Idea to take GForce to an entirely new level...

[mobile1]

chkman - I am thrilled someone is giving this a try.

[chkman]

Is the purpose of the genetic algorithm to test on only some but somehow come up with the best without viewing all possible combinations? Is that what it's all about? (Possible lighting of the bulb at this point )

Hmmmm.

I guess I was reading more into it than what you are talking about. I thought we'd have some kind of evolution of configs, meaning creating new ones that don't currently exist. I was generating tons of configs and was looking for an easy way of filtering thru them to find the ones worth watching. I like the variety of many possible combinations and was thinking more about just preventing the worse ones from occuring. Like a wave at the edge with an outward flow. But that wave with any inward flow would be okay.

I was going in a way different direction.

If you just want to know the best, here it is: Wave-Kaleidoscope, Flow-Tunnel Vision, Color-Dali and Particle-DT - ZoomPlanet.

I hope all this makes sense, and I hope my understanding of GForce as described above is correct.

Correct. It's slowly sinking in what you want to do. There are over 100 of each config type except particles. Unless we could run waves as particles as I said previously. I used to do it under 2.0 but can't under version 3. Just have to make do with less of them.

We can use a rating scale of 0-9 if you want to. But G-Force won't be sending us the choices. We have to think outside the G-Box for that. We need a web form to collect the data. The script would display the 'name of the individual' being evaluated and the users cast their votes in the browser and click submit.

Let's get the 4 going before we add any more dimensions. And transitions? I never saw a transition I didn't like. What about unions? If you want to make it more complex - unions. If you haven't purchased you haven't seen unions.

These two flowfields kinda illustrate what the unions do. In the 1st there is certain activity in the center. Then there are circles around it with a different flow.

In the second example there is a flow out from the center and the squares and rectangles with a plaid look to them represent the union of the second flow.

The union determines what areas of the screen keep the 1st flow and which areas do the new flow.







Well, enough talk. I have work to do....

[mobile1]

Thanks for your info - I sent you an email with my contact info.
To answer some of your question:

GA - yep you got it, with a GA you don't test all solutions.
It's actually pretty straight forward. Genetic Algorithms are one of the only algorithms that can (interatively) solve class NP problems - don't know how much you know about math - but this is a type of problems you normally can't solve with math or statistics since there are too many solutions, and no way to solve the problem with equations. One such example is Travelling Salesman Problem but there are many others. The larger the number of possible combinations the more powerfull is a Genetic Algorithm. Also GA's are excellent where there are many local (and probably global) optimums (which is the case with GForce).

[chkman]

The unions will require which union config is being tested. I'm not sure if an either or approach to transitions and configs is good. With transitions you have a few seconds of change and the new flow is what it is. With a union the new flow is a combination of the new and old. Unless you want the union to clear and then show only the new flow.

Having problems with the server so far. I'll keep you posted on progress.

[chkman]

don't know how much you know about math

It has something to do with numbers, right?

What if we test the concept with just waves and flows? A lot has to do with whether or not the flow needs activity in the center of the screen and whether or not the wave does anything in the center. Is the final generation going to be the top 20% as during the run?

In the next phase we could consider a wave/flow combination as one 'chromosome' to be used in more complex ratings. I am thinking that someone not liking a color or particle etc might skew the results of wave-flow. So if we determine the best combos of that first then we'd get better opinions on the other things you want to look at.

Is there a possibility to use a Java Applet

Them's fightin' words buddy, watch it!

Seriously, I don't of G-Force actually using 0-9. But not all media players will send keys to G-Force. Windows Media Player 10.0 doesn't seem to. As far as a program catching the keys before they even get to the player, I'm not sure. Sounds like I need to write my own player to have G-Force plug into. But I shouldn't be spending a lot of time on that. If the user is testing only 10 'entities' at a time a web form should work fine.
It will take awhile just to get this 'generation generator' going. It would easier as a desktop application but using the internet is a little more complicated.

[mobile1]

your right lets just test the concept with a very simple example.

[Linvincible]

hi guys,
the concept is very good,
but it sounds like too much involvement from the user to have it really work!
I mean, who is going to actually vote for the configs as they cycle through?
the number of votes could very well be insignificant in comparison to what would be necessary to achieve anything real
it's not at all like simulating basic pluri-cellular lifeforms
unless you can create an "environment" and define rules following which a wave / flowfield would live, the math funtions beeing the cells...
could be fun!
but seems to be a LOT of work...

just my opinion, congratulations if this gives any result!
and I find you all very brave to try

[mobile1]

hi guys,
the concept is very good,
but it sounds like too much involvement from the user to have it really work!
I mean, who is going to actually vote for the configs as they cycle through?

Well the first prove of concept will be a bit cumbersome when you need to click the submit button for each rating, however it should be enough to prove that it works if a couple of people would dedicate some time.

I think the end goal should be where the user simply presses a digit between 0 and 9 to rate it - nothing else required. I think there are enough people out there who would rate it. It would even work if you'd only rate the worst and the best ones. It would even work if only a handfull of people rate configs. At the end it comes down to statistics, if you have 1,000 people using the program and only 2% rate configs (it would still work).

If we can make this work the results will be mindblowing.

[chkman]

Linvincible

I mean, who is going to actually vote for the configs as they cycle through?

I was hoping you would.

define rules following which a wave / flowfield would live, the math funtions beeing the cells...

Yep, that's what I'd like to do eventually...

[mobile1]

By the way if you want to create rules - about what combinations people like and which they don't like, I could help you with that as well - I have quite a bit of data mining experience, which would be perfect.

However in order to give you rules, we would need a couple thousand ratings to come up with some reliable rules - and the only way to get these ratings is through this GA. So once we have this GA running, to do the rules is easy and basically a side product (I could simply mine all the ratings)

In the future GForce would ideally come with 3 modes.
-Random
-Rules Based
-AI Evolution (with optional rating - requires internet connection)

[Linvincible]

chkman

I was hoping you would.

I would, definitely!
I just hope there will be enough jury members

[chkman]

Linvincible

Thanks!

Mobile1

Should we use just standard configs or include those from the extras page?

[andy55]

Ok checked back again after a long time - (about every year) - always hoping someone is doing the Genetic Algorithm thing - but it looks like we are still not there yet.

I haven't been aggressive in responding to this thread, because I thought mobile1 and I covered it in that email correspondence back in Dec 2001. Please permit me to paste parts of my responses to you in the interest of saving time...

/snip/

Problem: user X may think c3-f3-w3 looks incredible while user Y hates it... In other words, you’re making the assumption all users like/value the same things... For example, my favorite pizza topping is onions and sausage while that’s someone’s least favorite combination. To finish the analogy, there is no winning pizza topping combination because each user has different tastes. Or in a more related medium, I don’t like jazz while many many others would die for it. There is no master/universal genre of music that’s the “best�, and likewise, there is no unilateral "good" config set (ie, trait). This is in contrast to natural selection.


> I mean if you look at nature. I believe that there
> is nothing more complex then nature...so if a
> Genetic Algorithm works for nature, I am sure it will
> definitely work for G-Force.

Don’t forget that Natural selection is still a “greedy� process. Like I said above, a trait will never be selected for if doesn’t play a role in an organism’s immediate survival. An immunity to an epidemic that won’t appear until 10 years from now will not be selected for today. Likewise, the highest elevation won’t always be found by traveling in the direction of higher ground—you’ll get a a high place, but you have no guarantee how it will compare to the highest ground.

...

GA doesn’t work for G-Force: none of the components are universal: I like onions and sausage, John hates them, while Mary loves them only if there’s mushrooms present. I love milk and cereal, and I love fruit, but I don’t like them together—but Rob does--he’ loves them *only* together. There’s no rhyme or reason to people’s fancies (presumably because artistic preferences haven't been selected for during the course of mankind's evolution).

/snip/

If you reject my pizza argument by saying that the GA will be applied only for a given person's preferences (and thus, only by their input), now you're forcing that person to endlessly rate something visually. Most people (myself included) are not interested in turning their music visualization experience into a data-entry job. You just want to sit back, dive into your music, and get taken away. Who want to be playing with a mouse or keyboard during that process? I would argue the two experiences are mutually exclusive.


> I think you guys need a major development step and take
> GForce and your company to a new level

We actually are going to a new level -- we're hard at work making new stuff. As much as I like G-Force, it's time to leave it where its at and move on to the next generation visual engine (project Gallium). We'll be bringing G-Force to OpenGL and adding some exciting new things that come along with that, but visually, it's fully matured. My sights and visions are much much higher than G-Force and have been for some time now.

> invest the couple of weeks it would take to test this GA idea....

This unfortunately is not realistic. You're right about core GAs being straightforward to make, but the rub is making the (successful) glue needed to make it "work" with the data sets in question. In the case of G-Force, the human-perceivable return of GAs would be low (despite what you may think). In the interest of time, I will justify this position by saying that the most of the rules and laws of how the human mind judge, value, and regard visual art are very very different from the black and white rules that measure and score computing and survival problems.

You seem to regard GAs as a magic wand. GAs are a tool, like any other, and the challenges of the "tasteful realtime visual art" problem can't be solved by throwing a single tool at them. They demand a unique and innovative approach designed to incorporate some of the high-level processes that people use when they create art (consciously or subconsciously). Such examples are the golden ratio, self-symmetry, color matching and complimenting, use of patterns/textures, and visual structure styles.

I think it's really great that you've taken such a strong interest in sharing your ideas and what you think may make G-Force a better experience. However, there are many other challenges to the realtime visual art problem that are outside the scope of GAs. The challenges of computer-created art are the content, style, and design, not the medium, mechanics, or tools (just as is the case with human-created art, ironically).

Andy

[chkman]

Andy

So what are you saying? Don't bother?

BTW, A pizza has to have pepperoni, toppings can include anything else except anchovies.

[JayPro]

Maybe what he's trying to say here is that this whole genetic algorithm thing is not at the point at which it could be simply applied to any sort of music visualization program even in the mid-range future.

As one sufficiently interested (yet not expert) in artificial life and cellular automata, I too would like to see how music can be applied to the *many* fields in this subject, aside from the somewhat shaky marriages between the two that I see existing now.

It's just that the ideas bandied about here at length offer scenarios that are too far down the road to do anything with now.

Maybe this now Project Nitrogen will in due course demonstrate the sort of versatility that can make it all happen.