Use Smoothiness to classify rules
Re: Use Smoothiness to classify rules
Guys.
I am sorry but I made a big mistake about the orientation of the NHT matrix. The input-output axis was flipped in previous posts. They should be fixed by now.
(The "reshape": command in matlab fill in the elements by column, not by row. )
I am sorry but I made a big mistake about the orientation of the NHT matrix. The input-output axis was flipped in previous posts. They should be fixed by now.
(The "reshape": command in matlab fill in the elements by column, not by row. )
Re: Use Smoothiness to classify rules
Here is one possible way to visualise the information contained in NHT:Bullet51 wrote:A somewhat difficult question: How much information can we extract from the transition matrix?
Plot the within-N entropy against the entropy weight of N (which is -F(N)*log2(F(N))), repeat this for every neighborhood to get a scatter plot. On such scatter plots, complex behavior looks fairly different from chaotic ones.
We should be able to develop some statistics from here. Note for complex behavior dots scatter onto x-axis with no height, while for chaos dots are not attracted to x-axis. This partially confirm my idea that "complex is more deterministic than chaos"
- Attachments
-
- stable.jpg (29.88 KiB) Viewed 11385 times
-
- complex4.jpg (42.84 KiB) Viewed 11385 times
-
- chaos1.jpg (50.71 KiB) Viewed 11385 times
Re: Use Smoothiness to classify rules
Here's somehow effective statistics to describe complexity.
(x,y) where x=variance explained by trend in sum(entropyA*entropyB) time series.
y=mean( sum(entropyA*entropyB) time series )
But manual inspection reveal that right-bottom corner do contain complex behavior falsely classifed as chaotic. BUT their scale is usually higher than the neighborhood taken.
(x,y) where x=variance explained by trend in sum(entropyA*entropyB) time series.
y=mean( sum(entropyA*entropyB) time series )
But manual inspection reveal that right-bottom corner do contain complex behavior falsely classifed as chaotic. BUT their scale is usually higher than the neighborhood taken.
- Attachments
-
- v1res.jpg (32.98 KiB) Viewed 11370 times
Re: Use Smoothiness to classify rules
Under the earlier notation, I did a manual scan based on the recommendation parameters. As it turns out, these statistics captures some aspects of the behavior, but not quite precise the emergent nature yet.
The complied list is partially color coded.
Yellow: abundant moving heads/gliders/stable recognisable recurrent separable objects
Orange: same as yellow, but plagued with chaotic mess, usually in the centre (complex behavior at edge)
Dark orange: more chaos
dark to light purple: These rules tend to fall stable quickly/ oscillator rich, but little long-range spatial communication.
Blue/Cyan: these rules are very stable, but they exhibit fractal behavior in some aspect.
Blank: Not interesting enough/ very ordered/ close related to some other rules here. They are kept mainly for comparative analysis between close related rules.
Let's vote for your favorite color!!!
UPDATE:The signatures for these rules have been complied into NHT_ambig. I am currently storing them as .jpg files. Would love switch to some 3D format if there is any recommendation.
The complied list is partially color coded.
Yellow: abundant moving heads/gliders/stable recognisable recurrent separable objects
Orange: same as yellow, but plagued with chaotic mess, usually in the centre (complex behavior at edge)
Dark orange: more chaos
dark to light purple: These rules tend to fall stable quickly/ oscillator rich, but little long-range spatial communication.
Blue/Cyan: these rules are very stable, but they exhibit fractal behavior in some aspect.
Blank: Not interesting enough/ very ordered/ close related to some other rules here. They are kept mainly for comparative analysis between close related rules.
Let's vote for your favorite color!!!
UPDATE:The signatures for these rules have been complied into NHT_ambig. I am currently storing them as .jpg files. Would love switch to some 3D format if there is any recommendation.
- Attachments
-
- ent_collect.xls.tar.gz
- (15.22 KiB) Downloaded 196 times
Last edited by shouldsee on August 14th, 2016, 6:33 pm, edited 1 time in total.
Re: Use Smoothiness to classify rules
Further to the aforementioned 'within-NH entropy" and "between-NH entropy/forward uncertainty" plot.
I found it helpful to also include the 'backward uncertainty' as well as 'forward uncertainty', and to record the scatter plot over time. This provides a better signature of the dynamics.
The collection is uploaded as NHT1
Let's extract some statistics from these distributions.
I found it helpful to also include the 'backward uncertainty' as well as 'forward uncertainty', and to record the scatter plot over time. This provides a better signature of the dynamics.
The collection is uploaded as NHT1
Let's extract some statistics from these distributions.
- Attachments
-
- B367S02358 _soup62.jpg (233.12 KiB) Viewed 11341 times
-
- B01234S012357_soup37.jpg (194.42 KiB) Viewed 11341 times
-
- B3S23 _soup2.jpg (140.88 KiB) Viewed 11341 times
Re: Use Smoothiness to classify rules
EDIT: Actually I messed up with forward uncertainty and backward uncertainty.
Let me be concise.
1.Forward predictability+chaos = complex
2.Backward predictability = order
3.No predictability = chaos
4. For complex behaviour, the states contributing to each element must directly inter-convert.
These are conjectures only and await more evidence.
Let's try to define 'interesting'.
One idea has knocked around since study on CA begins- we can use compressibility to define spaceship, because spaceships exhibit behavior periodic in time but aperiodic in space.
This idea naturally leads to the conjecture, that interesting rules are reversible. Imagine we got a universe of spaceships, we can easily deduce the configuration at an earlier time by reverse the speed of all spaceships. Thus, the more reversible behavior in a universe, the more interesting it is.
We can evaluate the reversibility by counting the states with a backward uncertainty near zero. For example, B01367S01 has a big collection of states near Bunc=0. B3S23 has a smaller collection. This reflects the fact that B01367S01 has more spontaneous spaceships/replicators than B3S23.
This criteria above is not enough to gurantee interesting behavior, it's only essential. Another essential is that these Bunc=0 states need to be connected to uncertain chaotic states that is unpredictable both forward and backward. Otherwise there will be no unpredictability in the universe.
Furthermore, for classical behavior like B3S23, we also need a stable background. This background will manifest as highly predictable (low uncertainty) both forward and backward, and constitute a large fraction of the universe (high entropic weight).
Let me be concise.
1.Forward predictability+chaos = complex
2.Backward predictability = order
3.No predictability = chaos
4. For complex behaviour, the states contributing to each element must directly inter-convert.
These are conjectures only and await more evidence.
Let's try to define 'interesting'.
One idea has knocked around since study on CA begins- we can use compressibility to define spaceship, because spaceships exhibit behavior periodic in time but aperiodic in space.
This idea naturally leads to the conjecture, that interesting rules are reversible. Imagine we got a universe of spaceships, we can easily deduce the configuration at an earlier time by reverse the speed of all spaceships. Thus, the more reversible behavior in a universe, the more interesting it is.
We can evaluate the reversibility by counting the states with a backward uncertainty near zero. For example, B01367S01 has a big collection of states near Bunc=0. B3S23 has a smaller collection. This reflects the fact that B01367S01 has more spontaneous spaceships/replicators than B3S23.
This criteria above is not enough to gurantee interesting behavior, it's only essential. Another essential is that these Bunc=0 states need to be connected to uncertain chaotic states that is unpredictable both forward and backward. Otherwise there will be no unpredictability in the universe.
Furthermore, for classical behavior like B3S23, we also need a stable background. This background will manifest as highly predictable (low uncertainty) both forward and backward, and constitute a large fraction of the universe (high entropic weight).
- Attachments
-
- B01367S01_soup11.jpg (209.42 KiB) Viewed 11295 times
-
- B3S23_soup202.jpg (147.26 KiB) Viewed 11295 times
Last edited by shouldsee on August 15th, 2016, 1:16 pm, edited 1 time in total.
Re: Use Smoothiness to classify rules
The rule B378S126 may be a good sample to test. How many states are there near Bunc=0?shouldsee wrote:We can evaluate the reversibility by counting the states with a backward uncertainty near zero. For example, B01367S01 has a big collection of states near Bunc=0. B3S23 has a smaller collection. This reflects the fact that B01367S01 has more spontaneous spaceships/replicators than B3S23.
Still drifting.
Re: Use Smoothiness to classify rules
The collection is uploaded as LogicBullet51 wrote:The rule B378S126 may be a good sample to test. How many states are there near Bunc=0?shouldsee wrote:We can evaluate the reversibility by counting the states with a backward uncertainty near zero. For example, B01367S01 has a big collection of states near Bunc=0. B3S23 has a smaller collection. This reflects the fact that B01367S01 has more spontaneous spaceships/replicators than B3S23.
Re: Use Smoothiness to classify rules
Still refining the definition of interesting.
UPDATE: As it came out, the points with forward/backward uncertainty of zero are largely due to the base period 2 in B0 rules, which has led me to falsely conjectured a connection between these points and the interestingness of rules. After normalising the plot, the points are scatter along the plane x=y.
If we consider a spaceship moving through space, the vain in front of it is destined to get occupied by the spaceship, thus they are forward predictable. The vain to the tail of the spaceship, however, is bound to be occupied by the spaceship previously, thus they are backward predictable. Between these two edges, something happens in between.
Thus, by measuring the propensity of transition between these edges, we can get to taste the nature of this rule.
Attaches is another round of interesting rules. For some reason my collection is enriched of B0 rules.
UPDATE: As it came out, the points with forward/backward uncertainty of zero are largely due to the base period 2 in B0 rules, which has led me to falsely conjectured a connection between these points and the interestingness of rules. After normalising the plot, the points are scatter along the plane x=y.
If we consider a spaceship moving through space, the vain in front of it is destined to get occupied by the spaceship, thus they are forward predictable. The vain to the tail of the spaceship, however, is bound to be occupied by the spaceship previously, thus they are backward predictable. Between these two edges, something happens in between.
Thus, by measuring the propensity of transition between these edges, we can get to taste the nature of this rule.
Attaches is another round of interesting rules. For some reason my collection is enriched of B0 rules.
- Attachments
-
- flux_collect.xls.tar.gz
- (15.38 KiB) Downloaded 220 times
Re: Use Smoothiness to classify rules
Some visualisation of the chaotic core in rulespace.
The red dots are complex rules from my test collection. The blue dots are random samples.
UPDATE: Legend added and B3/S23 pointed out.
FE=forward entropy
BE=backward entropy
flux are measured along the FE+BE, FE-BE axis, and weighted by the number of transition between corresponding input configurations. To remove the strobing effect, transition is taken every 6 steps.
A 3D version in form of matlab .fig can be found here
I have also updated the front page of this post.
raw statistics can be found here. Columns 2,3,4 are mf(FE+BE) mf(FE-BE), m(input entropy), 5,6,7 gives standard deviation std() instead of mean().
Sadly, the arm region indicated with red dots doesn't constitute an exclusive area, i.e interesting rules in curated collection are bound to lie in such region, but points in such region are not necessary interesting. This indicates that we may still need extra dimension to expand this 'arm' region.
The red dots are complex rules from my test collection. The blue dots are random samples.
UPDATE: Legend added and B3/S23 pointed out.
FE=forward entropy
BE=backward entropy
flux are measured along the FE+BE, FE-BE axis, and weighted by the number of transition between corresponding input configurations. To remove the strobing effect, transition is taken every 6 steps.
A 3D version in form of matlab .fig can be found here
I have also updated the front page of this post.
raw statistics can be found here. Columns 2,3,4 are mf(FE+BE) mf(FE-BE), m(input entropy), 5,6,7 gives standard deviation std() instead of mean().
Sadly, the arm region indicated with red dots doesn't constitute an exclusive area, i.e interesting rules in curated collection are bound to lie in such region, but points in such region are not necessary interesting. This indicates that we may still need extra dimension to expand this 'arm' region.
- Attachments
-
- BSrulespace_2d.jpg (81.71 KiB) Viewed 11227 times
-
- BSrulespace.jpg (82.6 KiB) Viewed 11228 times
Last edited by shouldsee on August 21st, 2016, 12:48 pm, edited 2 times in total.
Re: Use Smoothiness to classify rules
agar crawlers
Code: Select all
x = 69, y = 220, rule = B0345678/S0158:T69,220
2obo2bobo2bobobobobo2bobo2bobobobobo2bobobobobobobobobobobobobob2obo$
69o2$69o2$bobo2bobo2bobobobobo2bobo2bobobobobo2bobobobobobobobobobobob
obob2obo2$69o2$69o$obob2obob2obobobobob2obob2obobobobob2obobobobobobob
obobobobobobo2bobo$69o2$69o2$obob2obob2obobobobob2obob2obobobobob2obob
obobobobobobobobobobobo2bobo2$69o2$28o5b36o$bobo2bobo2bobobobobo2bobo
2b7obo2bobobobobobobobobobobobobob2obo$29obob37o$30bo$29o3b37o$27b2o3b
2o$bobo2bobo2bobobobobo2bobob2obobob3o2bobobobobobobobobobobobobob2obo
$29bobo$30ob38o$30bo$29obob2ob34o$obob2obob2obobobobob2obobobo2bobo4bo
bobobobobobobobobobobobobo2bobo$29obob37o$30bo$30ob38o$27b2o$obob2obob
2obobobobob2obob9obobobobobobobobobobobobobobobo2bobo$27b2o2bo$69o2$
69o$bobo2bobo2bobobobobo2bobobobo7bobobobobobobobobobobobobobob2obo$
30obobob34o2$69o2$bobo2bobo2bobobobobo2bobob13obobobobobobobobobobobob
ob2obo$28bobobobobo$69o2$69o$obob2obob2obobobobob2obo17bobobobobobobob
obobobobo2bobo$25obobobobobobobob29o2$69o2$obob2obob2obobobobob24obobo
bobobobobobobobo2bobo$23bobobobobobobobobobo$69o2$69o$bobo2bobo2bobobo
bobobo24bobobobobobobobobob2obo$24obobobobobobobobobobob24o2$69o2$bobo
2bobo2bobobobobobobob24obobobobobobobob2obo$26bobobobobobobobobobobo$
69o2$69o$obob2obob2obobobobobobobobo24bobobobobobobo2bobo$29obobobobob
obobobobobob19o2$69o2$obob2obob2obobobobobobobobobob24obobobobobo2bobo
$31bobobobobobobobobobobo$69o2$69o$bobo2bobo2bobobobobobobobobobobo24b
obobobob2obo$34obobobobobobobobobobob14o2$69o2$bobo2bobo2bobobobobobob
obobobobobob24obobob2obo$36bobobobobobobobobobobo$69o2$69o$obob2obob2o
bobobobobobobobobobobobobo24bobo2bobo$39obobobobobobobobobobob9o2$69o
2$obob2obob2obobobobobobobobobobobobobobob24o2bobo$41bobobobobobobobob
obobo$69o2$69o$bobo2bobo2bobobobobobobobobobobobobobobobo23bobo$44obob
obobobobobobobob6o2$69o2$bobo2bobo2bobobobobobobobobobobobobobobobobob
17obobobo$46bobobobobobobobo$69o2$69o$obob2obob2obobobobobobobobobobob
obobobobobobobo13bobobobobo$49obobobobob11o2$69o2$obob2obob2obobobobob
obobobobobobobobobobobobobobob7obobobobobobo$51bobobo$69o2$69o$bobo2bo
bo2bobobobobobobobobobobobobobobobobobobobobo3bobobobobobobo$69o2$69o
2$bobo2bobo2bobobobobobobobobobobobobobobobobobobobobobobobobobobobobo
2$69o2$69o$obob2obob2obobobobobobobobobobobobobobobobobobobobobobobobo
bobobobobo$69o2$69o2$obob2obob2obobobobobobobobobobobobobobobobobobobo
bobobobobobobobobobo2$69o2$69o$bobo2bobo2bobobobobobobobobobobobobobob
obobobobobobobobobobobobobobo$69o2$69o2$bobo2bobo2bobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobo2$69o2$69o$obob2obob2obobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobo$69o2$69o2$obob2obob2o
bobobobobobobobobobobobobobobobobobobobobobobobobobobobobo2$69o2$69o$b
obo2bobo2bobobobobobobobobobobobobobobobobobobobobobobobobobobobobo$
69o2$69o2$bobo2bobo2bobobobobobobobobobobobobobobobobobobobobobobobobo
bobobobo2$69o2$69o$obob2obob2obobobobobobobobobobobobobobobobobobobobo
bobobobobobobobobo$69o2$69o2$obob2obob2obobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobo2$69o2$69o$bobo2bobo2bobobobobobobobobobob
obobobobobobobobobobobobobobobobobobo$69o2$69o2$bobo2bobo2bobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobo2$69o2$69o$obob2obob2obo
bobobobobobobobobobobobobobobobobobobobobobobobobobobobo$69o2$69o2$obo
b2obob2obobobobobobobobobobobobobobobobobobobobobobobobobobobobobo2$
69o2$69o$bobo2bobo2bobobobobobobobobobobobobobobobobobobobobobobobobob
obobobo$69o2$69o2$bobo2bobo2bobobobobobobobobobobobobobobobobobobobobo
bobobobobobobobo2$69o2$69o$obob2obob2obobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobo$69o2$69o2$obob2obob2obobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobo2$69o2$69o!
Code: Select all
x = 243, y = 12, rule = B12348/S123468:T243,12
96obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobob2o$95bobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobo$94obobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobo$93bobobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobob2o$94obobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobo$82b2o9bobo
bobobobobobobob5obobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obo$94obobobobobobobobob7obobobobobobobobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobo
bobobobob2o$81b3o11bobobobobobobobob7obobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobo
bobobobobobobobobobobo$96obobobobobobobob5obobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobob2o$82b2o13bobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobo$98obobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobo
bobobobobobobobobobobobobobobobobobobobobobobobobob2o$o96bobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobob
obobobobobobobobobobobobobobobobobobobobobobobobobobobobobobobo!
Last edited by shouldsee on September 2nd, 2016, 2:04 am, edited 1 time in total.
Re: Use Smoothiness to classify rules
2 steps is enough.shouldsee wrote:To remove the strobing effect, transition is taken every 6 steps.
Points in such region, of course, are not necessary interesting, but most of them are interesting:shouldsee wrote:Sadly, the arm region indicated with red dots doesn't constitute an exclusive area, i.e interesting rules in curated collection are bound to lie in such region, but points in such region are not necessary interesting.
Code: Select all
G:Glider-Possible
N:Natural-Glider-Possible
S:Special
A:Ameoba dynamics
F:Fast-Decay
B357/S238 N
B013568/S0123 N
B012358/S01237 G
B3468/S028 G
B01245/S14 G
B0123457/S0126 N
B2346/S124678 S
[The negative rule is b358s01378, which is N]
B0127/S0135 G
B358/S136 N
B01367/S01 G
[Possibly N]
B123568/S123678 S
[The negative rule is b348s148, which is N]
B01347/S13 N
B0127/S0235 G
B2578/S0234678 S
[The negative rule is b37s24578, which is G]
B378/S024678 N
B01237/S236 S
[It has natural puffers, but not natural gliders]
B013678/S0 N
B012456/S023 N
B0123567/S02346 G
[Possibly N]
B378/S02458 G
[Possibly N]
B3678/S135 N
B012467/S1235 N
B358/S12678 N
B12/S12478 S
[Biphasic. May be regarded as complex behaviour]
B37/S0245 N
[Has a 4c/25 glider: b5o$o4bo$o4bo$b5o!]
B3568/S136 N
B012345/S0246 N
B0134568/S0234 N
[The natural ship is actually an infinite class of ships]
B013457/S0123 G
[There are too many replicators]
B013458/S02 G
[There are too many replicators]
B01267/S0125 A
B358/S1267 N
B123478/S24678 S
[The negative rule is b3578s238, which is G, and have a replicator]
B01258/S025 G
[Possibly N]
B024678/S04678 S
[The negative rule is b3567s3578, which is G]
B0134568/S1234 S
[It has natural puffers, but not natural gliders]
B01238/S1246 G
B012/S015 F
B34/S368 F
B012357/S14678 S
[The negative rule is b3578s238, which is N]
B01234567/S035 A
B0123578/S135 A
B01246/S01234678 S
[The negative rule is b3s0135, which is both N and puffer-natural]
B346/S0567 G
B3468/S01 G
[Sometimes explodes]
B012458/S4 A
B0123457/S023678 S
[The negative rule of b347s02, which is G]
B012567/S014 A
[It converges to some strange states]
B348/S037 G
B123458/S024678 S
[Negative of b357s128, which is G]
B012/S145 A
B356/S034 G
[Maybe F]
B1578/S234678 S
[Negative of b378s24568, which is full of gliders]
B0123456/S01235 F
[Has a glider]
B01278/S012468 S
[Biphasic. May be regarded as complex behaviour]
B0138/S4 N
B0123578/S1236 G
[Heavy ameoba dynamics]
B3568/S0146 N
Surely.shouldsee wrote:This indicates that we may still need extra dimension to expand this 'arm' region.
Would you like to share your program, since I think it's a good idea to use the program to search hexagonal rules:
Code: Select all
x = 16, y = 7, rule = B0126/S015H
5bo6bo$4bo6bo$obobob8o$o2bo2b8obo$bobobob8o$6bo6bo$8bo6bo!
Still drifting.
Re: Use Smoothiness to classify rules
I don't think 2 would be enough, since p3 strobing is quite frequent for cooperative oscillators. Use a different lagging length would also change the shape of the landscape. But no harm trying.Bullet51 wrote:2 steps is enough.shouldsee wrote:To remove the strobing effect, transition is taken every 6 steps.
Did you tried to bulid a classifer under inspection here? Or what's the shape recognised by your NN? DId you just used the 3 features I provided?Bullet51 wrote:Points in such region, of course, are not necessary interesting, but most of them are interesting:shouldsee wrote:Sadly, the arm region indicated with red dots doesn't constitute an exclusive area, i.e interesting rules in curated collection are bound to lie in such region, but points in such region are not necessary interesting.
The region is determined by a neural network.
After all we managed to filter some out, but not yet enough.Bullet51 wrote:Surely.shouldsee wrote:This indicates that we may still need extra dimension to expand this 'arm' region.
Sure. It just took some cleaning work. I have updated the link on the frontpage. Although to apply for hexagonal, you would need to undertake a extensive change.Bullet51 wrote: Would you like to share your program, since I think it's a good idea to use the program to search hexagonal rules:
Re: Use Smoothiness to classify rules
Yes.shouldsee wrote:DId you just used the 3 features I provided?
Somewhat hard to say, since I don't know how to plot the surface of f(x,y,z)=0.shouldsee wrote:Or what's the shape recognised by your NN?
Thanks!shouldsee wrote:Sure. It just took some cleaning work. I have updated the link on the frontpage. Although to apply for hexagonal, you would need to undertake a extensive change.
Still drifting.
Re: Use Smoothiness to classify rules
Try contour3.Bullet51 wrote:Somewhat hard to say, since I don't know how to plot the surface of f(x,y,z)=0.shouldsee wrote:Or what's the shape recognised by your NN?
You are welcome.Bullet51 wrote: Thanks!
Re: Use Smoothiness to classify rules
The NN may seem somewhat inaccurate, but in fact, it has a accuracy of 94.5%, just because I weighted the complex rules 94x higher.shouldsee wrote:Or what's the shape recognised by your NN?
Still drifting.
Re: Use Smoothiness to classify rules
Since my simulator does not distinguish between rules and their negative equivalents, we should classify postive and negative rules into the same class. i.e. if an S rule has a G negative counterpart, itself should be G instead of S. The plague of biphasic behavior is somewhat confusing, since there is not a general framework to classify within this discipline. Although I did observe some gliders, the heterogeneity in the randomly generated background make it hard to undertake systematic search, although I think randomagar should be able to cope with such things.Bullet51 wrote:[
Points in such region, of course, are not necessary interesting, but most of them are interesting:The region is determined by a neural network.Code: Select all
G:Glider-Possible N:Natural-Glider-Possible S:Special A:Ameoba dynamics F:Fast-Decay B357/S238 N B013568/S0123 N B012358/S01237 G B3468/S028 G B01245/S14 G B0123457/S0126 N B2346/S124678 S [The negative rule is b358s01378, which is N] B0127/S0135 G B358/S136 N B01367/S01 G [Possibly N] B123568/S123678 S [The negative rule is b348s148, which is N] B01347/S13 N B0127/S0235 G B2578/S0234678 S [The negative rule is b37s24578, which is G] B378/S024678 N B01237/S236 S [It has natural puffers, but not natural gliders] B013678/S0 N B012456/S023 N B0123567/S02346 G [Possibly N] B378/S02458 G [Possibly N] B3678/S135 N B012467/S1235 N B358/S12678 N B12/S12478 S [Biphasic. May be regarded as complex behaviour] B37/S0245 N [Has a 4c/25 glider: b5o$o4bo$o4bo$b5o!] B3568/S136 N B012345/S0246 N B0134568/S0234 N [The natural ship is actually an infinite class of ships] B013457/S0123 G [There are too many replicators] B013458/S02 G [There are too many replicators] B01267/S0125 A B358/S1267 N B123478/S24678 S [The negative rule is b3578s238, which is G, and have a replicator] B01258/S025 G [Possibly N] B024678/S04678 S [The negative rule is b3567s3578, which is G] B0134568/S1234 S [It has natural puffers, but not natural gliders] B01238/S1246 G B012/S015 F B34/S368 F B012357/S14678 S [The negative rule is b3578s238, which is N] B01234567/S035 A B0123578/S135 A B01246/S01234678 S [The negative rule is b3s0135, which is both N and puffer-natural] B346/S0567 G B3468/S01 G [Sometimes explodes] B012458/S4 A B0123457/S023678 S [The negative rule of b347s02, which is G] B012567/S014 A [It converges to some strange states] B348/S037 G B123458/S024678 S [Negative of b357s128, which is G] B012/S145 A B356/S034 G [Maybe F] B1578/S234678 S [Negative of b378s24568, which is full of gliders] B0123456/S01235 F [Has a glider] B01278/S012468 S [Biphasic. May be regarded as complex behaviour] B0138/S4 N B0123578/S1236 G [Heavy ameoba dynamics] B3568/S0146 N
The presence of fast decay rules is also very intriguing, since they are usually scattered on the other side of the distribution. One possibility is these F rules are closely related to some more interesting rules and thus inherited some characteristics.
Regarding the usage of neural network, there really can't do any better than human eye on this occasion. The shape of distribution is already recognisable by inspection. But NN might be a good tool when we have more than 3 features.
Re: Use Smoothiness to classify rules
It seems that some files are missing in your package:shouldsee wrote:Sure. It just took some cleaning work. I have updated the link on the frontpage. Although to apply for hexagonal, you would need to undertake a extensive change.
Code: Select all
Undefined function or variable 'update_rule'.
Error in CA_console_opt2 (line 296)
update_rule
There does exist some nearby rules, but I'm not sure if any characteristic is inherited.shouldsee wrote:The presence of fast decay rules is also very intriguing, since they are usually scattered on the other side of the distribution. One possibility is these F rules are closely related to some more interesting rules and thus inherited some characteristics.
Code: Select all
F Rule Nearby N rule
B012/S015 B0125/S015
B34/S368 B34/S0368
B0123456/S01235 B012456/S01235
Still drifting.
Re: Use Smoothiness to classify rules
Thanks for the notice, I forgot to include that in the package. It should be fixed now. I will test them before release next time.Bullet51 wrote: It seems that some files are missing in your package:
Also, if you are fitting a NN, definitly try to include the standard deviations of stats as well as means of stats as your feature, which should help discriminate "stable" complex and "transient“ complex.
Re: Use Smoothiness to classify rules
Dear Shouldsee,
I am puzzled by your definition of forward entropy and backward entropy. How are they different?
And what does the average of the 2-norm of [D_input' V_input(:,2)] mean? Why not just use the mean of the 2-norm of D_input' ?
I am puzzled by your definition of forward entropy and backward entropy. How are they different?
And what does the average of the 2-norm of [D_input' V_input(:,2)] mean? Why not just use the mean of the 2-norm of D_input' ?
Still drifting.
Re: Use Smoothiness to classify rules
They are obsolete statistics that are no longer recorded in the 'process_outputs'.Bullet51 wrote:Dear Shouldsee,
I am puzzled by your definition of forward entropy and backward entropy. How are they different?
And what does the average of the 2-norm of [D_input' V_input(:,2)] mean? Why not just use the mean of the 2-norm of D_input' ?
Code: Select all
dVV=[VV(1:3,:)-VV(4:6,:)];
wt=D_Tinput;
wt=wt/sum(wt);
v=[D_input' V_input(:,2)];
wtdVV=dVV(1:3,:).*repmat(wt,3,1);
v1=mean(dot(v,v,2)); %% this is obsolete
v2=mean(abs(wtdVV(1,:)));
v3=mean(abs(wtdVV(2,:)));
% v4=mean(wtdVV,2)'
%%these are actually recorded.
phaseX=v2;
phaseY=v3;
phaseZ=H_input;
;
Code: Select all
outnumber=8; %% number of scalar outputs in the stat file
scalars={'phaseX','phaseY','phaseZ','H_input','HD_linput'}; %% varaibles recorded during souping---v1 and v are not recorded
Another thing, I have finally figured out how to cluster a correlation matrix properly. The correlation is calculated between counts of a custom neighborhood (the central cell it self ONLY!!). And the cell state is determined by its 6-gen history. Effectively we are taking a 6*1 slice on the space-time dimension.
ignore 'sumcov=..', it's not really sumcov.
- Attachments
-
- B0145678S157_soup61.jpg (223.13 KiB) Viewed 11087 times
-
- B01234568S346_soup50.jpg (212.23 KiB) Viewed 11087 times
-
- B3S23 _soup1.jpg (182.15 KiB) Viewed 11087 times
Last edited by shouldsee on August 23rd, 2016, 8:13 am, edited 1 time in total.
Re: Use Smoothiness to classify rules
I have figured out that H_input is the entropy of {NH now, NH 1 step before}, but what is the meaning of dVV?shouldsee wrote: They are obsolete statistics that are no longer recorded in the 'process_outputs'.Code: Select all
dVV=[VV(1:3,:)-VV(4:6,:)]; wt=D_Tinput; wt=wt/sum(wt); v=[D_input' V_input(:,2)]; wtdVV=dVV(1:3,:).*repmat(wt,3,1); v1=mean(dot(v,v,2)); %% this is obsolete v2=mean(abs(wtdVV(1,:))); v3=mean(abs(wtdVV(2,:))); % v4=mean(wtdVV,2)' %%these are actually recorded. phaseX=v2; phaseY=v3; phaseZ=H_input; ;
Still drifting.
Re: Use Smoothiness to classify rules
Bullet51 wrote:
I have figured out that H_input is the entropy of {NH now, NH 1 step before}, but what is the meaning of dVV?
Code: Select all
HV=sum(VSD_Tinput,1)'; %% supposedly this should be forward entropy
HH=sum(HSD_Tinput,2); %% supposedly backward entropy
V_input=[HD_input' HV+HH HV-HH (1:18)']; %% combining into array
V=V_input(:,[2,3 ,1])'; %% reordering
VV=combvec(V,V); %% enumerate combination of points in the aforementioned 3d space
dVV=[VV(1:3,:)-VV(4:6,:)]; %% calulate vectors in 3d space, each corresponding to an NHT
wt=D_Tinput; %% read the number for each transition of NH (NHT)
wt=wt/sum(wt); %% normalise
wtdVV=dVV(1:3,:).*repmat(wt,3,1); %% weighting
v2=mean(abs(wtdVV(1,:)));
v3=mean(abs(wtdVV(2,:)));
Re: Use Smoothiness to classify rules
Sure. It seems that the 4th column of V_input is never used.shouldsee wrote:Code: Select all
HV=sum(VSD_Tinput,1)'; %% supposedly this should be forward entropy HH=sum(HSD_Tinput,2); %% supposedly backward entropy V_input=[HD_input' HV+HH HV-HH (1:18)']; %% combining into array V=V_input(:,[2,3 ,1])'; %% reordering VV=combvec(V,V); %% enumerate combination of points in the aforementioned 3d space dVV=[VV(1:3,:)-VV(4:6,:)]; %% calulate vectors in 3d space, each corresponding to an NHT wt=D_Tinput; %% read the number for each transition of NH (NHT) wt=wt/sum(wt); %% normalise wtdVV=dVV(1:3,:).*repmat(wt,3,1); %% weighting v2=mean(abs(wtdVV(1,:))); v3=mean(abs(wtdVV(2,:)));
Still drifting.
Re: Use Smoothiness to classify rules
It was a grouping index in rp_visual4.m for visualisation purpose.Bullet51 wrote: Sure. It seems that the 4th column of V_input is never used.