes modeled in science will be continuous, so even a discrete stochastic state model (such as a Markov model) is insufficient.
Another problem is that many relationships are not described in terms of probability, and thus even a Markov model is virtually impossible.
The problem is that we have supersensitivty and extreme complexity in the systems Warren wishes to model. The problem of trying to describe progress of such systems is that they quickly expand into an ensemble of a very large number of possible states. The best way to analyze those states will turn out, I suspect, to be almost exactly what biological studies of evolution already are showing, precisely because they are based on precisely that sort of analysis! (It many not be done directly in the terminology I described, but the studies that are done can be traced to being equivalent to what I described.)
quote:
Some individuals may prefer not to describe as evolution as a search and some may prefer not to describe evolutionary processes as teleological. Such preferences are not relevant to the discussion here where evolution is viewed as a teleological process controlled by ‘laws of nature’ that are ‘dynamic and teleological’.
Of course you may invent your own world. Is it relevant? That is the issue to be analyzed. And to show that it is a relevant view it has to conform to observation. A tremendous amount of observation has been cataloged in the sciences. Now show some conformance, and predictivity in your concept! Show that they predict a relationship that is not predicted by current evolutionary theory, and which is furthermore reliably in conformance with observation!
quote:
This, I suggest, explains the lack of reliable, testable, predictive theories in the life sciences.
So I presume that quantum mechanics is not a “testable predictive theory” because it predicts a relationship rather than a given single deterministic result? You have already acknowledged that predictive relationships must include stochastic one to many relationships, why to you deny that biology’s theories of precisely that form are not “testable” or “predictive”?
With regard to How can the "survival value" of any aspect be tested?
quote:
This gets into the technical aspects of testing a predictive theory involving dynamic and teleological causal relationships. In a dynamic and teleological theory, the goal or purpose G, or more accurately the expected outcome G, is one of the ‘input variables’. Given the range of values for S and R, and given G, you calculate the F with the highest ‘expectation’ of achieving G’. From this information you predict the value of R that would be generated under ideal conditions by the biological system.
Warren, did you not say that of your concept “The mechanism I described is fairly simple and mechanical”? My question was not one that can be answered by validating the conformance with reality, as it is a point that the concept is not well-defined. Are you describing a system that works by way of external influences beyond the local causal relationships? If not, then the question remains, how does an actual physical system effect a selection that fits these constraints (to the degree that the sentences even make any sense)?
Here is an aspect of a question of a so-called “predictive” concept being well-defined and scientifically relevant. Suppose a concept is given that biological evolution follows a principle of a goal of systems that work functionally with interrelating functional subsystems. Then we observe that all biological systems do work functionally with interrelating functional subsystems. Is the concept validated as a scientific concept? Certainly not, as there is only a prediction of a relationship existing, not an explanation of the physical relationships of the real world that physically explain the relationships. That the world might be caused (for example) by a God who created the physical relationships of physical reality to have such a goal would neither be validated nor contradicted. Further physical relationships of that physical reality (whether or not intended as a goal by God) also explain that all biological systems must conform to the relationships stated. So these are not a prediction of a theory because no causal explanatory proposition has been made. Such a concept fails not because it is not validated as consistent with observation, but because it is not in the form of a scientific concept, and it is not distinguished from other concepts by that observation. Part of the reason for this is that the statement of the functionality being a “goal” is a statement that involves extensions beyond the physical entities being studied -- and thus not differentiated by the observation. A relationship such as gravity (without any further “causal” reasoning being proffered) is a scientific concept because it is a testable relationship that does not depend on elements outside of the physical system being described. Now a “goal” by an external system capable of creating such a “goal” and with testable linkages establish could be an example of a testable scientific proposition.
[ 12. February 2003, 16:39: Message edited by: gedanken ]
IP: Logged
|
|
warren_bergerson
Member
Member # 262
|
posted 13. February 2003 10:30
RBH,
Quote: But it is not universally true that the "fitness environments have been manipulated to fit the search routine." When GAs are used in applied search techniques, it is true that if the fitness environments (the several fitness landscapes on which the population of a GA resides) all have uncorrelated surfaces then the GA reduces to stochastic search. But many many fitness landscapes are naturally correlated, and therefore a GA is not performing a random stochastic search.
An interesting viewpoint. There are, however, a lot of different types of fitness landscapes associated with life forms and a lot of search routines. The probability that the right landscapes usually gets matched with the right search routine would be a miracle.
In observing biological systems we often find extremely fast and extremely effective search routines. This, it seems likely, is the result of the evolution or adaptive change in the search routine, not because of a natural affinity between search routine and fitness landscape. This ‘evolution’, it will be noted, will only occur it the same search is performed many, many times. If we follow this logical path, we not only need to explain the evolution of a very improbable event, but we need to explain enough repetitions of the evolutionary change to explain the evolution of an efficient evolutionary change process.
Quote: I will note that in the GAs my firm develops and deploys, we have little control of the "fitness environment" - in particular, we cannot "manipulate" it. It is piped into the computers from the outside world
One would not expect that your firm has the ability to manipulate the laws of the universe so that your search routines will produce results. However, you can and undoubtedly do select which type of problems you attempt to solve, or the techniques you use for solving a particular problem, based on the nature of the fitness environment you input into your computer.
Quote: Further, no one to my knowledge has argued that evolution in either GAs or biology is "random." Gedanken's post, in speaking of "random variation-selection schemes," is explicitly identifying a non-random process:
For reasons that seem to have been abandoned and forgotten, neo-Darwinian genetics, was based on the concept of random variation. However, if you abandon the concept of random variation for directed or non-random variation you are faced with the problem of explaining where did the direction come from? Conventional biology, in abandoning insistence on random variation, has conveniently avoided explaining ‘where the non-random’ searches can from.
Quote: However, the question is not, as warren claims, whether one "prefers" to regard evolution as a search process, it is whether one can create a veridical model of biological evolution by so regarding it. Modeling biological evolution as a directed search process is to risk producing a false model, a model that does not represent the phenomena it purports to represent.
You are misinterpreting my position. I am saying we can analyze and model biological change processes in terms of the mathematical concept of variation-selection search processes. This would include ‘open searches’ where the sets of Stimuli and Responses can increases as well as searches of closed spaces. I am saying that if you look at an established biological search processes you will find ‘non-random directed searches’. This, I suggest, is the result of biological change processes operating on biological change processes.
I am not saying that evolution is a directed search process, I am saying that evolution utilizes directed search processes, and that evolutionary processes can generate these directed search processes. It is traditional evolutionary theory that is trying to ‘sneak’ directed, non-random processes into explanations of evolutionary change without explaining how they got there.
It is important to repeat once again that there is a big difference between ‘scientific teleology’ and what could be labeled ‘metaphysical teleology’. Metaphysical teleology defines a goal or purpose in terms of a future event. Scientific teleology defines a goal as ‘an expectation of a future event based on past events’. You can denote the scientific concept of purpose or goal by GS (goal scientific) and the metaphysical concept by GM (goal metaphysical). I am proposing that mathematical predictive theories of life forms can be expressed in the form F(S, GS)=R . Conventional biology incorrectly argues that F(S, GM)=R is not a legitimate form or theory, therefore no teleological theory is valid.
Gedanken,
Quote: The science of physics has predictions, and in quantum mechanics those predictions are of the form of relationships, not of functional maping of an "always" or deterministic nature.
Predictions produced by predictive scientific theories need to satisfy three general requirements. First, the predictions must fit past observations. Second, predictions must be valid for current testing and validation. Third, there must be a reasonable expectations that the theory will produce reliable predictions in the future. Because of the success of Newtonian physics, this criteria got interpreted to mean science was only concerned with the analysis of permanent and universal causal relationships.
In the last 100 years the dogma of the permanent and universal interpretation of scientific determinism has been modified with the acceptance of stochastic theories. As you suggest, physics in accepting quantum mechanics is also apparently accepting some non-traditional form of scientific theory.
My proposal is to introduce teleological predictive theories or theories with the mathematical form F(S, GS)=R into the analysis of life forms. From a technical perspective there would appear to be no reason to reject teleological theories. At least no one has presented a technical reason for rejecting such theories. The resistance to teleological theories appears to be based entirely on an outdated dogma.
IMO, the only valid mathematical predictive theories in the life sciences are ‘limited scope’ theories (theories defining relationships under very restrictive sets of conditions). Anytime someone attempts to formulate a theory describing a general ‘permanent and universal’ relationship relating to life forms, one finds ways that life forms can change to invalidate the general theory. If you or anyone knows of a ‘valid, precisely defined, general mathematical predictive’ theory in the life sciences I would be happy to discuss it. The discussion here, however, does not depend on the existence or non-existence of permanent and universal life science theories. The discussion here depends on the technical soundness and acceptability of ‘scientific teleological theories’. The experimental paradigm being described here is design to develop and test teleological theories.
Quote: One problem is that any discrete stochastic mapping involving description of the relationships must be done in terms of a Markov model, or some other way of describing the various next states that could occur. The simplistic mapping is very inadequate for this. And furthermore many processes modeled in science will be continuous, so even a discrete stochastic state model (such as a Markov model) is insufficient.
Note that the modeling proposed here involves ‘discrete deterministic’ rather ‘discrete stochastic’ relationships. The modeling proposed is based not on modeling the ‘states of the systems during assembly but on modeling ‘discrete change of state’ operations’. The ‘automated assembly’ criteria is used to define the adequacy or completeness of the discrete set of change of state operations.
Quote: The problem is that we have supersensitivty and extreme complexity in the systems Warren wishes to model.
Again you are making a priori assumptions about what finding the experimental design will produce. I am suggesting that if you look at a life form at the micro level of assembly instructions you do not find a lot of seemingly random fluctuations that only produce meaningful results at the macro level. What you find, I suggest, at the micro level, is complex information processing with a rather incredible degree of precision.
Your suggestion on what type of analysis will produce productive results is entirely contrary to the prevailing trend. Research in almost all fields of behavior, physiology, genetics, and evolution started with the assumption that biological systems were relatively simple and would be explained by relatively simple models and processes. Time and again the assumed simplicity turned out to be extremely complex.
The experimental paradigm being proposed here is based on the assumption that if we want to understand evolution, we need to start the level of the lowly assembly or operating instruction. It will, I predict, quickly be shown that these assembly instructions are not ‘permanent instructions’ evolved over millions of years and passed from generation to generation. Instead, assembly instructions are highly dynamic operations that are continually being modified by complex, very powerful information processing mechanisms or machines. Despite the power and complexity involved, these information processing machines are explainable, I predict, in terms of known physical and chemical processes.
IP: Logged
|
|
|
Topic Closed
Printer-friendly view of this topic
