Schafkopf_RL

Reinforcement Learning applied on the Bavarian card game 'Schafkopf'

Approach:

Player 1 (RL bot) is learning, while players 2-4 are acting random. Afterwards update players 2-4 so they act according to the Q Network while player 1 is learning again.

Set up:

module	content
rules.py	definition of cards, scores, games, rewards and helper methods
state_overall.py	states which are valid for every player
state_player.py	states which are valid for single player
train_select_game.py	hyperparameters for NNs, training Q Networks
train_select_cards.py	hyperparameters for NNs, training Q Networks
QL_select_game.py	Q Network architechture and Memory for selecting game
QL_select_card.py	Q Network architechture and Memory for selecting cards
interface_to_states.py	managing states for each player and overall states
script.py	script
helper_functions.py	e.g. plot functions

Next steps:

Action
set state davongelaufen	identify if player davongelaufen in play()
QL for selecting cards
QL for doppeln
Stock
select game process simplified right now
Tout, Sie
Contra

Results:

Reward ~ epochs after implementing QL for choosing game

Player 1 = choosing game based to dealed cards as a state and NN, but playing cards random
Players 2-4 = acting completely random