Blablaland Powa Index du Forum
Blablaland Powa
BIENVENUE SUR LE FORUM DE BLABLALAND,ICI VOUS POURREZ PARLER DE VOUS,PRESENTEZ VOS SKIN ET MEME GAGNER DES CODES!! ATTENTION: IL FAUT VOUS ENREGISTRER POUR POSTER ET PARTICIPER
 
Blablaland Powa Index du ForumFAQRechercherS’enregistrerConnexion

:: Problem discussion: Card Game ::

 
Poster un nouveau sujet   Répondre au sujet    Blablaland Powa Index du Forum -> Les mini-jeux du forum -> Jeux de cartes
Sujet précédent :: Sujet suivant  
Auteur Message
JoeClark


Hors ligne

Inscrit le: 18 Oct 2017
Messages: 90

MessagePosté le: Lun 18 Déc - 15:16 (2017)    Sujet du message: Problem discussion: Card Game Répondre en citant

Hello,


I am new to fann but not to anns at all. I am currently working on a trick based card game for which I develop an AI based on reinforcement learning. The card game uses 24 different cards, each twice, makes 48 cards. I represent the problem by an high dimensional input vector, e.g. 125 dimensions. In this case I have 48 dimensions for each card in hand (1 = I have this card on my hand, 0 I dont). The next 24 dimensions represent the card I am about to play (so out of the 24 inputs there is 1 one and 23 zeros) the next 48 cards are the cards in the middle (0 to 3 ones, rest zeros) and the last five inputs represent the color that the first card had (necessary to determin which card wins the trick because you have to play the same color if you can, else your card is usually worthless). I train the net on the reward that is generated from that round in the game (it has 12 rounds) til the end. There are 240 points in the game and I normalize the result to the maximum points that can be made in the remaining rounds (40 is max per round). 


I hope I made the theory clear. 
Now for the difficult part: 


I have 3 testplayers which just play random cards that are valid in that situation, thats my benchmark. My net usually wins up to 62% of the games. On the other hand I trained an SVM instead of an ann and it wins up to 70% of the games. I wonder if I could make it possible to improve my representation or my use of FANN to improve the success of my nets. I tried different layers, different neurons per hidden layer, different activation and error functions but I never got close to the SVM. 


Maybe someone of you has an idea to improve my solution? 


please help


Thanks !


I didn't find the right solution from the internet.


References:
http://leenissen.dk/fann//forum/viewtopic.php?f=1&t=1343&view=next


Promotional Video Campaign


Revenir en haut
Publicité






MessagePosté le: Lun 18 Déc - 15:16 (2017)    Sujet du message: Publicité

PublicitéSupprimer les publicités ?
Revenir en haut
Montrer les messages depuis:   
Poster un nouveau sujet   Répondre au sujet    Blablaland Powa Index du Forum -> Les mini-jeux du forum -> Jeux de cartes Toutes les heures sont au format GMT + 2 Heures
Page 1 sur 1

 
Sauter vers:  

Portail | Index | Panneau d’administration | créer forum | Forum gratuit d’entraide | Annuaire des forums gratuits | Signaler une violation | Conditions générales d'utilisation
HalloweenOclock © theme by larme d'ange 2006
Powered by phpBB © 2001, 2005 phpBB Group
Traduction par : phpBB-fr.com