Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[PTTG??]

There will certainly be the potential for two creatures to fight, but that will be far from the point. If anything, it might represent a failure state. In this way it's like letting elf traders walk away unharmed- yes, it's possible, but...

[Puzzlemaker]

well I just want him to track energy, so I can breed creatures that eat the others.  Om Nom Nom.

[Neonivek]

There will certainly be the potential for two creatures to fight, but that will be far from the point. If anything, it might represent a failure state. In this way it's like letting elf traders walk away unharmed- yes, it's possible, but...

Universal peace huh?

Grendels attack!

[PTTG??]

Hi everyone. I just got back from the GPTP workshop in Ann Arbor, MI. Exhausted but buzzing.

GPTP stands for Genetic Programming in Theory and Practice, and for those who don't know (which is probably most people - GP is a pretty esoteric subject!) Genetic Programming means the use of artificial evolution to solve problems by evolving computer programs.

At least, sort of. In practice GP has become focused on expression trees rather than full-blown programs with loops and arrays. The chief thrust of research at the moment is into Symbolic Regression, which means evolving mathematical functions (such as y=sqrt(sin(x/4.7)cos(x))+1) whose output can approximate a complex curve from some large data set. Cancer genes, paint formulations, symptoms of illness, that sort of thing. Evolving full-blown programs is still in its infancy (partly because programming languages are innately brittle), so most researchers have headed for symbolic regression because that can solve real problems and hence attract grants.

By the way, GP has some distinctions from GAs (Genetic Algorithms - something more people have heard of), which evolve data rather than executable functions and code. They're a different community of scientsts but there's also some overlap. Either way, it's all very esoteric and abstruse mathematical stuff. Although I'm very familiar with the basic concepts my head was spinning after every talk because I'm not part of that field and haven't grown up with the jargon. There was some interesting research, though, and we had some good discussions. I didn't give a very good talk myself, partly because I'm out of practice and partly because I work in an entirely different field, but I was giving the keynote, so my job was basically just to wake everyone up (which is ironic, given that I was jet-lagged and it was still 6am in my head!) and talk about something wider than their usual narrow academic interests. They were nice people, only a few of whom I've met before, and I expect some of us will stay in touch. It was nice being back in an academic environment after several years in the wilderness, although my head feels like it has exploded now. Conferences and workshops are incredibly intense.

Actually, I might put in a paper myself next year, because I woke up this morning with an idea about how their work in symbolic regression could help this project.

@Deansher - you might be especially interested in this!

One thing that has always intrigued and puzzled me is instinctive behavior. With Creatures I was faced with the problem of how to instill simple instinctive responses into norns, because although I knew how to represent them as rules in the neural network that *I* had designed, I had no way of knowing how their brains might evolve. I couldn't just store the neural wiring pattern in the genes, because if later norns' brains ended up working in a fundamentally different way, the wiring would no longer make sense. In the end I solved the problem (rather elegantly, I thought) by having the genes trigger sensory inputs, force the creatures' brains to behave AS IF they'd chosen a certain behavior, and then inject reward or punishment chemicals into the system. So, for instance, I could temporarily (before birth) make the norn feel bored, trigger the "walk" action as if it had been chosen by the creature itself, and then reward the network for 'making that decision.' After it was born, if the creature got bored (and hadn't learned to overrule this instinct) it was likely to walk somewhere and hence discover a novel object, which it then might learn something useful about. Without that instinct a newborn norn would just stand there until it starved, because it would never even discover that it could have an effect on the world. Anyway, the clever thing about this method was that it worked regardless of how the brain actually functioned internally.

But these creatures have vastly more complex brains and it's not obvious how they can be wired up to make some behaviors likely to occur without the help of experience. And it matters even more, this time. If a natural creature falls over, for instance, this tends to trigger a complex set of reflexes that make it likely the creature will be able to stand up again, even before it has learned how to do this for itself. Getting up when you fall over is a crucial capacity if you're going to be spending your time blundering around learning about how not to bump into things or step off edges. If all the creatures fall over at the slightest opportunity and can't figure out how to get back on their feet, this is going to be a pretty boring game! As it is, I suspect you might have to help the infants of some species quite a bit at first, but they at least need the rudiments of an instinctive response, and since it's likely to involve many muscles simultaneously, it's likely to be complex. How do I wire that into an already complex brain?

In humans, instinctive behavior has always intrigued me. For instance it's pretty crucial that we can recognize members of our own species and also the opposite sex. We can do this without trial and error, so it's instinctive. How do you identify whether someone is male or female? How can this be done in such a way that it can be wired into the architecture of the brain by no more than a handful of genes (that are probably involved in many other tasks simultaneously)? Obviously there are straightforward cues like smell (pheromones), the pitch of the voice, etc. but what about visually? A cartoonist can suggest a powerful hint of the female or male form with the mere flick of a pencil. Just the simple existence of certain kinds of curvature, even in non-human circumstances like the shapes in a landscape, can be enough to trigger an association with one or the other sex. In fact I think this natural tendency to associate certain curves and acceleration profiles with gender has had a powerful influence on human culture and mythology. Anyone who has seen Glastonbury Tor, in England, will be left in no doubt that it was the home of a goddess - it even has a nipple! But I think this neurological cueing goes much deeper and more subliminal than this. It seems likely to me that there is something in the mathematical geometry of male and female body shapes that is encoded in a primitive part of our brains and enables us to begin to tell the one from the other. In essence I'd say that male curves tend to be arithmetic while female curves tend to be geometric (e.g. polynomial). You can see hints of this in the shapes of ammonites and nautiloids - some are Archimedian spirals, which to me look 'male', while others are logarithmic and seem more 'female'. La Tene and Belgic Celtic art diverge from each other in a similar way. But I digress...

So anyway, everything inside the new creatures' brains is represented by neural *surfaces* - 3D curves encoded in the synaptic patterns of layered sheets of neurons. There are three kinds of surfaces: one of them essentially classifies some aspect of the environment - how one thing tends to lead to another; the second classifies the response relationship - how the creature can react in order to change the situation from one of these states to another; and the third biases the homeostatic response by making it more likely that the creature will avoid some circumstances and seek others, even when no top-down goal is present. In his blog, Deansher is speculating that these surfaces might be represented as mathematical functions. I don't actually think that's the best way to think about them when it comes to learning and behavior, but there's no doubt that for each learned surface there is, in principle at least, a complex function that approximates it.

And this is what symbolic regression does in genetic programming. It uses evolution to find mathematical expressions that approximate some curve with n dimensions. These expressions tend to have transcendental functions (sines, cosine, logs, etc.), and nature is very good at generating transcendental phenomena. Waves, for instance, are sinusoidal. Two interfering waves are the sum of two sine curves. Leaky integrators turn up throughout nature and they produce exponential decay curves. So perhaps it is a set of physical phenomena like these that guide the development of synaptic networks in the brain to produce instinctive perceptions and behaviors? Given that the curves needed to drive instincts are likely to be quite smooth compared to learned behavior, it may not take many interacting wave or wavelet-like disturbances to produce the necessary bowls and troughs and ridges in the cortical or subcortical maps. So evolution should be able to find them comparatively readily.

It might, therefore, be possible to evolve instincts, in which genes define simple combinations of transcendental functions by producing interfering waves, which in turn modulate the initial topography of the creatures' brain maps. The result would bias their initial behavior without preventing it from being overruled by later learning. In nature, these patterns would be evolved through natural selection. This is very hard to do, because many of them would be evolving at once, along with the rest of the brain and body, so evolution would find it hard to separate out the variables. But in this game it might be possible to define simple explicit fitness tests and evolve specified instincts deliberately, ahead of the game's release.

Which is where you testers and gurus come in. I might need you to run many instances of the program, long before the game is ready, specifically to evolve solutions to key instinctive responses. I could send out special copies of the game (with partially-built creatures and a specialised virtual world) and ask you to send me back the creature with the greatest ability to get up after it falls, or the one that most reliably makes a noise when pictures of its own species are presented to it. If I can define a fitness test, the evolution will be measured automatically, but for some things I might need you to manually score individuals.

We'll see. I only thought of it this morning and there's a lot to think about yet. But in general, although mathematical functions are not the best way to think about the formation of complex neural surfaces through learning, they may be the answer for instilling the relatively simple rules that drive instincts. So genetic programming by symbolic regression might actually turn out to have been a handy thing for me to have been discussing with the experts. I'll keep you posted.

Okay, enough academia. I still have to sort out rewards and stuff, but for now, if you'll forgive me, I'm itching to get back to work and build myself some test bodies!

[lordcooper]

I swear I understood at least half of that 

This project seems like it could really be something groundbreaking and I'm happy to have had the opportunity to support it.

[Areyar]

Sorry, but I thought recognising ones own species is learned through imprinting rather than as an instinctual development of the brain.
This enables chicks to adopt a bunny as parent or a baby to be raised by wolves, etc.

But as stated a lot of thinking (and research!) is still needed.

[Neonivek]

Sorry, but I thought recognising ones own species is learned through imprinting rather than as an instinctual development of the brain.
This enables chicks to adopt a bunny as parent or a baby to be raised by wolves, etc.

But as stated a lot of thinking (and research!) is still needed.

Well there are natural human instincts that don't need to be taught.

Plus there is something to be said about learning something through imprinting rather then through trial and error or being taught.

I swear I understood at least half of that 

It isn't that hard to understand. The secret about complex language is that at heart it is truely simple.

[Areyar]

Actually, the imprinting process could be called an instinct.

It just does not include own species recognition, rather it is a flexible self-identity learning mechanism. Learning to recognise creatures that care for you as your parents and thus your own species.

instincts are like a gull-chick pecking at parent's beaks in order to get fod and learning that the orangespot is the target to trigger make mommy puking. :&

[PTTG??]

Hoooo boy... So, the latest from the private forum... this is one of the other donators. I believe he is German, so excuse his... grammatical errors.

How would you control yourself in the world? As far as I remember, Steve likes the idea to become one of the creatures – but how to control them?
Maybe you got a good idea out of a other game or just know how it could work for you – tell it to us!

For me I would like 3 options;
- A animated Avatar that looks like a creature, you click on symbols for things you do – I haven’t thought much about this because I like my other ideas more
- Teaching your avatar. You would get a complete creature as avatar, with working brain. But it would just know two things (instead of hunger pain sexdrive etc) good and bad. You would get unique information’s for commands. Those could be combinated by a symbol of your choice and now you would have to teach the creature what each command means. You would also have to use the good and bad thing after you send a command. You would also fight with your mind against your avatars mind. Oh and in case this happens, importing and exporting avatars would become important! And also a start collection of pre-teached avatars. I guess many people would also like to train a good avatar.
- Brain-control, this would be the same like the previous idea – but your avatar would also need to eat drink and similar. And also feel those feelings (but we would have to see those needs, too)
I’m wondering what you guys prefer? ^^

[Neonivek]

I am looking forward to when Grandroids finally comes out for public use.

Though I have no idea when that will be. Wasn't there originally a time when it would become public?

[lordcooper]

Not really.  Originally the dev time was to be limited to around a year, but he exceeded the kickstarter target by a significant amount, meaning we'll probably end up with a higher quality game a little later.

[Neonivek]

Not really.  Originally the dev time was to be limited to around a year, but he exceeded the kickstarter target by a significant amount, meaning we'll probably end up with a higher quality game a little later.

Hmm I was originally under the impression that the game wouldn't be close to done by the time he made it public anyway.

Though if what your saying is correct we could end up waiting quite longer before he is willing to let us normies try (though you probably got a copy von yon cooper)

[lordcooper]

I have literally nothing on my HD yet.  I'll be sure to let y'all know when something playable-ish gets passed around and I'll give as many details as ethically possible.

[The13thRonin]

Are we thinking that this will be released some time in 2012 or some time in 2013?

[PTTG??]

Well, there's nothing to share now. The creatures are no more complete than a completed transmission and brake system sitting in a pile of raw rubber trees and iron ingots is a completed car.

I understand that as things progress more information will become public.