Something I would love to see in one of these evolving critters projects people work on is this: Food that isn't just worth X amount of energy, but food that has X amount of Stuff A, Y amount of Stuff B, and Z amount of Stuff C... and so on. Then the creature would have organs don't don't require energy to function but require a specific Stuff to work right. The specific Stuff any given organ would need to work right would be based on it's genetics, and what the organs do would again be based on the genetics. Each creature would have also a max amount of stuff it could hold in it's body at any given time. So you could then start having organs that develop that do things like, using a bit of Stuff A as energy convert any Stuff B you have in your body into Stuff C. Or another organ that says, using a bit of Stuff C as energy expel from your body some of the Stuff C you are currently holding. Depending on how it was done you might get complex things emerging like: When I'm ready to breed start expelling Stuff A and B from my body to attract a mate, and look for Stuff C while mating because it just so happens that there is another type of creature that evolved along side me that when it's ready to mate expels C from its body when it is ready to mate. You might also have creatures that develop "toxins" as well. What I mean is that your world might have only Stuff A, B, and C in it naturally but this creature might develop an organ that converts a bit of stuff C into stuff D. Nothing has yet developed an organ though that can do anything with stuff D though, so when some random creature comes over and eats the creature that has built up a bit of stuff D... now this random creature has stuff D filling up it's stomach and has no organ to convert D into a useful Stuff and nor does it have an organ that can expel it, so this random creature effectively now has a smaller stomach for eating the foods it wants to eat.
Great to watch - thank you. Firstly is there a modern update to this anywhere with updates ? I.e. Where we are now. I know this is a very old video so my comments won't really be relevant - but my ideas for development * make maximum life span dependent on size , as well as energy storage and maximum brain development , although have a longer gestation period period for spawning - if it's bigger it should live longer, have more capacity for neural links. This would hopefully generate more complex 'critters'. * being ten years on the sandbox should be able to be a lot bigger with the option to run two or more separate species - evolution may then create a carnivore species and less likely a cannibal. *the only real update to the environment I think that's needed is a day / night cycle - promoting possible dormant night time evolution with food only being spawned during the day. Using this a dark permanent 'cave' environment could be created to see if a subspecies(s) develop in or around that area . (Simple organisms in a half covered Petrie dish).
this guys perfectly adequate to getting the point across. Not to mention that he obviously knows exactly what hes doing and how to do it. He even answers the less important questions with what i would call as ease.
Evolution does favor complexity, but I'm not sure the system described could actually see this. If they made it so that creatures produced waste, it would go a long way towards helping. The drive towards complexity naturally is due to the fact that every living thing makes changes to its environment. For instance, with bacteria, they metabolize something and turn it into a waste product. So even if their food source regenerates, the buildup of waste creates a pressure. It creates a situation that greatly favors the survival of an organism which could thrive both in the world the original bacteria thrives in AND in the waste. Even development of a symbiotic organism which could survive solely from the waste, which might be seen as equivalent in complexity since it's just surviving on something else, would be inferior to that more complex organism because its survival would be restricted by its food source - the original bacteria.
+Dustin Rodriguez Theres obviously a ceiling on the complexity allowed in any system. Its also balanced by the 'most direct route' and 'least energy' thing. This is why when you remove a barrier to continence in an environment any modifications that overcame it revert to the way they were before such a barrier existed. The lose complexity. Interesting to note that GA's have this reversion built in to the function.
Mick Wright You say "there is obviously a ceiling on the complexity allowed in any system." That is not at all obvious to me. What measure of complexity are you supposing? Going by Stephen Wolfram's Principle of Computational Equivalence, there does seem to be a ceiling... but that ceiling is hit almost immediately. Even extremely basic systems with very simple forms of interaction between a small number of types of parts are capable of being Turing Complete, the highest level of complexity we believe possible. I'm not sure what relevance the 'ceiling' has if it means that weather systems, down to the level of quantum particles, are on exactly the same level as 3 bodies in a vacuum interacting under Newtonian gravitation, or a rule 30 1-dimensional cellular automata.
Dustin Rodriguez Well although a complex system especially one like say evolution or climate or weather is non deterministic it is possible to calculate the maximum level of complexity deterministically. Now without getting too complex about it, (sorry couldn't resist that) you could started with a CAS that is rather simple. Say for example a simple cellular automata with just ten squares. Its simple enough. And yes we can calculate how much complexity it possible there. In other words how many combinations or derivations are possible. It'll be a huge number granted. But its a ceiling. So if theres a lowest ceiling... in a really simple system... lets say a cellular automata of two squares by two... and a more complex one of ten by ten... then a billion by a billion has a ceiling too! Its a huge ceiling granted... but its also a limit. So a brain with 90 billion neurons with an average of say 5,000 neural connections... and lets assume any one neuron can connect to any other (which is not the case but hey its a hypothetical) well the ceiling is high...but there is a ceiling.
I understand now. I'm not sure that waste is relevant to this presentation beyond being a single additional factor (pressure) that affects complexity. Energy usage of neuron activation is already the mechanic that enforces the ceiling. He's saying that evolution (trait passing with minor mutations) doesn't drive complexity, but natural selection absolutely caps it.
"Natural selection" is used as an analogue. It's not really natural, but it is selection. You can't have any kind of selection (or evolution) without a cost or fitness function. In nature there are sets of probabilities that act as the "fitness function" although not explicitly stated as a "function"
evolution : natural selection (of groups) : 1) groups with altruistic members vs. 2) groups with no altriustic members... obvisously groups with altruistic members has certain advantages in times of extreme external conditions such as wars or famine. Thanks for the thought provoking video, it has helped me understand myself and my position in the world.
The question of the complexity of the environment limiting the development was interesting as was that point that a designer could implement a constructed agent that could easily out compete the evolved entities. Perhaps the element that should increase environmental complexity should be a constructed agent that scales forcing evolution to keep up with it. In essence that is the true goal, not to just evolve a moderately efficient system but to evolve a system that could not be designed.
One of the variables for the environment could possibly be variations in temperature in different region of the environment which could lead to different adaptations.
+qed qubit Yeah. just look up some stuff. Or look up "Species-game". It's a great alpha game that is in development, and I think it simulates brain evolution along with actual evolution.
Hello everybody, Could anyone pleasee explain to me how the connection matrix ( 38:30 ) works? I have a simple program with similar neural networks how do i build such a connection matrix for them
Yes, if you allow interaction between agents, agents become a part of the environment (i.e, coevolution). My understanding of polyworld is that agents *are* allowed to interact.
Game designers should be using genetic programming for artificial intelligence. Imagine a game where the enemies you face are always learning which techniques are more effective against you?! It would lend to a very challenging game. I'm sure the enemies would learn to do things that the game designers never even intended.
Nuclear decay does not have to be cause-effect induced. A particular atom may decay now or wait 14 billion years. Arguably perhaps we don't know how to define the atom well enough to prove rigorously that it is also cause-effect grounded. There are other examples along these lines that involve "quantum decisions".
Interesting. My hunch is that the relationships are there, but it's a matter of scale. We don't know enough about really small things or really big things to see the realtionships clearly. I seem to remember reading about the use of hamiltonians to achieve some insight into quantum systems. In any case, I like the comment. I give it six points.
You're talking about something different. I'm not talking about AI *learning* what you do, and then trying to counter it, I'm talking about AI that uses a randomly mutating genetic algorithm for behavior and then a fitness function to naturally select more successful behavior. This is a significantly different model for AI, and you are right, it is entirely possible, and maybe someone has tried it already.
evolution is an algorithm: given only variable population, selection, reproduction with occasional errors, you get evolution. the n-armed bandit problem in machine learning is essentially the same idea. You tolerate errors for exploration, and keep the same ideas for exploitation.
if being prone to self sacriface is a genetic weakness, then the term of genetic weakness needs to be examine, what one might term "defective" others might see it as the opposite
Brain and genes together search for optimal solution. While brain costs energy to run, the genes are cheaper from energy point of view. So it makes that genes would become more complex and brains less complex.
there's an interesting discussion on usenet about the morality of evolving artificial intelligence, destroying "unfit" intelligences, putting them through harsh testing processes, etc. it also asks what the finished, intelligent product of this process would think about its origins, and how it might respond. The discussion starts with Johan Larson's post at groups google com/group/rec.arts.sf.written/browse_thread/thread/db51c30169c7bc58/66d85b18a3235b10?hl=en
True, I was probably a bit harsh with my comment. I think it's really important for academics to be trained in public speaking. From my experience they are, and I know what it's like talking in front of groups with last minute preparations of material. I wouldn't quite say he was an amateur. Given that he's a grad student that would suggest that he's seriously considering a career in this kind of stuff. I think in time he'll present his ideas well and should become a lot more interesting!
Well, I've seen numerous sample codes and projects that specifically "wire up" neural nets based on evolutionary environments since 15-20 years ago. Just saying from experience.
i think defenition of "alive" is being self aware. knowing that you are you. i know that doesnt really apply to plants but if anything artificial was to be considered alive it would have to be selfaware.
Having reviewed the literature, neuroevolution was born in the early 1990's for simple optimization problems like pole balancing. Karl Sims made a huge breakthrough with his work in 1994, but his experiments didn't allow for much coevolution between agents (other than simple 1v1 competitions). Polyworld extended this by allowing coevolution between multiple agents, which is huge for evolution of species diversity (see Cambrian explosion). Sadly, seems Polyworld kinda died out for now.
Awesome vid. I'd assume that Virgil G. checks out the comments, so I just wanted to say good job. It's easy to tell that you love the subject, and the love is infectious. The presentation kept me interested all the way through.
I would add some blocks of different kinds to use as tools. The speaker at the end had very salient points regarding specieation. Overall an excellent talk. Hats off applause. :-)
a little confused about the experiment where he turns off evolution - aren't the genes that are randomly used still a result of selection? That is, you can only randomly select from creatures still alive - thus selected. There may be a scale difference in complexity but it's still evolution. Perhaps simply cloning with no cross-over or mutations is evolutionless?
Sure, the whole number eigen values of the hamiltonian express the basic states of quanta in 3d1t space and it builds up from there. Granted I don't have a DS, but I would still say we have a pretty good understanding of the math involved at that level. Human ai isn't special, especially if you watch Jerry Springer.
Looking through the comments it becomes immediately clear to me that the commenters here doesn't know about aspergers or autism. I believe the reason he is jittery and seem so stressed is because that's just how he is in his head. When he is explaining something his brain is going in overdrive so to speak and every time he says "so" it's to summarize in his head, giving a step by step breakdown of something the only way he knows how. The time is too short and he is too smart to explain it well.
Is shooting for 'neural' and high brain functions that humans might have the way to go? Is there any research done on studying the brains of early fish or the insect brain?
Genetic algorithms allow for solutions that a human could not/ would not logically conceive. because of the random nature of the algorithm. For example If a human was only capable of placing a petri dish the likelihood would be that a patch of some bacteria/life-form would develop on the dish through the random chance of a cell landing on the dish. Penicillin was developed accidentally. This might be a horrible example now I look back, if anyone can give a better one please do.
Yes. Even if it wasn't smart enough to determine which exact solution was the most efficient/advantageous, by brute forcing random mutations and micro adjustments to its own code it would eventually become excellent at editing itself creating an exponential growth in efficiency, APS, and intelligence. With these methods and more we've already utilized a program which analyzes everything digital that humans create. It helped us determine Bin Laden's location look it up. But has this program already created it's road map for our future?
I have built my own simulation with much more complex adaptive nueral networks, and I would like to visualize the brains the way he visualized them using the connection matrix at 38:10. an explanation on how this type of visualization works would be highly appriciated
need more variables in the program if you want inteligent creatures to evolve you need food to come about in a non random way. Maybe come about more like a plant which leads to food growing in cycles, spreading, evolving as well as being able to be over eaten lowering the availablilty of plant food. this would lead to inteligence for finding food and a greater niche for hunters. random food just leads to random foraging
Maybe if they added an output neuron that gave signal to an additional input neuron, the organisms could have a short pseudo-memory. I know nothing of these systems, just a thought.
I believe the reason the neural complexity plateaus is because the question the environment places is still lacking a key part. While a variety of factors are addressed, the fact remains this is a 2D world. Apparently, a 2D world is incapable of creating intelligence. A 3D world with the ability for different bodies instead of the generalized 'strength' gene is necessary. This allows a much more complex environment... promoting a complex intelligence.
The way to view 2) is to realize that it's genes, not organisms that are the unit of replication. For my genes, it will make sense to inject me with a suicidal impulse to sacrifice myself for X close family members. Where X varies depending on how closely related they are to me. The tradeoff point is where it becomes likely that my sacrifice will cause more copies of my own genes to survive by dying than by living. So, it's not really altruism - it's gene-level selfishness.
It is not like I want to care, I just can't help it. I agree that I need to work this out, but it is not easy to care only about the content of what people say. If you can do this then congratulations, you will become a very good engineer.
Actually, this is pure natural selection, not evolution and adaptation depends on the living or dying of the polys. Among it's not intelligent, it could become part of a more complex simulation for an IA, as an ideas test for giving an absolute answer or idea. But for that there should be more things thought... just gessing here. (Love this stuff)
Noticed those horsetail tablets or was it a plastic bottle of some sort of smart drug (?) on top of the podium stand as the speaker nervously twisted the cap of his ice tea bottle then ironically plunked the audience into a segment on nervous systems.
Di...... Di... Digimon.... Digimon.... Di...... Di... Digimon.... Digimon.... Digimon, Digital Monsters, Digimon are the CHAMPIONS! That kind of reminds me of Digimon Tamers, in which a group of college students created artificial life in their computer with the basic function of survival, similar to this, and many years later Digimon became a big thing, and then came to life. Who knows, maybe with the Higgs Boson something crazy could happen. Oh the day that humanity meets its end. ^-^
@dadalinN In reality, complexity increases (further, it may stop when balance is reached), so sg is wroing with this model. I think in reality the growth of complexity is driven by the need for more effective usage of the environment's energy. And, as these beings evolve, they stay on species, they can still mate with eachother. If this would change, I think there would be motive to further improve complexity.
If it is possible to re program Polyworld with gpu cycles instead of cpu cycles the evolution can go a lot faster because gpu are a lot faster in linear processing than cpu are. End when you look at rainbow calculations a gpu is really 200 times faster than a cpu: a passport like "23P#d3" can be hacked in 7 minutes with gpu processing and a cpu needs 23 hours to do the same linear calculation. I hope someone who can code gpu programs wants to look at it. Greets Pieter Schreurs
these things should have family trees. some traits skip a generation. some things are too closely related to breed. sometimes a defect can change up the game. also family probably cultivates familiarity. if the blinking function had a code that could be inherited and particular families blinked certain codes, like birds sing certain songs, then they might start to use that thing. I wouldn't only take the genes of the parents. I wish there was a family tree aspect to it. probably too late >.>
valken666...definately..too many of these topics have people who say "uh" "um" a lot, and go off on tangeants, explaining in strict specifics with general concepts without diving into actual processes.
the project models live very well and it needs to model of geographical isolation better ; humans initially elvoved in isolated parts of the world and have different ethnicity, later as technology improves, humans mixes their genes inter-racially, producing more robust offsprings in most cases, in the model, the critters are simply assigned a colour, neglicting all the various attributes of racial type.
It's been argued that there are no altruistic members of any group really. Those that share surplus with those that are in shortage do it expecting return when they are in need and those they shared with were in surplus. Purely altruistic members of a society would simply die during a famine since they would share even when they are in need and there is no penalty when those that take charity do not return it later on.
Why should defective genes be entirely purged? It may not be a single gene, but a group of genes that makes a defective behavior. Anyhow, why does Muscular Sclerosis exist, instead of purged? TRUE altruism is extremely rare. You also need to look at it from the point of view is that you're a robot designed to replicate genes that cause you to replicate those genes. Everything else is just built around that, some genes can promote replication in SOME environments, others hinder it.
I challenge the claim that Polyworld is pure natural selection with no fitness function. Virgil mentions that there's a cost function applied to the networks. Obviously, if there is no cost function, then the networks will be filled with a lot of neurons that do nothing, as long as they are being "carried" by a useful neuron. The problem is the cost function is arbitrary. Unless the cost function is subject to natural selection, this is not purely natural. It is essentially a fitness function.
This presentation should have subtitles in English. Did they get this presenter from a random person on a city bus? "...so like zinna sints heh, i'm gonna put this lay mit sonna mits here..."
While interesting, it will not be useful until you can "harvest" parts of the neural network. The visual subsystem would be useful, if it was 3D. However the self preservation that will evolve will just get in the way (I mean in the short term, way before the system rises up in revolt)
If somebody is interested in testing out a neuro-evolutionary algorithm: I created a game where you can train your own drivers (their neural networks) by a genetic algorithm, it is still in beta, but you can have a look at www.kongregate.com/games/sem23/neurodriver to play the game.
+Peter Rudolph Training drivers to drive isn't really that impressive though. Personally, I'm convinced that you actually used neural networks in the way described in accepted scholarly articles, but what if skeptical people think that you made neural networks that don't evolve efficiently when faced with difficult tasks or that you just made an algorithm for generating pre-recorded sequences of inputs and programmed a simple system for getting back on track, then made a variable for how precisely the cars follow the sequence that increases with training? For the driving task, if the cars don't collide with each other, they might just settle for an algorithm that's nearly perfect and stop evolving significantly, but if they can collide, then a car could get unlucky even though it's doing the right thing because other cars just happen to want to run into it, and balancing out luck with many tries will take tons of CPU again. In the swimming test in this video, cross-breeding would likely mess stuff up so it'll be many tries before there's successful crossbreeds, and for the cube-grabbing demonstration, although it's 2-player and zero-sum, the two sides have to determine their shapes and settings without knowing anything, so it's definitely not a perfect information game. I can provide you with criticism, but I can't think of a simple way for you to do better show off the potential of your system.
39:30
I like how quickly he went from nervous to annoyed.
39:51
and then annoying
Something I would love to see in one of these evolving critters projects people work on is this:
Food that isn't just worth X amount of energy, but food that has X amount of Stuff A, Y amount of Stuff B, and Z amount of Stuff C... and so on. Then the creature would have organs don't don't require energy to function but require a specific Stuff to work right. The specific Stuff any given organ would need to work right would be based on it's genetics, and what the organs do would again be based on the genetics. Each creature would have also a max amount of stuff it could hold in it's body at any given time.
So you could then start having organs that develop that do things like, using a bit of Stuff A as energy convert any Stuff B you have in your body into Stuff C. Or another organ that says, using a bit of Stuff C as energy expel from your body some of the Stuff C you are currently holding.
Depending on how it was done you might get complex things emerging like: When I'm ready to breed start expelling Stuff A and B from my body to attract a mate, and look for Stuff C while mating because it just so happens that there is another type of creature that evolved along side me that when it's ready to mate expels C from its body when it is ready to mate.
You might also have creatures that develop "toxins" as well. What I mean is that your world might have only Stuff A, B, and C in it naturally but this creature might develop an organ that converts a bit of stuff C into stuff D. Nothing has yet developed an organ though that can do anything with stuff D though, so when some random creature comes over and eats the creature that has built up a bit of stuff D... now this random creature has stuff D filling up it's stomach and has no organ to convert D into a useful Stuff and nor does it have an organ that can expel it, so this random creature effectively now has a smaller stomach for eating the foods it wants to eat.
Great to watch - thank you.
Firstly is there a modern update to this anywhere with updates ? I.e. Where we are now.
I know this is a very old video so my comments won't really be relevant - but my ideas for development
* make maximum life span dependent on size , as well as energy storage and maximum brain development , although have a longer gestation period period for spawning - if it's bigger it should live longer, have more capacity for neural links. This would hopefully generate more complex 'critters'.
* being ten years on the sandbox should be able to be a lot bigger with the option to run two or more separate species - evolution may then create a carnivore species and less likely a cannibal.
*the only real update to the environment I think that's needed is a day / night cycle - promoting possible dormant night time evolution with food only being spawned during the day. Using this a dark permanent 'cave' environment could be created to see if a subspecies(s) develop in or around that area . (Simple organisms in a half covered Petrie dish).
24:40
They like the random, because it breaks behavior patterns which could be used by another creature against them. It also helps in exploration.
this guys perfectly adequate to getting the point across. Not to mention that he obviously knows exactly what hes doing and how to do it. He even answers the less important questions with what i would call as ease.
This is one of the most understandable AI presentations I've ever seen.
Evolution does favor complexity, but I'm not sure the system described could actually see this. If they made it so that creatures produced waste, it would go a long way towards helping. The drive towards complexity naturally is due to the fact that every living thing makes changes to its environment. For instance, with bacteria, they metabolize something and turn it into a waste product. So even if their food source regenerates, the buildup of waste creates a pressure. It creates a situation that greatly favors the survival of an organism which could thrive both in the world the original bacteria thrives in AND in the waste. Even development of a symbiotic organism which could survive solely from the waste, which might be seen as equivalent in complexity since it's just surviving on something else, would be inferior to that more complex organism because its survival would be restricted by its food source - the original bacteria.
+Dustin Rodriguez Theres obviously a ceiling on the complexity allowed in any system. Its also balanced by the 'most direct route' and 'least energy' thing.
This is why when you remove a barrier to continence in an environment any modifications that overcame it revert to the way they were before such a barrier existed. The lose complexity.
Interesting to note that GA's have this reversion built in to the function.
Mick Wright You say "there is obviously a ceiling on the complexity allowed in any system." That is not at all obvious to me. What measure of complexity are you supposing? Going by Stephen Wolfram's Principle of Computational Equivalence, there does seem to be a ceiling... but that ceiling is hit almost immediately. Even extremely basic systems with very simple forms of interaction between a small number of types of parts are capable of being Turing Complete, the highest level of complexity we believe possible. I'm not sure what relevance the 'ceiling' has if it means that weather systems, down to the level of quantum particles, are on exactly the same level as 3 bodies in a vacuum interacting under Newtonian gravitation, or a rule 30 1-dimensional cellular automata.
Dustin Rodriguez Well although a complex system especially one like say evolution or climate or weather is non deterministic it is possible to calculate the maximum level of complexity deterministically.
Now without getting too complex about it, (sorry couldn't resist that) you could started with a CAS that is rather simple. Say for example a simple cellular automata with just ten squares. Its simple enough. And yes we can calculate how much complexity it possible there. In other words how many combinations or derivations are possible. It'll be a huge number granted. But its a ceiling.
So if theres a lowest ceiling... in a really simple system... lets say a cellular automata of two squares by two... and a more complex one of ten by ten... then a billion by a billion has a ceiling too! Its a huge ceiling granted... but its also a limit.
So a brain with 90 billion neurons with an average of say 5,000 neural connections... and lets assume any one neuron can connect to any other (which is not the case but hey its a hypothetical) well the ceiling is high...but there is a ceiling.
You did see how he showed the graphs of complexity then the graph of genetic complexity right?
I understand now. I'm not sure that waste is relevant to this presentation beyond being a single additional factor (pressure) that affects complexity. Energy usage of neuron activation is already the mechanic that enforces the ceiling. He's saying that evolution (trait passing with minor mutations) doesn't drive complexity, but natural selection absolutely caps it.
He always makes us surprised and excited. Looking forward to seeing what you will do next, Virgil.
"Natural selection" is used as an analogue. It's not really natural, but it is selection. You can't have any kind of selection (or evolution) without a cost or fitness function. In nature there are sets of probabilities that act as the "fitness function" although not explicitly stated as a "function"
evolution : natural selection (of groups) :
1) groups with altruistic members vs. 2) groups with no altriustic members...
obvisously groups with altruistic members has certain advantages in times of extreme external conditions such as wars or famine.
Thanks for the thought provoking video, it has helped me understand myself and my position in the world.
The question of the complexity of the environment limiting the development was interesting as was that point that a designer could implement a constructed agent that could easily out compete the evolved entities. Perhaps the element that should increase environmental complexity should be a constructed agent that scales forcing evolution to keep up with it. In essence that is the true goal, not to just evolve a moderately efficient system but to evolve a system that could not be designed.
Agreed... his delivery was very jittery and not flowing... which leads to his speech requiring excessive concentration.
Amazing talk. I love the way he articulates his ideas, much better than most "good" speakers imo.
One of the variables for the environment could possibly be variations in temperature in different region of the environment which could lead to different adaptations.
Hes so smart in his field already I don't expect him to be a great public speaker also
"Our speaker is Virgil Griffith. He's talking about the theory of relativity."
True. Biological brains (even the smallest) are still far ahead of anything man has yet devised.
it's 2015 now, has anybody seen something new like this ?
+qed qubit Yeah. just look up some stuff. Or look up "Species-game". It's a great alpha game that is in development, and I think it simulates brain evolution along with actual evolution.
+Yawning Gull thanks for that comment
We are older and a bit similar :-)
Hello everybody, Could anyone pleasee explain to me how the connection matrix ( 38:30 ) works? I have a simple program with similar neural networks how do i build such a connection matrix for them
Not everyone was born with excellent public speaking skills. If you grew up with him, like I did, you would see he is doing a very good job.
Yes, if you allow interaction between agents, agents become a part of the environment (i.e, coevolution). My understanding of polyworld is that agents *are* allowed to interact.
Props to Virgil Griffith, the presenter. He also created the WikiScanner. Stay awesome, dude. :)
I sure do wish that all of us could have a copy of this program so that we could see what we could do with the program...
Game designers should be using genetic programming for artificial intelligence. Imagine a game where the enemies you face are always learning which techniques are more effective against you?! It would lend to a very challenging game. I'm sure the enemies would learn to do things that the game designers never even intended.
Why would they bother making this in 3d? it would be much easier to understand if it was 2d
Nuclear decay does not have to be cause-effect induced. A particular atom may decay now or wait 14 billion years.
Arguably perhaps we don't know how to define the atom well enough to prove rigorously that it is also cause-effect grounded.
There are other examples along these lines that involve "quantum decisions".
Interesting. My hunch is that the relationships are there, but it's a matter of scale. We don't know enough about really small things or really big things to see the realtionships clearly. I seem to remember reading about the use of hamiltonians to achieve some insight into quantum systems. In any case, I like the comment. I give it six points.
You're talking about something different.
I'm not talking about AI *learning* what you do, and then trying to counter it, I'm talking about AI that uses a randomly mutating genetic algorithm for behavior and then a fitness function to naturally select more successful behavior. This is a significantly different model for AI, and you are right, it is entirely possible, and maybe someone has tried it already.
evolution is an algorithm: given only variable population, selection, reproduction with occasional errors, you get evolution. the n-armed bandit problem in machine learning is essentially the same idea. You tolerate errors for exploration, and keep the same ideas for exploitation.
It has been 4 years. 2011.
What are the most impressive achievements of this brilliant program?
is there a new software?
yeah, it would be nice if they had a video form of abstract or a 10 minute compressed form of what they want to say.
Watching this while letting my virtual creatures evolve :) They are way more simple though. Great talk! This is just perfect!
Watching virtual creatures evolve is better than porn.
@Niskinatorn, what little creatures? you have your own program? can you share it?
Shared to Facebook & Twitter : : THANKS GoogleTech
if being prone to self sacriface is a genetic weakness, then the term of genetic weakness needs to be examine, what one might term "defective" others might see it as the opposite
Brain and genes together search for optimal solution. While brain costs energy to run, the genes are cheaper from energy point of view. So it makes that genes would become more complex and brains less complex.
there's an interesting discussion on usenet about the morality of evolving artificial intelligence, destroying "unfit" intelligences, putting them through harsh testing processes, etc. it also asks what the finished, intelligent product of this process would think about its origins, and how it might respond.
The discussion starts with Johan Larson's post at groups google com/group/rec.arts.sf.written/browse_thread/thread/db51c30169c7bc58/66d85b18a3235b10?hl=en
True, I was probably a bit harsh with my comment. I think it's really important for academics to be trained in public speaking. From my experience they are, and I know what it's like talking in front of groups with last minute preparations of material.
I wouldn't quite say he was an amateur. Given that he's a grad student that would suggest that he's seriously considering a career in this kind of stuff. I think in time he'll present his ideas well and should become a lot more interesting!
on that last question i would like to see what would happen if the max life was lower than the average and see wat effects that has on the creatures
Interesting thesis. It will be interesting to see how it comes to fruition.
What a time to be alive!
Well, I've seen numerous sample codes and projects that specifically "wire up" neural nets based on evolutionary environments since 15-20 years ago. Just saying from experience.
i think defenition of "alive" is being self aware. knowing that you are you. i know that doesnt really apply to plants but if anything artificial was to be considered alive it would have to be selfaware.
I got a question, if matting actually takes energy, what do they gain for doing so, or did you placed some kind of instinct?
Having reviewed the literature, neuroevolution was born in the early 1990's for simple optimization problems like pole balancing. Karl Sims made a huge breakthrough with his work in 1994, but his experiments didn't allow for much coevolution between agents (other than simple 1v1 competitions). Polyworld extended this by allowing coevolution between multiple agents, which is huge for evolution of species diversity (see Cambrian explosion). Sadly, seems Polyworld kinda died out for now.
Awesome vid. I'd assume that Virgil G. checks out the comments, so I just wanted to say good job. It's easy to tell that you love the subject, and the love is infectious. The presentation kept me interested all the way through.
I would add some blocks of different kinds to use as tools. The speaker at the end had very salient points regarding specieation. Overall an excellent talk. Hats off applause. :-)
I didn't CZcams had videos longer than 10 minutes... When did this upgrade happen?
a little confused about the experiment where he turns off evolution - aren't the genes that are randomly used still a result of selection? That is, you can only randomly select from creatures still alive - thus selected. There may be a scale difference in complexity but it's still evolution. Perhaps simply cloning with no cross-over or mutations is evolutionless?
The presenter is also on King of the Nerds season 1...
Sure, the whole number eigen values of the hamiltonian express the basic states of quanta in 3d1t space and it builds up from there. Granted I don't have a DS, but I would still say we have a pretty good understanding of the math involved at that level.
Human ai isn't special, especially if you watch Jerry Springer.
guess this guy is thinking too fast for his mouth to follow.genius
Looking through the comments it becomes immediately clear to me that the commenters here doesn't know about aspergers or autism.
I believe the reason he is jittery and seem so stressed is because that's just how he is in his head. When he is explaining something his brain is going in overdrive so to speak and every time he says "so" it's to summarize in his head, giving a step by step breakdown of something the only way he knows how.
The time is too short and he is too smart to explain it well.
Is shooting for 'neural' and high brain functions that humans might have the way to go? Is there any research done on studying the brains of early fish or the insect brain?
The current version does not appear to work on Snow Leopard. It loads and immediately crashes. I'll try the linux version later.
Can somebody explain how the artificial neural network is programmed? Why are the neurons put into groups if they can connect however they want?
The smartest AI would be the one that nobody had any idea existed or operated.
Genetic algorithms allow for solutions that a human could not/ would not logically conceive. because of the random nature of the algorithm. For example If a human was only capable of placing a petri dish the likelihood would be that a patch of some bacteria/life-form would develop on the dish through the random chance of a cell landing on the dish. Penicillin was developed accidentally. This might be a horrible example now I look back, if anyone can give a better one please do.
Yes. Even if it wasn't smart enough to determine which exact solution was the most efficient/advantageous, by brute forcing random mutations and micro adjustments to its own code it would eventually become excellent at editing itself creating an exponential growth in efficiency, APS, and intelligence. With these methods and more we've already utilized a program which analyzes everything digital that humans create. It helped us determine Bin Laden's location look it up. But has this program already created it's road map for our future?
the kaiser soze of A.I.
I have built my own simulation with much more complex adaptive nueral networks, and I would like to visualize the brains the way he visualized them using the connection matrix at 38:10. an explanation on how this type of visualization works would be highly appriciated
need more variables in the program if you want inteligent creatures to evolve you need food to come about in a non random way. Maybe come about more like a plant which leads to food growing in cycles, spreading, evolving as well as being able to be over eaten lowering the availablilty of plant food. this would lead to inteligence for finding food and a greater niche for hunters. random food just leads to random foraging
Maybe if they added an output neuron that gave signal to an additional input neuron, the organisms could have a short pseudo-memory. I know nothing of these systems, just a thought.
My favorite part of the video is the title.
I believe the reason the neural complexity plateaus is because the question the environment places is still lacking a key part. While a variety of factors are addressed, the fact remains this is a 2D world. Apparently, a 2D world is incapable of creating intelligence. A 3D world with the ability for different bodies instead of the generalized 'strength' gene is necessary. This allows a much more complex environment... promoting a complex intelligence.
The way to view 2) is to realize that it's genes, not organisms that are the unit of replication.
For my genes, it will make sense to inject me with a suicidal impulse to sacrifice myself for X close family members. Where X varies depending on how closely related they are to me. The tradeoff point is where it becomes likely that my sacrifice will cause more copies of my own genes to survive by dying than by living.
So, it's not really altruism - it's gene-level selfishness.
Sounds like something worth pursuing as part of a graduate program! Hop to!
It is not like I want to care, I just can't help it. I agree that I need to work this out, but it is not easy to care only about the content of what people say. If you can do this then congratulations, you will become a very good engineer.
Actually, this is pure natural selection, not evolution and adaptation depends on the living or dying of the polys. Among it's not intelligent, it could become part of a more complex simulation for an IA, as an ideas test for giving an absolute answer or idea. But for that there should be more things thought...
just gessing here. (Love this stuff)
The beginning of the video should be on a Mac commercial.
So Google Techtalks are on CZcams exclusively now? Wow, that really sucks! No more downloading to my iPod, too :-(
Noticed those horsetail tablets or was it a plastic bottle of some sort of smart drug (?) on top of the podium stand as the speaker nervously twisted the cap of his ice tea bottle then ironically plunked the audience into a segment on nervous systems.
Di...... Di... Digimon.... Digimon....
Di...... Di... Digimon.... Digimon....
Digimon, Digital Monsters, Digimon are the CHAMPIONS!
That kind of reminds me of Digimon Tamers, in which a group of college students created artificial life in their computer with the basic function of survival, similar to this, and many years later Digimon became a big thing, and then came to life. Who knows, maybe with the Higgs Boson something crazy could happen. Oh the day that humanity meets its end. ^-^
Dang!
This guy talks faster than Chip & Dale!
I need to take notes!
did he say that? perhaps he meant once a nervous system evolved? But i guess how would he prove that... still interesting stuff...
I imagine this kid saying (As Daniel Tosh once said):
“Because I'm not in the 7th grade and I've got things to do."
@dadalinN In reality, complexity increases (further, it may stop when balance is reached), so sg is wroing with this model. I think in reality the growth of complexity is driven by the need for more effective usage of the environment's energy. And, as these beings evolve, they stay on species, they can still mate with eachother. If this would change, I think there would be motive to further improve complexity.
If it is possible to re program Polyworld with gpu cycles instead of cpu cycles the evolution can go a lot faster because gpu are a lot faster in linear processing than cpu are. End when you look at rainbow calculations a gpu is really 200 times faster than a cpu: a passport like "23P#d3" can be hacked in 7 minutes with gpu processing and a cpu needs 23 hours to do the same linear calculation. I hope someone who can code gpu programs wants to look at it. Greets Pieter Schreurs
This is great stuff, though. Very intriguing. Too bad I use Windows.
They aren't going to use the light without adaptive behavior from practical cause and effect analysis.
B+ concept
B+ application
Should be categorized as Entertainment.
these things should have family trees. some traits skip a generation. some things are too closely related to breed. sometimes a defect can change up the game. also family probably cultivates familiarity. if the blinking function had a code that could be inherited and particular families blinked certain codes, like birds sing certain songs, then they might start to use that thing. I wouldn't only take the genes of the parents. I wish there was a family tree aspect to it. probably too late >.>
@rainingblood92 nax linux is better for simple coding because it has open code or somthing like that, not sure if you ment that
valken666...definately..too many of these topics have people who say "uh" "um" a lot, and go off on tangeants, explaining in strict specifics with general concepts without diving into actual processes.
the project models live very well and it needs to model of geographical isolation better ; humans initially elvoved in isolated parts of the world and have different ethnicity, later as technology improves, humans mixes their genes inter-racially, producing more robust offsprings in most cases, in the model, the critters are simply assigned a colour, neglicting all the various attributes of racial type.
Is there somewhere i can download this for windows?
WHOA! 1 hour long!
It's been argued that there are no altruistic members of any group really.
Those that share surplus with those that are in shortage do it expecting return when they are in need and those they shared with were in surplus.
Purely altruistic members of a society would simply die during a famine since they would share even when they are in need and there is no penalty when those that take charity do not return it later on.
Why should defective genes be entirely purged? It may not be a single gene, but a group of genes that makes a defective behavior.
Anyhow, why does Muscular Sclerosis exist, instead of purged? TRUE altruism is extremely rare.
You also need to look at it from the point of view is that you're a robot designed to replicate genes that cause you to replicate those genes. Everything else is just built around that, some genes can promote replication in SOME environments, others hinder it.
I challenge the claim that Polyworld is pure natural selection with no fitness function.
Virgil mentions that there's a cost function applied to the networks. Obviously, if there is no cost function, then the networks will be filled with a lot of neurons that do nothing, as long as they are being "carried" by a useful neuron.
The problem is the cost function is arbitrary. Unless the cost function is subject to natural selection, this is not purely natural. It is essentially a fitness function.
Fascinating stuff. Nice work.
This presentation should have subtitles in English. Did they get this presenter from a random person on a city bus? "...so like zinna sints heh, i'm gonna put this lay mit sonna mits here..."
beanblossom . in . us / larryy / PolyWorld . html
It's also on sourceforge...
How can anyone not see this video is a spoof after the announcer pronounces "polyworld" like Will Ferrell is doing a Bush junior impression?
Finally someone figured it out! Been looking at ALife for creating AI for a long time.
While interesting, it will not be useful until you can "harvest" parts of the neural network. The visual subsystem would be useful, if it was 3D. However the self preservation that will evolve will just get in the way (I mean in the short term, way before the system rises up in revolt)
If somebody is interested in testing out a neuro-evolutionary algorithm: I created a game where you can train your own drivers (their neural networks) by a genetic algorithm, it is still in beta, but you can have a look at www.kongregate.com/games/sem23/neurodriver to play the game.
+Peter Rudolph great man
+Peter Rudolph Training drivers to drive isn't really that impressive though. Personally, I'm convinced that you actually used neural networks in the way described in accepted scholarly articles, but what if skeptical people think that you made neural networks that don't evolve efficiently when faced with difficult tasks or that you just made an algorithm for generating pre-recorded sequences of inputs and programmed a simple system for getting back on track, then made a variable for how precisely the cars follow the sequence that increases with training? For the driving task, if the cars don't collide with each other, they might just settle for an algorithm that's nearly perfect and stop evolving significantly, but if they can collide, then a car could get unlucky even though it's doing the right thing because other cars just happen to want to run into it, and balancing out luck with many tries will take tons of CPU again. In the swimming test in this video, cross-breeding would likely mess stuff up so it'll be many tries before there's successful crossbreeds, and for the cube-grabbing demonstration, although it's 2-player and zero-sum, the two sides have to determine their shapes and settings without knowing anything, so it's definitely not a perfect information game. I can provide you with criticism, but I can't think of a simple way for you to do better show off the potential of your system.
How cool! I saw in the tutorial section you'd put it on Android, so I just got it am looking forward to checking it out.
His minds thinking faster then his mouth can keep up with obviously a genius lol.
wheeeew watched it whole, it was well worth it
incredibly interesting... 5 stars :)
Very good!!! Carl Sims best student!
It would only be a matter of time before these AI programs reach Singularity and conquer the world!
Awesome idea, awesome research, awesome relaxed presentation.
is that first video anywhere else on youtube?
?