| Blogs > Qzy |
|
Qzy
Denmark1121 Posts
Does anyone understand neural networks and how they work with multiple layers? Got a bunch of questions for it, to even being able to understand it slightly - most scientific texts on neural networks are very strong in math, but not doing a good job explaining what the... is going on.  | ||
|
Tabbris
Bangladesh2839 Posts
| ||
|
Glacierz
United States1245 Posts
| ||
|
ZBiR
Poland1092 Posts
 | ||
|
meeple
Canada10211 Posts
| ||
|
Qzy
Denmark1121 Posts
Okay here's a few questions. I get the basics of it - but ... How many neurons should you use, with 2 inputs? Do you HAVE to use 2 neurons in the hidden layer, or can you simple use 1? Can you use 5? What's the benefit for using less/more? You can have more hidden layers - but with what benefits? Should it have the same amount of neurons as the other hidden layers? When teaching the NN how to play, for instance, tictactoe, do I give it training examples: This is the input, i expect this output...? How many examples does it need to adjust to play decent? Can it generalize, once it has seen a few examples? | ||
|
Glacierz
United States1245 Posts
| ||
|
Qzy
Denmark1121 Posts
On September 17 2010 23:33 Glacierz wrote: Based on your questions, I suggest you start out with Bayesian networks first before getting into neural networks. Can't :/. I'm following my specialization course. This week it's neural networks, where we have to make a ludo player (in 1 week), next up is generic algorithms and then reinforcement learning. Sigh. | ||
|
Cambium
United States16368 Posts
On September 17 2010 23:16 Qzy wrote: Thanks for already answering. Okay here's a few questions. I get the basics of it - but ... How many neurons should you use, with 2 inputs? Do you HAVE to use 2 neurons in the hidden layer, or can you simple use 1? Can you use 5? What's the benefit for using less/more? You can have more hidden layers - but with what benefits? Should it have the same amount of neurons as the other hidden layers? When teaching the NN how to play, for instance, tictactoe, do I give it training examples: This is the input, i expect this output...? How many examples does it need to adjust to play decent? Can it generalize, once it has seen a few examples? It's been a while since I took ML, so I don't remember too much about NN, I'll give it a shot You should have at least N+1 nodes in each layer where N is your # of input nodes. You can, of course, build a layer with any number of nodes, you just won't reach the accuracy you desire. I remember at one point, this made intuitive sense to me, but I don't remember it well enough to describe it back to you. There is no hard restriction on the number of neurons in each layer, this is something you have to experiment by running the NN multiple times wrt your training data, and choose the one with the highest accuracy. You can also do this with any number of hidden layers. The reason why you shouldn't use excessive number of neurons and hidden layers is to avoid overfitting (I think...). Each layer can have a different number of nodes, and the optimal number of layers depend largely on your input data and your activation function (such as gradient descent, sigmoid). Too few nodes cause underfitting, and too many nodes cause overfitting (again, I think...). Tic-Tac-Toe is actually a difficult problem to solve with NN (I'd use decision tree actually) since it's adaptive Your first task would be to digitize all of the moves in a given game (so every game is one piece of training datum), and the output would be win, lose and tie. Alternatively, you can assign state a value (much like Chess) so that every move can be a row in your training data. I hope this helps. | ||
|
illu
Canada2531 Posts
| ||
|
Cambium
United States16368 Posts
| ||
|
Qzy
Denmark1121 Posts
On September 17 2010 23:55 Cambium wrote: It's been a while since I took ML, so I don't remember too much about NN, I'll give it a shot You should have at least N+1 nodes in each layer where N is your # of input nodes. You can, of course, build a layer with any number of nodes, you just won't reach the accuracy you desire. I remember at one point, this made intuitive sense to me, but I don't remember it well enough to describe it back to you. There is no hard restriction on the number of neurons in each layer, this is something you have to experiment by running the NN multiple times wrt your training data, and choose the one with the highest accuracy. You can also do this with any number of hidden layers. The reason why you shouldn't use excessive number of neurons and hidden layers is to avoid overfitting (I think...). Each layer can have a different number of nodes, and the optimal number of layers depend largely on your input data and your activation function (such as gradient descent, sigmoid). Too few nodes cause underfitting, and too many nodes cause overfitting (again, I think...). Tic-Tac-Toe is actually a difficult problem to solve with NN (I'd use decision tree actually) since it's adaptive Your first task would be to digitize all of the moves in a given game (so every game is one piece of training datum), and the output would be win, lose and tie. Alternatively, you can assign state a value (much like Chess) so that every move can be a row in your training data. I hope this helps. Thanks it clears it up a bit.. Right now, I have to implement a ludo player - with a lot of possible states (like chess). Do i simply give it a few examples (inputs, and expected output), and it can generalize from these examples, once properly trained (output has reached desired)? | ||
|
Cambium
United States16368 Posts
On September 18 2010 00:02 Qzy wrote: Thanks it clears it up a bit.. Right now, I have to implement a ludo player - with a lot of possible states (like chess). Do i simply give it a few examples (inputs, and expected output), and it can generalize from these examples, once properly trained (output has reached desired)? Well, you first need to classify your inputs and outputs. In a game of tic-tac-toe, say you are red, it would be along the lines of, how many red on each line (attack), how many blacks on each line (defence), and maybe a few more. Your output would be the quantification of the state after you place your piece. If you win or prevent a loss, it would probably be max, and go from there. You would need a lot more data than a "few lines"; I would think in the order of hundreds if not thousands. I would try to find existing data for tic-tac-toe and see how experts classified the game. The best way to obtain data is to either find it, or to create an online version and ask your friends to play so that you can classify their respective inputs and outputs. | ||
|
Qzy
Denmark1121 Posts
On September 18 2010 00:09 Cambium wrote: Well, you first need to classify your inputs and outputs. In a game of tic-tac-toe, say you are red, it would be along the lines of, how many red on each line (attack), how many blacks on each line (defence), and maybe a few more. Your output would be the quantification of the state after you place your piece. If you win or prevent a loss, it would probably be max, and go from there. You would need a lot more data than a "few lines"; I would think in the order of hundreds if not thousands. I would try to find existing data for tic-tac-toe and see how experts classified the game. The best way to obtain data is to either find it, or to create an online version and ask your friends to play so that you can classify their respective inputs and outputs. Then it's good i got 3 more days to come up with a ludo player :D God I love university with their "1 week to understand 50 years of AI, and implement it - kkthxbye" . So lost in this - how's it possible to create a working neural network in a week.. seriously. | ||
|
Glacierz
United States1245 Posts
| ||
|
AcrossFiveJulys
United States3612 Posts
On September 18 2010 00:09 Cambium wrote: Well, you first need to classify your inputs and outputs. In a game of tic-tac-toe, say you are red, it would be along the lines of, how many red on each line (attack), how many blacks on each line (defence), and maybe a few more. Your output would be the quantification of the state after you place your piece. If you win or prevent a loss, it would probably be max, and go from there. You would need a lot more data than a "few lines"; I would think in the order of hundreds if not thousands. I would try to find existing data for tic-tac-toe and see how experts classified the game. The best way to obtain data is to either find it, or to create an online version and ask your friends to play so that you can classify their respective inputs and outputs. I wouldn't recommend using a neural network for learning a game unless integrated as part of a self play system. If you want to read about a way to kick some serious ass, look up reinforcement learning backgammon. Cadmium is suggesting that you use the NN as a state utility evaluator (and then presumably use those evaluations to implement minimax?). That is a viable approach, but I think you could implement the evaluation step yourself and get pretty far and skip the NN part. If you want to do this the simplest way possible, get someone (or a program) that plays the game pretty well and collect a lot of data saying "at this state, do action a". It will be important to choose your inputs wisely. You could describe the entire board state as your inputs, but that will make it harder for the NN to generalize. Instead you should consider coming up with some features of the state that are interesting. As for the parameters, such as the number of hidden nodes, hidden layers, learning rate, momentum term, how much data you need, etc, you have to understand that there isn't hard theory that says you need exactly this much. In practice getting neural networks to work is a form of black magic: you must empirically determine a good parameter setting through your own intuition and lots and lots of experimentation. | ||
|
Qzy
Denmark1121 Posts
On September 18 2010 01:03 Glacierz wrote: Tic-tac-toe is easily solved by alpha-beta pruning, no need for complex frameworks like NN You mean minmax  not alpha-beta pruning. But alpha-beta pruning does help to speed up the search. But this is neural networks - always 2nd best method, but very suitable, when search space becomes too big, ie chess.I'll try training my neural network with some common states in my game, where i know some good outputs - and then i hope it can generalize this to more states which looks like it. Is this an okay way to do it? It's pretty slow to "hand feed" it everything. | ||
|
ToxNub
Canada805 Posts
When people used NNs to play chess or whatever, it's really not just a NN. The NN in those games is really just a function approximator that tells you whether or not a given gamestate (board) is "good". Both chess, backgammon, and tic-tac-toe all rely on sequential moves, which means you've added a temporal element. ML techniques like temporal difference learning have successfully been applied in the form of TD-gammon, TD-Chess (and probably TD-tic-tac-toe) and are more suited to your needs. | ||
|
Qzy
Denmark1121 Posts
On September 18 2010 01:11 ToxNub wrote: I've written a few neural networks by hand before. I don't really understand how (why) you would get a neural network to play tic-tac-toe tho. NNs are supervised learning algorithms, which means you need to already know the answer in advance. Your network just sorta "remembers" the answers you've trained it to remember. Granted, if you get enough data, you can provide a novel input and you MIGHT get novel output, but for your purposes you will likely just be "remembering" what to do given gamestate a, b, c... Boring. Not an interesting application of NN at all :p When people used NNs to play chess or whatever, it's really not just a NN. The NN in those games is really just a function approximator that tells you whether or not a given gamestate (board) is "good". Both chess, backgammon, and tic-tac-toe all rely on sequential moves, which means you've added a temporal element. ML techniques like temporal difference learning have successfully been applied in the form of TD-gammon, TD-Chess (and probably TD-tic-tac-toe) and are more suited to your needs. Actually I have to implement a ludo player for my AI course - but tictactoe is more simple to start out with (yes even tho it's complex solution to a simple question). Like I wrote earlier, I'll try giving it some states, ie start state, and tell it: The correct output here is to move this piece. And some more states, and hope it can generalize from it. Is that even possible? | ||
|
Glacierz
United States1245 Posts
Had to do it for reversi player for my AI class, just used minimax with A-B pruning. | ||
|
ToxNub
Canada805 Posts
If the game is retardedly simple, and you can play it optimally, sure, you can tell it "given gamestate a, make move 3" in your training examples. This is the trivial case where the NN just memorizes your input/output mappings. Depending on your teacher, he may not accept this. If you have enough of them, you might be able to give it a state it's never seen, but it's hard to say if it will be able to generalize. Usually you need a lot of training data for that, and it sounds to me like you would have to generate it on your own. If you can't play it optimally yourself (and hence can't provide good training data) or you actually want it to LEARN how to play ludo (and not tell it how to play ludo), you need to come up a way to provide feedback. You need to be able to say "this is your goal", and let it figure out how to get there on its own. The goal of learning, at its heart, is to get the program to do something you did not explictly write. If you just give it all the answers, you really just have a sounding board. So for tic-tac-toe you need to abstract a "move". What defines a bad move and what defines a good move? Obviously a move that ends with a loss is a bad move, and a move that wins the game is a good move. Unfortunately, there is much more to it than that. What defines a move that leads to a good move? What defines a move that is good now but loses you the game next move?  These are not simple questions, and I'll stop rambling on and confusing you. Basically, this is where you have to get creative. | ||
| ||
StarCraft 2 StarCraft: Brood War Dota 2 League of Legends Other Games |
|
Replay Cast
Wardi Open
OSC
Solar vs MaxPax
ByuN vs Krystianer
Spirit vs TBD
OSC
Korean StarCraft League
OSC
OSC
OSC
uThermal 2v2 Circuit
|
|
|
EPT
