Problem Statement
The goal of this project is to design a neural network that can learn to play the Swedish card game Seven (Sju:an in Swedish). The network will produce evaluations of the given game state using regular forward propagation. However, the network will train itself using a modified version of the Temporal Difference algorithm where all previous game states are incorporated in the back propagation of the neural network after each play.
Rules for 3-Player Seven
Take a standard deck of 52 playing cards and shuffle them. The dealer then deals 17 cards to each player, starting to his/her left. There will be one card left that goes the player who holds the Seven of Hearts. The Seven of Hearts is then played on to the table and the game play continues clockwise.
Valid moves are either playing a Seven of the remaining three suits (7S, 7D, or 7C) or building either up (6->A) or down (8->K), in the same suit, from a previously played Seven. If a player has no moves to make, every other player gives that player a card.
The game ends win a player runs out of cards or on his/her turn can play every card in his/her hand aka "going out".