|
6/14/09 forum
Taken as a game, the goal of itatsi is to
score points. Points are a surrogate for reproductive success. Taken
as a simple simulation of evolution, a population that includes
scoreable words is healthy and successful.
In the column to the right, we have the results of a demo run in
itatsi, a hundred generations with mothers
picked by a fitness algorithm. It is reasonable to ask whether the
fitness algorithm has a target or goal.
There are about eight billion possible combinations of seven letters.
Of these eight billion, about 31,000 are English words in the
itatsi dictionary. The chances of producing
a seven letter word by random toss of the dice is about one in
260,000. The Demo player produced a seven letter word in 23
generations and proceeded to find several more rather quickly. The
31,000 seven letter words in the dictionary are not known to the
Weasel mutation algorithm, nor are they known to the algorithm that
scores "fitness" and selects the best candidate in each generation.
The fitness algorithm knows about two and three letter embedded
"words." No attempt is made to check the current population against
all of the possible 31,000 "targets."
Even if the attempt were made to score a generation of
itatsi children against 31,000 possible
targets, the time and computer resources required would be
prohibitive. The Training Weasel runs on an anonymous server at an
unknown location, with time purchased for a few dollars a month. Its
capabilities are not known, but it probably isn't a super computer.
One of the oddest phenomena in the history of the evolution debate is
the persistent misunderstanding of how evolution achieves "goals."
The simplest and most accurate response to the problem of goals is
that there are none. But this requires some explanation.
Television and popular science articles have trained us to think of
evolution as producing adaptations. We often see phrases like
"environmental pressure" or "selection pressure." We have been led to
think of evolution as an active hunt for solutions to problems.
Evolve or die.
This is a sad misunderstanding of what is going on and an unfortunate
opportunity for deliberate misrepresentation of the process of
evolution. The two most prominent critics of Evolution, Michael Behe
and William Dembski, have spent the greater part of their
professional careers trying to prove that Darwinian selection is an
inefficient and inadequate process for finding "targets" in a sea of
possibilities. Behe, for example, has issued a challenge to
experimental biologists to demonstrate that a bacterial flagellum can
evolve under controlled laboratory conditions. Dembski has written
books and papers arguing that evolutionary algorithms are no more
efficient than random searches.
All the discussion of targets and searches is a silly and
unproductive diversion. Humans can have targets. They can, for
example, have breed standards for pedigreed dogs. Players of
itatsi can have a word in mind as a goal.
Dawkins' Weasel program has a target. (It is worth noting that long
before evolution critics jumped on the target aspect of Weasel,
Dawkins anticipated the criticism.) In The Blind Watchmaker,
(Norton, 1996, p. 72) he writes:
Life isn't like that. Evolution has no long-term goal. There is no
long-distance target, no final perfection to serve as a criterion for
selection.
The argument [bacteria in a laboratory are not likely to evolve a
flagellum: therefore evolution can't produce complex
structures], is logically equivalent to arguing that a lottery
winner cheated or received divine assistance because he or she can't
do it again under controlled laboratory controls. The term for
calculating the odds against something happening after it has already
happened is Retrospective Astonishment. It makes some sense
if you have reason to suspect fraud, but the history of science has
few exhibits of weighted ping pong balls among natural phenomena. The
logical fallacy lies in assuming that a flagellum, or a particular
lottery winner was a prior goal or target. The math may be done
correctly, and the odds against may be astonishing, but odds are
irrelevant if there is no target specified in advance.
So what is going on in the game? If the object of the game isn't to
find a target, what is the object? The answer is both simple and
subtle. when you play itatsi, you shape a
population. You do so in much the same way you shape a hedge or a
topiary, by pruning. It would be nice to have the resources of some
of the serious evolution simulators, the ability to track tens of
thousands of individuals over thousands of generations. But we don't
have that. We have one small population and one mother per
generation. But within those confines, we can see the shaping of a
population by pruning.
Without knowing more than one language, you can run the demo and see
a random seed shaped in a few generations into a population that
resembles a selected language. The shaping occurs not because there
is a search or a target, but because the demo selects the child of
each generation that most nearly resembles the selected language
(based almost entirely on letter pairs). Looking at the Demo listing
to the right, you might ask yourself whether the
itatsi Demo selector had the word "TESTING"
in mind when it chose "RYSCATA" as more fit than "AYSCATA" in the
first generation.
Itatsi may look like a search for a target,
but there is an important difference.
Itatsi can produce children that look like
members of a language family, but have never been in a dictionary and
have never been in print - words like "COBLING" and "CHALING" and
"SULTING." Nothing separates these constructions from real words
except context. In another environment or another language, these
might be scoreable words.
The itatsi demo can shape a population into
something that looks like words in the language of choice, without a
specific target. It can make new and original language objects. A
selection process implies criteria, but requires neither intelligence
nor targets to produce novel and complex structures. The entire
argument regarding targets is a smokescreen. An oddly transparant
smokescreen.
|
Gen Pts Word Lang
1 04 SCAT En [AYSCATA]
1 01 CAT En [AYSCATA]
2 SCAT [RYSCATA]
3 SCAT [YSCATAY]
3 01 TAY En [YSCATAY]
4 SCAT [TSCATAY]
5 CAT [TOCATAY]
5 01 OCA En [TOCATAY]
5 01 TOC En [TOCATAY]
6 CAT [TOCATAH]
7 01 OAT En [TOOATAH]
7 01 TOO En [TOOATAH]
8 OAT [TOOATAB]
8 01 TAB En [TOOATAB]
9 OAT [TOOATOB]
10 04 ATOK En [TOOATOK]
11 ATOK [COOATOK]
11 01 COO En [COOATOK]
12 04 ATOP En [COOATOP]
12 01 TOP En [COOATOP]
13 COO [COOATON]
13 01 TON En [COOATON]
14 COO [COOATOR]
14 01 TOR En [COOATOR]
15 COO [COOATNG]
16 COO [COOAONG]
17 01 CON En [CONAONG]
18 04 HONG En [CONHONG]
18 01 HON En [CONHONG]
19 04 HING En [CONHING]
19 01 HIN En [CONHING]
20 04 NAIN En [XONAING]
20 01 AIN En [XONAING]
21 NAIN [IONAING]
21 01 ION En [IONAING]
22 NAIN [PONAING]
23 50 PONKING En [PONKING]
23 04 KING En [PONKING]
23 04 PONK En [PONKING]
23 01 KIN En [PONKING]
24 50 BOCKING En [BOCKING]
24 04 BOCK En [BOCKING]
25 BOCKING [BOCKING]
26 KING [KOCKING]
27 50 ROCKING En [ROCKING]
27 04 ROCK En [ROCKING]
27 01 ROC En [ROCKING]
28 KING [FOCKING]
29 KING [FOCKING]
30 04 JOCK En [JOCKING]
31 50 COCKING En [COCKING]
31 04 COCK En [COCKING]
32 50 CORKING En [CORKING]
32 04 CORK En [CORKING]
32 01 COR En [CORKING]
33 04 LING En [CORLING]
33 01 LIN En [CORLING]
34 50 CORNING En [CORNING]
34 09 CORNI En [CORNING]
34 04 CORN En [CORNING]
35 CORNING [CORNING]
36 01 COL En [COLNING]
37 50 CONNING En [CONNING]
37 16 ONNING En [CONNING]
37 04 CONN En [CONNING]
38 CON [CONCING]
39 CON [CONCING]
40 LING [CONLING]
41 LING [CONLING]
42 04 BLIN En [COBLING]
42 01 COB En [COBLING]
43 50 COOLING En [COOLING]
43 04 COOL En [COOLING]
44 50 COWLING En [COWLING]
44 16 OWLING En [COWLING]
44 04 COWL En [COWLING]
44 01 COW En [COWLING]
44 01 OWL En [COWLING]
45 COWLING [COWLING]
46 BLIN [COBLING]
47 50 COALING En [COALING]
47 04 COAL En [COALING]
48 16 HALING En [CHALING]
48 04 CHAL En [CHALING]
48 01 CHA En [CHALING]
49 HALING [CHALING]
50 HALING [LHALING]
51 50 WHALING En [WHALING]
51 01 WHA En [WHALING]
52 LING [WLALING]
53 LING [FLALING]
54 LING [OLALING]
55 LING [ELALING]
56 LING [WLALING]
57 LING [LLALING]
58 LING [LCALING]
59 LING [ACALING]
60 LING [DCALING]
61 LING [OCALING]
62 LING [PCALING]
63 LING [PIALING]
63 04 PIAL En [PIALING]
63 01 PIA En [PIALING]
64 LING [PIELING]
64 01 PIE En [PIELING]
65 04 PIET En [PIETING]
65 04 TING En [PIETING]
65 01 TIN En [PIETING]
66 TING [PIGTING]
66 01 PIG En [PIGTING]
67 TING [PIOTING]
68 TING [PUOTING]
69 TING [SUOTING]
70 TING [SULTING]
71 TING [SULTING]
72 TING [SUCTING]
73 09 STING En [SUSTING]
73 01 SUS En [SUSTING]
74 STING [CUSTING]
75 STING [CUSTING]
76 50 RUSTING En [RUSTING]
76 04 RUST En [RUSTING]
77 50 MUSTING En [MUSTING]
77 04 MUST En [MUSTING]
77 01 MUS En [MUSTING]
78 STING [MESTING]
78 01 EST En [MESTING]
78 01 MES En [MESTING]
79 STING [DESTING]
80 STING [PESTING]
80 04 PEST En [PESTING]
80 01 PES En [PESTING]
81 STING [PESTING]
82 09 CESTI En [CESTING]
83 CESTI [CESTING]
84 STING [XESTING]
85 STING [PESTING]
86 50 TESTING En [TESTING]
86 04 TEST En [TESTING]
86 01 TES En [TESTING]
87 TESTING [TESTING]
88 STING [FESTING]
88 04 FEST En [FESTING]
89 STING [FESTING]
90 STING [FOSTING]
91 STING [FOSTING]
92 STING [DOSTING]
92 04 DOST En [DOSTING]
92 01 DOS En [DOSTING]
93 STING [DOSTING]
94 STING [ZOSTING]
94 01 ZOS En [ZOSTING]
95 STING [ZOSTING]
96 TING [ZOTTING]
97 50 LOTTING En [LOTTING]
97 01 LOT En [LOTTING]
98 TING [IOTTING]
99 TING [IOTTING]
100 TING [AOTTING]
|
|
5/26/09
Basically the Training Weasel is a very good player, a kibitzer who
stands over your shoulder and makes suggestions. So far, there are
two generations of the Trainer. The first,
is a simple daemon that looks at the words after they are created and
scores them by letter and position frequency. The
second requires some explanation.
First of all, it is important to understand that all versions of
itatsi use the same child generator, a simple Dawkins'
Weasel program that makes multiple copies of a parent while applying
an occasional mutation to letters as they are copied. The central
core of itatsi is an imperfect replicator that knows
nothing of the past or future or goals. All children are generated
before any colors are applied or any children are evaluated for
fitness. Colors are added to make game play easier. They do nothing
and imply nothing that could not be done by a human player. Colors
are applied to help human players quickly identify the children that
have been mutated.
The Trainer, however, has a special place. The silver highlighted
child has been evaluated by a fitness daemon and designated as the
most fit of the population. So what does that mean?
First, some terminology . Itatsi thinks of the children
as genomes having some level of fitness. They are instances of a
genetic code. Codes have no meaning in and of themselves. They are
interpreted. In the game, the ultimate interpreter is the scoring
program, which awards points for children that are in the
itatsi dictionary. But there are lower level
interpreters, the players who select children based on their
resemblance to words. The question addressed by the Training
Weasel is: can there be a simple algorithm that evaluates progress
toward wordness without having a massive database of
26x26x26x26x26x26x26 possible children?
The answer proposed here is yes, by applying a bit of
pseudo-biological thinking. First, itatsi thinks of its
children as genomes composed of two letter genes, with each gene
having 26x26 possible variations or alleles. A child having n letters
has n-1 genes (due to the way language users interpret the
code). So the genes of "WEASEL" are "WE", "EA", "AS", "SE", and "EL".
The Training Weasel algorithm breaks down each child into its genes
and scores each gene by the occurrence frequency of its allele.
There are 26x26 possible alleles formed by letter pairs using the
English alphabet. In reality, about 80 percent of possible alleles
are used.
But not with equal frequency. Itatsi knows how often
each allele occurs in each possible position, and uses this
information to score genes for fitness. The fitness of a child is the
sum of the fitness scores for each gene. The database has only a few
thousand entries, far fewer than the number of dictionary words, and
enormously fewer than the number of possible children.
So is this smuggling information into the Weasel program? If the
question applies to the imperfect replicator, the answer is
absolutely not. All the intelligence is in the selectors, the
Training Weasel and the human player. And isn't this what the theory
of evolution has said all along?
The Training Weasel has a number of interesting behaviors. It can
form words, but doesn't necessarily recognize them. It quickly forms
pronounceable letter strings, most of which are not words. It often
scores the same string several times as best without knowing that it
is no longer eligible for game scoring. It can get stuck, scoring
non-words as best, because their genes, taken individually, are best.
The Training Weasel is blind to the ultimate selecting environment.
|
