Warning on NN threads!

OMG maybe you should write a NN to automatically close all NN-related threads!!!

NN? Non-Nude?? Ohhh Neural Network!

I love them, I made one that controlled a car autonomously that nailed me a 1st.... sweet.

Seriously, folks. What is up with all the grumpy moderators?

Quote:

Original post by Hodgman Seriously, folks. What is up with all the grumpy moderators?

C'mon... this NN stuff is the equivalent of wandering into the graphics forum and suggesting that we experiment more with the Apple 2 low-res, 16-color graphics and trying to make it more like the Apple 2 hi-res, 8 color, 2-palette graphics. That wouldn't go over too well, would it? Why? Because it is old-school and sucks at doing anything productive compared to other techniques.

*sigh*

Seriously, enough with the NNs. GAs are the only tool you ever need. They work like nature so they must be smart!!

On the other hand NNs are simple to understand and provide a smooth introduction to soft computing. In particular, many of the better alternatives people suggest are more complex to understand and are less well covered by the literature. Perhaps more needs to be done to explain the alternatives.

Edit following Sneftel's post: I actually make a lot of use of GAs. Again, for a specific task it's unlikely that they are anything like the optimal solution but they're a pretty flexible tool that is fairly easy to tweak to your needs once you understand the concepts. There's a lot to be said for a tool that is easy to understand and adapt, even if better ones exist.

I actually think both NNs and GAs can be very useful. The problem is when people assume that "AI" means "AI algorithm", and that that means "brainy algorithm", and that that means "NN". So some guy who wants his agents to seek out food spends a week figuring out NNs and backpropagation and what his inputs and outputs should look like and deciding on a reasonable layout, and then after a week, he's done, and it TOTALLY GOES TOWARDS THE FOOD!

Hey, guy, spoiler alert:

if(food.x < agent.x) {    go_left();} else {    go_right();}

Yes, I know these toy problems are ways to learn about neural networks while keeping complexity under control. But the whole POINT of NNs is that you use them when the model is too complex to understand. The considerations are totally different. The approaches are totally different. The results are totally different. The domains are TOTALLY DIFFERENT. If you want game AI, pick an approach that fits the problem. And if you're going to cut your teeth on machine learning, pick a problem that fits the approach.

The same thing happens with GAs, though GAs have the disadvantage of not having "neural" in the name, so people jump for them less.

Actually, I've just had a brainstorm. Guess what, guys? I've just come up with a new AI technique. It's called Brain Intellect Vectors, and it's the same as k-nearest neighbors, only it has a name that'll make people think it's so smart. After all, if you do have just one tool in your ML toolbox, kNN isn't a bad choice for that tool, and it's easy to understand and apply. (Or maybe SVMs... I'm open to suggestions.) We just need to convince everyone that BIVs are based on how the human brain works, so that they can make decisions in the same way that human brains make decisions! What's a neuron? I forget! BRAIN INTELLECT VECTORS!

Quote:

Original post by Kylotan On the other hand NNs are simple to understand and provide a smooth introduction to soft computing. In particular, many of the better alternatives people suggest are more complex to understand and are less well covered by the literature. Perhaps more needs to be done to explain the alternatives.

Yeah... someone should write a book describing how to construct behavior algorithms using stuff like utility theory, response curves, weighted randoms, and stuff.

Sup'

I guess I'm one primary target here, as I've seen a lot of NNs threads pop up after mine this week.

Well I'm sorry if just seeing a NN thread makes you angry.
I had hoped to be quite polite in mine... I even understood what you told me and, eventually, agreed with you :p

I'm not sure that anything related to NNs deserves such a warm welcome...
Not even sure that anything-related-to-NNs-and-written-by-a-newbie deserves it.

But if NNs are such a dead end (and you can't assume that everyone knows that), simply add a mention of it in the FAQ... much like what Game Design section does with MMO related threads.

Now, maybe this section would indeed be better with the same query for an A* algorithm over and over again...

Respectfully,
* A sad Piou *

Quote:

Original post by TiPiou I guess I'm one primary target here, as I've seen a lot of NNs threads pop up after mine this week.

Nah... this goes well beyond you. You were just one of many over the past... 7 years? 8 years?

Quote:

Well I'm sorry if just seeing a NN thread makes you angry.

Not angry... just exasperated, I suppose.

Quote:

But if NNs are such a dead end (and you can't assume that everyone knows that), simply add a mention of it in the FAQ... much like what Game Design section does with MMO related threads.

The problem is, there are educational institutions that keep teaching their students that NNs are vitally important despite the fact that they really don't go anywhere. Why would a school teach people that? An FAQ can't compete with that.

Quote:

Now, maybe this section would indeed be better with the same query for an A* algorithm over and over again...

Don't get me started!

I suppose it is a shame that people waste time trying to fit a glorified regression and fancily named blind search when stuff from decision theory & trees would be much clearer, appropriate and arguably more elegant for their games. While for more impressive things, Reinforcement Learning is far more interesting.

Although it is obvious that neural networks are the most intelligent since they are neural.

Quote:

Original post by Daerax I suppose it is a shame that people waste time trying to fit a glorified regression and fancily named blind search when stuff from decision theory & trees would be much clearer, appropriate and arguably more elegant for their games.

This.

The problem with NN is the name. It lures people in thinking its a simulation of how the brain works. Actually, the name came decades after the concept was first published. Its an old academia trick: take something old, re-publish it with a better name, get funding­.

Couldnt you write a script that would replace the words "NN" (and equivalent) by "regression" and "GA" (and equivalent) by "gradient descent" in every post ?

As someone who used to teach neural networks (albeit for vision systems and text recognition, about 10 years ago) I take umbrage with the idea that they're useless. They may well be useless for most game AI (apart from some augmented reality work), but not useless in general.

Also, as an ex-neurological system modeller, McCulloch and Pitts neuron are a good starting point for (simple) mathematical models of the neural system.

Quote:

Original post by gregs As someone who used to teach neural networks (albeit for vision systems and text recognition, about 10 years ago) I take umbrage with the idea that they're useless. They may well be useless for most game AI (apart from some augmented reality work), but not useless in general.

That's the point, though, isn't it? This is "GameDev.net" and people are coming here often to ask about how to use NNs in a game situation. The examples you cited are both pattern recognition algos... which is only in a broad sense related to decision algos.

Is there a problem with people trying things out for themselves? Surely you tried them once as well?

Quote:

Original post by Side Winder Is there a problem with people trying things out for themselves? Surely you tried them once as well?

Methinks you have missed the point entirely.

but... NN's are occasionally used in games, or at least in the tool sets behind games.

Maybe the usefulness of NNs (and other more common techniques) could be put in the Forum FAQ, so then these 'annoying' posts could be simply responded to with "RTFFAQ"

I would argue that those who use neural networks in games used them because of the amount of publicity they get. Neural Networks are best used when you want to learn data (typically for classification) whose structure you don't understand and where the whyfor of the learner's output is neither important. And even in those cases Support Vector Machines have been shown to often outperform ANNs. And then in the other case where it is important to have some understanding of the numbers of the system probabilistic graphical models are far better - markov chains and my favourite - Bayes Nets.

Neural Network have a place but as we develop machine/statistical learning that place is becoming to be seen to have a smaller and smaller place. As for games, ANNs make little sense to me. Having agents trying to maximise their expected utility and tweaking their rationality and braveness by adjusting their risk curves and throwing in some random numbers will lead to some awesome behaviour and for more enjoyable than the black art of spending countless hours training, tweaking and training your Neural Network to duck for cover - only to find it only ducks when the gun shoots lasers.