Printable View
Hi everybody.
I would like to invite you to beta test my latest extension: the Multilayer Perceptron Neural Network object, available here.
The MLP Neural network object is mainly intended for machine learning and AI. This object may be a bit advanced for the average user, refer to the included FAQ & examples for some quick hints.
This is still a beta version so please e-mail me any bugs you encounter.
Enjoy.
Oh wow... Cant wait to give this a try
Finally got some time to play around with it, and I have to say, this is amazing. Seems to be very easy to use, and so far I haven't run into any bugs.
I think the developer is looking for good, illustrative examples that would make it clear to many users how the extension works and learns in practical situations.
This is a great chance to get creative and help show others what this kind of tool can do. Once others see it in action more and it is a bit demystified, who knows where this might go?
An animated illustration, (similar to the object's icon) could really be useful for those unfamiliar with the concept.
I am glad you like it Shadoku.
As Novabrain mentioned, I'm very interested to see what power users can do with this object.
Wow, this is excellent work :)
This is a really exciting extension! I managed to make a simple program where one object was learning to set its Y position to the Y position of another object. It was interesting.
I hope to see some interesting examples of this object's capability. I'm going to try some more :)
Wow! This would be perfect for an virtual pet sim:)
Even if your experiments are basic and simple, we hope that everyone will consider posting an .mfa in the file section for others to learn from and work with.
I am glad to see the interest shown so far. Right now, it is a matter of getting the ball rolling ;) Everybody is welcome to pitch-in and give it some momentum.
I'll polish and comment some of my work later.
The examples are confusing me. What are they demonstrating? At first I thought (in example 1) that it would calculate how to hit the left wall, but it doesn't always, so I'm not quite sure.
It is learning to detect where the ball is going. Note that depending on how it learns it may or may not be able to predict accurately in all cases.
Typically a network would not be used to accurately do the fine calculations, but rather decide on where approximately the ball is going and then some finer calculations would be done to move in the direction of the ball or something. Just like you can not manually predict which exact pixels the ball would hit, but you can make reasonable estimations on the region it will hit.
Mokhtar - One thing I was wondering about the examples is, why did you make the user of the applications train the network manually? Especially in the second example. Isn't it much more efficient to pick a random coordinate in that box to fire the ball towards, rather than have the player manually choose spots to fire at?
Sure, you're right. In a real game you'll probably do all your training automatically offline and only train a little bit while the actual game is running.
I know this is an old topic, but I was just wondering if there have been any developments? It seemed really promising!
I think the lack of suggestions about what to implement or applications of it, stopped it..
link broken
Yes the link is broken. And I don't know where to get the extension anymore. I tried emailing Mokhtar but he hasn't replied yet.
Users who see the potential of this and want to see it developed published eventually should support it by testing and suggesting. There is a lot that could happen if that comes about.
Let's hope that Mokhtar is still around and interested in continuing, in that case.
anyone has a copy of it? since the link is broken... i'd like to try it!
Now I feel bad because I was looking forward to it, however I didn't post enough!
I think this extension could be really amazing when it comes to AI. We would be able to program AI that can learn and teach itself the best way to do things, rather than hard coding every situation.
Well, here's hoping it can continue. From previous discussions, this extension was, at least, a first step in that kind of direction.
I don't feel comfortable with distributing the extension file without Mokhtar being around or acknowledging it. Since it is not available online, that could mean he pulled it or, it could something else.
I see ^^ it's ok, tnx :)Quote:
Originally Posted by Novabrain
I just got netninja's email. The extension was not pulled, it's only that the "www" in the URL should be removed now. Here is the updated URL.
The object did not receive any further updates because there were no bug reports or feature requests. So why fix something when it isn't broken?
It could be nice to have more examples for it. :)
I agree, I'd really like to see some interesting examples.
Umm seems the link is down, it appears I can't download your A.I. Extension or whatever it is, idk.Quote:
Originally Posted by Thunder-Power.com
Mokhtar should really edit the first post or fix his website to allow www. (or at least redirect from www. to the plain url).
Look a couple of posts up, he posted a fixed link.
After having a look at the MLP Extension it doesn't seem to make any sense what so ever.
In the Bat & Ball Example, I coudn't do anything and the ball just moved, whereas both Bats just sat completely still the entire time.
Now for the second example, the bat moved one stop closer to where the ball was going to, isn't it supposed to move all the way there?
I haven't looked, but I'd assume that the bats are learning, and they need to fail repeatedly before they get any good.
This seems interesting, if im thinking correctly it could be used to make a npc find it's way around a maze area etc?
I never saw this! It's brilliant. You could really design nice AI routines based on how the player plays and no longer have hard/easy modes. I'll try messing around with this, but please remove the pop-up at the beginning x).
Can I make chatbots with it? Conversation programs, if you want...
Unlikely, natural language parsing is far too complex for a neural network. Neural networks take input and give output as numbers, so you'd really struggle with text.
This is suited to learning movement. Give it enough inputs and outputs, and you can get something a lot like an ALife, that moves about to feed and runs from predators etc.
Well, if I use the Dictionary object, combine with some lists, I think I will do that. Oh, and a "good" button to teach, if network will give correct output. Cross your fingers for me!
//EDIT: And author should remove this irytous message, when application start (this is beta version...)
So when is it coming out of beta?
How bout a nural net calculator. This would be fun.
I have idea for another learning scheme:
1. Teacher set, um... call it "Question"
2. Teacher set a list of correct "answers".
3. Teacher test network with another parameters, to check network's "knowledge". (as in school ;) )
4. Teacher rate answer.
Of course network should be a good student ;).
And MLP object, what can input/output both strings and numbers will be great.
Well if you want to teach your Application to learn language i'd suggest only basic words like you get in Oddworld like hello, goodbye, etc, you shoudn't put too many words in.
Figuring out how the Human Brain works might help you figure out how to better the extension, Mohktar but one thing you have to remember is that the Human Brain has many, many Inputs but few Outputs, also loads of neurons.
He's not reproducing the human brain, I believe.
Theoretically, though, you can (with enough processor power).