Lenia – Biology of Artificial Life | Hacker News Reader

Lenia – Biology of Artificial Life

arxiv.org

89 pointsp1esk36 comments

Hm, sounds like mathematician giving fancy biological-sounding names to some procedurally generated abstract art forms.

Frankly, I had hoped for more, given the title. There's some really cool work that has been done with digital artificial life that has real biological significance. For example, Richard Lenski's group at Michigan State University developed a platform called AVIDA (https://en.wikipedia.org/wiki/Avida) that has been used for evolutionary biology research. Slightly different, 2012 saw the publication of the first whole-cell computational model (https://www.sciencedirect.com/science/article/pii/S009286741...). And in the past 30 years, computer simulations of ecosystems have slowly been gaining momentum and acceptance as a tool for addressing ecological questions that are very hard to study in real life (this happens to be the field I work in).

Lenia, on the other hand, is nothing more than an abstraction of Conway's Game of Life. A pretty mathematical pastime that is fun to play around with and gives you some interesting patterns - but nothing that merits the description "biology". The whole paper just reeks with misapplied terms: from "taxonomy" to "ecology" and even "physiology". (They don't have anything to do with their real counterparts.) In short, I fail to find any significance for our knowledge of life on earth, or any other life, for that matter.

Pretty graphs, though.

Molecular biology: https://youtu.be/_SGoDIkscXg?t=39

Game of Life: https://youtu.be/xP5-iIeKXE8

Sure, one is 2D and discrete, the other is 3D and continuous... oh wait, Lenia is 2D and continuous. That’s interesting.

    2D and discrete   -> Conway's Game of Life
    2D and continuous -> Lenia
    3D and discrete   -> Minecraft ?
    3D and continuous -> Molecular biology

Sure, being continuous is cool, but your parent is absolutely right that calling anything about this "biology" is way out of line.

My comment had nothing to do with the number of dimensions or whether the world is continuous or discrete. (AVIDA and many other biological models are both 2D and discrete.)

The two core characteristics of life are metabolism and reproduction. Metabolism requires some kind of input, a resource, that is somehow consumed or transformed as the organism grows or is active. Reproduction places an organism in a long line of descent, possibly with mutation and evolution, akin to what we like to call Life.

Sure, living organisms show (symmetrical) organisation - but that is because this morphological organisation enables metabolic functions and reproductions. The morphology does not arise because of some random mathematical rules, but because it fulfills a specific need. It is not an end in itself. This teleological perspective is completely lacking in cellular automata such as Lenia.

(edited for clarity)

Avida is dead, correct? Last release was 5 years ago. I wonder why, especially given that today it could run 100 times faster on gpus...

> Avida is dead, correct?

I'm not sure about that. You're right that there doesn't seem to have been much development going on in the last few years, but a quick Google Scholar search indicates that there are still papers being published with it.

I don't know if performance was ever much of a problem with AVIDA. It's a pretty light model, as far as I recall.

It is an intriguing system. I love the idea behind it and had good fun playing around with it back in secondary school. Also some interesting theoretical work that was done with it, biology-wise. Though of course, you always have to take model results with a big grain of salt (and have a careful think about what can and what can't be applied to Real Life).

I don't know if performance was ever much of a problem with AVIDA

What if interesting things only start happening when you run billions of organisms for billions of generations? Kinda like what's been happening in neural networks field since GPUs enabled training of huge models on huge datasets. Even small scale experiments, if you can run them 100 times faster, you can explore 100 times more configurations.

I also played with Avida a while back. It feels like the key people behind it just lost interest and moved on to other things, and there have been no others determined and talented enough to push it forward.

Youtube video showing some of the persistent life forms: https://www.youtube.com/watch?v=iE46jKYcI4Y

Here is an implementation: https://chakazul.github.io/Lenia/JavaScript/Lenia.html

Here is the github: https://github.com/Chakazul/Lenia

Strange and fascinating. It raises interesting questions about we consider "life". The implementation [0] contains a whole taxonomy tree of variations, of which Conway's Game of Life is but a branch. It's certainly an ecosystem of mathematical "life" forms.

Reminds me of one of my favorite books, Vehicles: Experiments in Synthetic Psychology [0].

[0] https://chakazul.github.io/Lenia/JavaScript/Lenia.html

[1] https://mitpress.mit.edu/books/vehicles

The title reminds me of the work done by Karl Simms and published in SIGGRAPH94

http://www.karlsims.com/papers/alife94.pdf

http://www.karlsims.com/papers/siggraph94.pdf

(video here: https://archive.org/details/sims_evolved_virtual_creatures_1...)

I wonder if cellullar automata could exhibit at least some approximation of classical mechanics, i.e. mass, inertia, forces etc.? From my understanding, motion as seen here is purely due to the structure of the objects.

Yes. See here:

https://journals.aps.org/prl/pdf/10.1103/PhysRevLett.56.1505 Frisch, Hasslacher, Pomeau - Lattice Gas Automata for the Navier-Stokes Equation (1986)

Citations (5k+) here:

https://scholar.google.com/scholar?cites=5993777633272177716

See also my other comment here:

https://news.ycombinator.com/item?id=18757946

Huh, they extend the Game of Life with a complex neighbor kernel, and get more complex patterns emerging. This is a cool experiment, but hardly "biology of artificial life".

A short, sweet summary of my opinion :D

>We describe the methods of constructing and studying Lenia, including its mathematical definition, in silico simulation [...]

It's funny how different fields of science have all invented their own names for "computer program." Physicists and many engineers call them "codes," and biologists love fitting in the buzzword "in silico."

I would guess it's a play on in situ.

Or 'in vitro' ('in glass' ie. 'in a test tube'), as opposed to 'in vivo' ('in life').

Yes, in biology we use in silico to extend this previous distinction between in vivo and in vitro.

In my field (analog hardware for deep learning) "in situ" (e.g. in situ training) means operations are executed in hardware, while "ex-situ" means a simulation of the hardware running the model is done in software (on CPU/GPU).

I honestly don't know if this a simulation of life or an example of it. Or if those are mutually exclusive.

From the video, they seem to be more like stable moving patterns than life. They don't appear to self-reproduce or have behaviours.

That is in the future works section in the pdf.

8. Do self-replicating or pattern-emitting lifeforms exist in Lenia?

It seems kind of a stretch to call it "biology" then.

"I designed a new life form! points to lego brick It exists and you can move it! Future work will involve making it self-replicate and have artificial intelligence."

Figure 20 from the paper is particularly interesting. Life solves multiple problems, and for unicellular life, one of them is the problem of locomotion. At least here Lenia and real world biology seems to have some amount of overlap.

If there’s any overlap at such a simplified model then the simplified model just might be a good place to start figuring out the actual system.

All due respect - I think that for pretty much any simple shape you'll find an actual organism that "looks just like it".

I do tend to agree that unicellular life has a mind boggling diversity of forms and you can find similarities to most things, simple or not-simple :)

That said, humans have a very specific notion of "simple shape", intuitively incorporating symmetry into it. Symmetry is far from guaranteed in mathematics (although mathematicians, being human, also tend to search and study symmetric things). Even in the most symmetric objets - let's say groups, there are many more non-abelian (i.e. not commutative ab != ba) groups than there are commutative groups.

Pet Peeve Rant:

Just like in the case of SmoothLife (which Lenia generalizes), the square grid used for this implies a lack of isotropy(we could also say that it "introduces anisotropy"; but that makes it sounds like a good thing), EVEN IF you turn the entire thing continuous like done here and use . This last fact becomes relatively obvious if you mash the "random" button in the demo, you'll see 'square' noise.

This really ought to use a triangular grid (or the dual of it, a hexagonal one) to ensure isotropy; Frisch, Hasslacher & Pomeau showed this in 1986 while working on two dimensional Navier-Stokes equations:

https://journals.aps.org/prl/pdf/10.1103/PhysRevLett.56.1505 (Ctrl+F isotropy)

Section 4.4.4 of the OP paper DOES acknowledges that one could generalize this further (which includes generalizing to a hexagonal grid, doesn't mention a triangular grid however), but unfortunately it doesn't mention isotropy even once, anywhere, not even in section 3.1.1 on Spatial invariance (albeit, I suppose, the section does semi-address it indirectly, but what's described there to address it seems like a hack, and the 'empirical' evidence from the random button would seem to agree with it.).

Having said all that:

I still consider Lenia DAMN impressive, and I think other people have pointed out this lack of isotropy to the author before. I suspect he'll follow up with more generalizations, but, based on the Github activity (See here: https://github.com/Chakazul/Lenia - last commit in July, long open & relatively trivial pull requests, etc.) and when the OP paper got published to arXiv, I suspect he's been busy writing the paper & with other academic duties and that he could almost certainly use a lot of help with this project. I lack the time (and admittedly, skills) to do so, but I figure if I point this out here, maybe some people might want to pick up on it, especially since, as the author points out, Lenia & further generalizations of it could serve as a great benchmark for various machine learning related matters, which Hacker News users seem to have a great interest in.

Edit: I've decided to us this opportunity to do some renewed literature research:

If you want to read more about Euclidean automatONS beyond the citation #27 from the OP paper, you might want to check here (I suspect the author of the OP paper either overlooked the existence of this book chapter, or hasn't had time to study it yet):

https://link.springer.com/chapter/10.1007/978-1-84996-217-9_... Pivato, Marcus - RealLife (Chapter from "Game of Life Cellular Automata" edited by Andrew Adamatzky)

Which seems like a follow up work to: https://arxiv.org/abs/math/0503504 Pivato, Marcus - RealLife: the continuum limit of Larger Than Life cellular automata

B: There also seems to exist a separate idea ALSO called Euclidean AutomatA, first described here:

https://www.aaai.org/ocs/index.php/SSS/SSS14/paper/viewFile/... Kornai, András - Euclidean Automata

Citations here: https://scholar.google.com/scholar?cites=2906127536931309801

The author of this seems unaware of the works by Pivato, the same applies to this follow up work:

https://ruj.uj.edu.pl/xmlui/bitstream/handle/item/51661/EA01... Gyenis, Zálan - Skeleton in the Euclidean closet

2. More on Triangular & Hexagonal automata:

Research by Carter Bays:

https://link.springer.com/referenceworkentry/10.1007%2F978-1... Cellular Automata in Triangular, Pentagonal, and Hexagonal Tessellations

"First Online: 28 November 2018" - but this is basically a republication of a 2009 version of an expanded & updated version of http://wpmedia.wolfram.com/uploads/sites/13/2018/02/15-3-4.p..., which stems from 2005!

Citations of the 2005 version (doesn't seem directly listed in Google Scholar):

https://scholar.google.com/scholar?q=%22A+Note+on+the+Game+o...

Citations of the 2009 version:

https://scholar.google.com/scholar?cites=1062581570709413178...

And there's also an even older version from 1994, only addressing triangular tesselations:

http://wpmedia.wolfram.com/uploads/sites/13/2018/02/08-2-4.p...

Citations here: https://scholar.google.com/scholar?cites=9274761406110052964

Research by Andrew Wuensche:

http://users.sussex.ac.uk/~andywu/downloads/papers/self_rep.... Self-reproduction by glider collisions: the beehive rule

Citations here:

https://scholar.google.com/scholar?cites=1226081815033511695...

On an off-topic side note: Navigating citation space is an absolute mess, as the above shows. Researchers constantly have to cut corners when it comes to studying the literature because it's too damn hard to navigate, which leads to constant overlooking of convergently evolved research. A Peter Landin quote comes to mind:

"Most papers in computer science describe how their author learned what someone else already knew."

If this bothers you even nearly as much as it bothers me, I suggest you see how you can help the Initiative for Open Citations:

https://i4oc.org/

So tools like http://citationgecko.com/ & http://cluster.cis.drexel.edu/~cchen/citespace/ can actually help put an end to this mess.

Great comments! Thanks for posting. I've written a hex grid automata for wolfram simulations in java, js & dart before.. maybe I can get it contributed to the Lenia guys.

https://github.com/pablo-mayrgundter/freality/tree/master/ph...

TLDR; "Artificial life" is a misnomer for efforts like this which are simply play on cellular autometa type variations. This phrase is extensively used by biologist to constructing actual living physical cells using non-living material. This paper is not about that and I hope folks stop calling this "artificial life".

Important bit:

Life can be defined as the capabilities of self-organizing (morphogenesis), selfregulating (homeostasis), self-directing (motility), self-replicating (reproduction), entropy reduction (metabolism), growth (development), response to stimuli (sensitivity), response to environment (adaptability), and evolving through mutation and selection (evolvability). Lenia, the subject of this paper, is able to achieve many of them, except self-replication that is yet to be discovered.

Above definition wouldn't be agreed upon by many biologist because the entire concept of environment that is as complex as universe is missing. It's easy to create something that can "thrive" in 2D grids with bunch of simple rules but quite another thing that can thrive in a computer called universe which is simulating 10^80 atoms at once where all quantum and space-time rules aren't even known to us.

However you would prefer the term be defined and used, for decades, people have used the term "artificial life" to include cellular automata.

Sources:

https://en.wikipedia.org/wiki/Artificial_life

https://en.wikipedia.org/wiki/History_of_artificial_life

Further, the study of how life arises has to start somewhere very simple, like drastically simple/simplified 2D rule systems.

Too late. The term has been co-opted. There are Artifical Life conferences that have been running for more than 15 years now. http://2018.alife.org/accepted-papers-oral/

Confusingly, biologists use the term "artificial life" to talk about both digital and chemical new-life experiments. Compare Richard Lenskis AVIDA platform: https://en.wikipedia.org/wiki/Avida

Reply on news.ycombinator.com