Railroad Ink is a game where you draw roads and railways on a grid. You get points for connecting exits, creating unbroken roads and railways, using the center of the board, and you lose points for open paths. The game is played over 7 rounds, and at the start of every round you roll dice to see the roads and railways you must draw.
This is a program that plays that game. It uses Monte Carlo Tree Search to evaluate moves, and a neural network to evaluate the board state. It uses tree pruning to reduce the size of the tree it needs to search. I did try to use Rapid Action Value Estimation to evaluate the board state, but I think my implementation was bad, and it didn't work very well.
Inspired by Tiltak, an AI for the board game Tak.
Let the AI play some Railroad Ink. By default, it will play a single game, spending a second to evaluate each move. The moves will be printed to the console.
--count - The number of games to play. Default is 1.
-d, --duration - The time to spend evaluating each move, in seconds.
-i, --iterations - The number of iterations to run the MCTS algorithm. If you specify this, the program will search the tree for the specified number of iterations, instead a specified duration.
Train the AI. Sorta. This is all just a pile of things that are vaguely useful for me as I go along.
-t, --train - A boolean flag that determines if the program should train the neural network it uses to evaluate moves. Defaults to true.
-g, --generate-training-data - A boolean flag that determines if the program should run some games to create more data to train on. Defaults to false.
-l, --loop-training - This boolean flag tells the program to never stop training.