Symbolic AI Systems & Neural Nets
Author
Hi it''s me again, the guy doing the original research proj.
Anyway, i''ve put alot of thought into the project and i''ve decided that my project will be a comparison between Symbolic AI Systems and Neural Nets. I''ll do this by implementing the methods within a tic-tac-toe game (or something similar)...
Anyway, what do you guys think? Suggestions, comments, etc..
Digital Radiation
Just one comment about tic-tac-toe, which I''m sure you already know: there is no ideal solution. The closest you''ll ever get to one is a draw. To those who don''t know this (are there any?), play the game, but this time, sit and think about your move. Play it like chess. Two good players will always draw against each other.
The_Minister
1C3-D3M0N Interactive
The_Minister
1C3-D3M0N Interactive
I''ve an even better idea, how about comparing the two systems in playing Go ;-).
This will highlight the problems of symbolic AI''s search methodologies, although I severly doubt you''ll get an NN system up to a good standard without long term genetic algorithm use.
I heard a quote that there was in the region of 10^390 possible games of Go. It''s played on a board of 19x19 squares. That''s 361 squares, meaning that in the first move you have 361 possible squares in which to place a piece (the board starts empty, you fill it up one square at a time, one player playing black, the other white).
Move Number of different possibilities
1 361 = 361
2 361x360 = 129,960
3 361x360x359 = 46,655,640
4 361x360x359x358 = 16,702,718,976
Given that there''s almost 17 billion possible combination of the first four moves it provides a compelling case against Deep Blue style search algorithms. Chess, by comparison, is a doddle.
Not that search style AI doesn''t have a whole host of uses but they''re limited to more feasible search spaces.
Anyhow, that''s beside the point. You wanted to play tic-tac-toe. The total number of possible games is factorial 9. This works out to 362880. Working through that search space would take your computer a fraction of a second, even if it were a 486.
So what I''m trying to say is that tic-tac-toe may be a forgone coclusion. Even if you got a neural net up to standard it would be forced into a draw by a simple search algorithm, guaranteed.
Chess then maybe? Not that someone hasn''t already suggested that (I just thought I''d give some reasons).
Although, if you did solve the Go problem and beat a first Dan player (the bottom level of the professional circuit, hardly a Kasparov), you would be in the money to the tune of $1,000,000.
Think about it.
Mike
This will highlight the problems of symbolic AI''s search methodologies, although I severly doubt you''ll get an NN system up to a good standard without long term genetic algorithm use.
I heard a quote that there was in the region of 10^390 possible games of Go. It''s played on a board of 19x19 squares. That''s 361 squares, meaning that in the first move you have 361 possible squares in which to place a piece (the board starts empty, you fill it up one square at a time, one player playing black, the other white).
Move Number of different possibilities
1 361 = 361
2 361x360 = 129,960
3 361x360x359 = 46,655,640
4 361x360x359x358 = 16,702,718,976
Given that there''s almost 17 billion possible combination of the first four moves it provides a compelling case against Deep Blue style search algorithms. Chess, by comparison, is a doddle.
Not that search style AI doesn''t have a whole host of uses but they''re limited to more feasible search spaces.
Anyhow, that''s beside the point. You wanted to play tic-tac-toe. The total number of possible games is factorial 9. This works out to 362880. Working through that search space would take your computer a fraction of a second, even if it were a 486.
So what I''m trying to say is that tic-tac-toe may be a forgone coclusion. Even if you got a neural net up to standard it would be forced into a draw by a simple search algorithm, guaranteed.
Chess then maybe? Not that someone hasn''t already suggested that (I just thought I''d give some reasons).
Although, if you did solve the Go problem and beat a first Dan player (the bottom level of the professional circuit, hardly a Kasparov), you would be in the money to the tune of $1,000,000.
Think about it.
Mike
Hey, this is the guy that''s been looking for a project idea. I''ve decided to do something on neural nets but 1) All of the docs I''ve read on neural nets are really really confusing and 2) I don''t know anything I could change about the neural net to produce a change in behavior. Reason 2 sorta proves how confusing all these docs about NNs are. Anyone wanna give me a quick summary of neural nets? (and PLEASE make it in plain english) Thanks
Author
I thought i might run into problems with the tic-tac-toe idea. Anyway, i''ll look into Go, sounds interesting. Any ideas on other games i could look into?
Ronald Forbes -
1) I know what you mean. Try www.generation5.org
2) No clue, i''m still a newbie at this.
Digital Radiation
Ronald Forbes -
1) I know what you mean. Try www.generation5.org
2) No clue, i''m still a newbie at this.
Digital Radiation
Hi
You dont need anything elese but GO
GO is the best game ever...its the universal best test of inteligence...
Its a known fact that the first man to make a program of GO that will beat a humman level 1 Dan has actually reinvented MAN...this goal is the ULTIMATE AI GOAL...
However GO is a very strange and complicated game, dont be fooled by its apparent simplicity...its the ultimate test for a man''s inteligence...so take care to understand GO at least at level 11 kyu before ever attempting to make Go program...
I have tried to do the same ... and i will try again...hope to make a good GO program before i die
All Go programs i have ever seen are EXPERT systems. they do a great job IMHO but they never fool a medium player not to speack of a 1 dan player....
however they have "advanced form recognition","strategic and tactical goals", "heuristic killers", "proverbs interpretation" and some other incredible things that chess programms never even consider...
Good luck
Bogdan
You dont need anything elese but GO
GO is the best game ever...its the universal best test of inteligence...
Its a known fact that the first man to make a program of GO that will beat a humman level 1 Dan has actually reinvented MAN...this goal is the ULTIMATE AI GOAL...
However GO is a very strange and complicated game, dont be fooled by its apparent simplicity...its the ultimate test for a man''s inteligence...so take care to understand GO at least at level 11 kyu before ever attempting to make Go program...
I have tried to do the same ... and i will try again...hope to make a good GO program before i die
All Go programs i have ever seen are EXPERT systems. they do a great job IMHO but they never fool a medium player not to speack of a 1 dan player....
however they have "advanced form recognition","strategic and tactical goals", "heuristic killers", "proverbs interpretation" and some other incredible things that chess programms never even consider...
Good luck
Bogdan
If you havent seen it yet, check the Machine Learning in Games site. It has a page on research done into using neural networks for 9x9 go.
If you want my opinion (and it is _occasionally_ worth having), give the network inputs (the board), outputs (one space on the board), constraints (you cannot put one piece on top of another) and _NOTHING_ else.
Then evolve it with the ability to adapt and play it against itself.
You''ll have to find the right network components, but it can be done. After all, we do it ;-).
Mike
Then evolve it with the ability to adapt and play it against itself.
You''ll have to find the right network components, but it can be done. After all, we do it ;-).
Mike
I true test of AI would be to let the programs play tic-tac-toe and wait to see which one gets "bored" and quits.
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
