Femto Ai

;;; -- scheme --

(use-modules (haunt utils) (ice-9 rdelim))

`((title . “Fembot: A naive AI for the game of femto”) (date . ,(string->date* “2016-07-31 15:40”)) (tags “project” “AI” “learning” “lua” ) (summary . “An AI for a simple card game”) (project . “fembot”) (content ( (h3 “Background”) (p “ Senior year of highschool I was introduced to a simple card game called “ (a (@ (href “http://nrich.maths.org/1179”)) “Femto”) “ by a good friend. The game is so simple, that with a little practice you can even play it in your head with no cards at all. An interesting property arises as a result of this simplicity. The game is simple enough that a person can reason about the possible outcomes of a decision fairly effectively and in a reasonable period of time. However, it’s not so simple that this is easy to to without a pen and paper though which makes it an excellent game of thought. “)

(h3	"Game/Rules")
(p	"
	There are eight cards in the deck: 2,3,4,5,6,7,8, and 10. Each player is dealt
	four cards. Each round both players play a card face down. The cards are
	revealed and the higher card wins, unless the lower card is less than half of
	the value of the higher card. The player of the winning card chooses one of
	the two cards and keeps it in front of him. The other player places the
	remaining card in his hand. When one player has no cards left the game is over.
	The player with the greatest total value of cards in front of him at the end of
	the game is the winner. 
	")
	
(h3	"How it Works")

(p	"
	This is a very simple AI. It was designed mostly so that
	I had something to play against and only needed to be good enough to beat me.
	Ideally an AI for this game would use something like Minimax/Alpha-Beta pruning.
	Even better would be working out the game theory and \"solving\" the game. Currently
	the AI just looks at every possible outcome that the game can take and chooses the one
	that has the most branches that win. It's sort of like Monte-Carlo/Rollouts, but
	there's nothing random about it. It's simple enough to try everything, so it does.

	The AI also has code for training and playing against a neural network using lua-fann.
	There are two networks. One for choosing which card to play and one for choosing which
	card to take if the round is won. The neural networks didn't pan out and I was too busy
	playing against my simple AI to fiddle with them to get them working. Perhaps if I come
	back to this project I'll fix up the neural-nets or implement a better algorithm.
	")

(h3	"Give it a Try")

(p	"You'll need lua-fann and lua5.2 or luajit to make this work."
  (pre
  		"git clone "
	(a (@ (href "https://github.com/rbryan/fembot"))
		"https://github.com/rbryan/fembot\n")
 		"cd fembot\n"
	"lua fembot.lua"))

)))

Written on July 31, 2016