|
Post by windfire on Aug 4, 2023 13:55:24 GMT
tsh73 I'll have to study it But looks like something i can figure out. Not sure if the learning is about saving each individual set of weights of each pattern output or what? My problem seems to be not knowing exactly what it is I'm looking for as inputs after it is trained. It's hard to process without knowing exactly what the goal is. Which set of weights is saved for the overall learned state. Don't know if I'm making sense cause i don't, know exactly what I'm looking for. Normal patterns as My patterns as [0,1] or [1,0,1] or [0.xxyyy,0.yyxx,0.yyxx,0.yyxx] [1,0] [1,0,0] [unk1$,unk2$,unk3$,unk4$] [0,0] [0,1,1] [unk3$,unk4$,unk1$,unk2$] [1,1] [1,1,0] [unk4$,unk3$,unk2$,unk1$] In any case Fantastic Response tsh73!! Thank's
|
|
|
Post by windfire on Aug 4, 2023 14:08:40 GMT
To plus,You would not believe the rabbit hole I'm in !! lol Thou it may appear I'm repeating without progress such is not the case. I think tsh has given me a way to find the answer I'm looking for. I appreciate and can use any help I can get.Thanks ya'll.,{my Texas showing}
|
|
|
Post by plus on Aug 4, 2023 14:43:43 GMT
Hi windfireWell I've been down a few rabbit holes myself: an Interpreter that does adding like this Add[a, b] instead of a + b so I could do strings without enclosing them in double quotes; And a simple GUI for QB64 that works allot like JB's maybe easier to use if you can follow me down the rabbit hole of complexity while I develop the code, good luck picking it up after 2000 LOC even if does have a manual/cheat sheet! ... too esoteric for the people I hang out with LOL Best to stick with code snippets and fun little graphics at forums. After say 300 LOC it's too much to follow and maintain interest and understanding. So for big projects you are pretty much on your own. I am sure it made a huge difference to tsh73 that you posted your source link what was inspiring your question and your attempt to port to JB. He even said he used that. b = b + ...
|
|
|
Post by Rod on Aug 4, 2023 17:47:48 GMT
A self learning program should be just that. You won’t know what is important at the start and so what the inputs should be.
You should have a set of data that exists before and a set of data that exists after.
The starting data is the input and the resultant data is the output.
First step is to value the single data inputs in relation to the desired output. Then all combinations of those inputs till you have sorted table of the worth of all inputs or input combinations.
Taking the top set of inputs lets you calculate the likelihood of those inputs predicting the outcome.
A system such as this can never really predict the actual outcome. It can only predict the likelihood in percentage terms.
So you never can predict the actual outcome of s specific case just the likely percentage success or failure.
Then it becomes a pricing/profit issue about where to set your decision point.
So while they sound sexy and AI ish they are in reality limited and flawed. Especially if you are human and pushed into the decline group.
|
|
|
Post by tsh73 on Aug 4, 2023 21:02:14 GMT
windfire,
learning is changing weights so network responds to inputs like wanted In that example, weights are w() and b() (I've read that bias b() could be considered as as special form of weight)
So it you make it learn, then save w() and b(), then reload them - you'll get learned model
See second post in example thread - I make it learn several goals instead of one.
(now I wonder. There should be some limit as to how much it can learn given this much of neurons? Information should be stored somewhere, here it is w() + b() just 11 floating point numbers, that's all it got. What is the limit? And will it be possible to teach similar small network (how small?) to play TicTacToe?
EDIT: the answer to last one is Yes, examples are plenty. )
|
|
|
Post by windfire on Aug 5, 2023 11:52:12 GMT
Thanks to all!! I see that I was looking at neurons inside one neuron , thinking of it as a network. It appears to be a network_inside a network.But that requites double looping all the way around, to teach and learn. Quite confusing! However, I think now, I have and understanding of it I can work with. I wanted to understand it, so I could relate it to another perspective,I have, of doing it differently,and better. And I believe it will work,I have to exsperiment with it,to see. Many thanks for you patients and help.Means a lot to me.My comunication with writing skill suck,I know.
|
|
|
Post by windfire on Aug 6, 2023 12:59:27 GMT
tsh73 Your examples cleared up a lot of questions I had. I think I have a handle on it now. I hadn't realized each pattern Op_Y had it's own goal.
|
|
|
Post by Rod on Aug 14, 2023 17:32:39 GMT
What was interesting? I don’t expect an answer from this bot fishing for, what?
|
|