MAKHFI.COM
Fascinating World of Neural Nets  
Neural Network Forums
Home PageHome Page : Neural Networks for Pattern Recognition : multi-player multi-skill game model
  You are currently not logged in. You can view the forums, but cannot post messages. | HOME | Log In | Register | Search | Help
Post a Reply on This Topic Post a Reply on This Topic

Author Topic: multi-player multi-skill game model
zerocred Posted: 28-Jan-11 08:26
  Edit Edit
 
Email the Author Mail   View Author's Profile Profile  
Hi how to model this problem...

From a population of beauty contestants each with, say, 4 characteristics e.g. height, weight, hair colour and fitness...

We pick about 10 at random who will play-off and a winner is selected for relative beauty - (depending on whatever criteria the judges were using of course).

Then another ~10 or so contestants (some might have picked before) and do the same... and over and over again.

How to train NN to pick the winner if we have the results of past games?

Do we train by applying all parameters of each contestant simultaenously (~40 inputs) to NN and randomize the order (by permuting).

Or, do we test each player individually and compare the NN outputs score with other contestants output scores...

The problem is, because the winner of this game depends on the relative asttributes of the other contestants I think we need to apply all paramaters to NN but with 10! different combinations of inputs (4 characteristics remain in same order) just for one play-off it might take a while to train. As it would be necessary for each players hair colour - for example to have the same weighting, because we don't know what order the contestants will be presented in.

Or, put another way - if we always presented the known winner in inputs 1,2,3,4 player 2 in NN inputs 5,6,7,8 then the NN would quickly work out the winner was always position 1 - cos we trained it that way. But if we randomized the inputs and put the same competitors in a differetn order then it would remove that bias - but 10! different training combinations to present for just a singel playoff...

Doing each contestant alone does not assess his or her relative strengths to the other contestants...

What would the NN look like? Anyone come across this problem before?

Or consider a simple card game where two players draw a card - highest card wins (ace = highest) except that an ace can only be beaten by a 3 (actually loosely based on the poker like game, Brag) if the competitor is not an ace then 3 reverts to its normal value and gets beat if the competitor has 4 - K. If the NN was trained with each hand in isolation it would never realize 3 beats an Ace - but if it could be trained with both cards at same time then it /might/ just figure it out. Its just that when the number of players holding several cards goes up so does the permutations leading to impossible to train NN - or maybe not...

Ideas & comments welcome....
Z
 
pejman Posted: 28-Jan-11 10:17
Delete Delete    Edit Edit
 
Email the Author Mail   View Author's Profile Profile  
Z,

the short answer is to train the network with all the parameters simultaneously (otherwise the correlation won't happen).

The most important issue will be the encoding mechanism of parameters (characteristics).

Feel free to send me some details and I might be able to help.

Best,
-Pejman

 
zerocred Posted: 29-Jan-11 05:14
Delete Delete    Edit Edit
 
Email the Author Mail   View Author's Profile Profile  
Thanks - do you have an email address you can share with me?
My email zerocredibility at gmail .

Basically I think the input weightings of the players should be the same for the same parameter - for my example there should be 10 identical sets of input weightings (height and fitness should have same weighting for each contestant). But how to train something like that?
 

Post a Reply on This Topic Post a Reply on This Topic
 

Copyright © 2001-2003 Pejman Makhfi. All rights Reserved.