mcts-2048

A Monte Carlo Tree Search AI for the game 2048. Using bitboards and customizable strategies.

Compile:

git clone git@github.com:thomasahle/mcts-2048.git
cd mcts-2048/2048
javac -cp src/ isrc/dk/ahle/thomas/mcts2048/Main.java

Run:

java -cp src dk.ahle.thomas.mcts2048.Main

...

7598.17444219067 / 493
7610.59842519685 / 508
9 11 12 9
5 6 8 5
1 3 4 3
2 1 2 1

UCTStrategy Win%: 1.0, Avg: 7598.0, StdVar: 0.0, Min: 4096.0, Max: 4096.0, ms/m: 115.93982627007108

In this casse the maximum tile achieved was 4096. The sum of tiles on the board was 7598. Note the miniature board uses the log-2 base to display tiles, so 12 means 2^12 = 4048.

The default configuration above usually achives a max tile of 4,048 or 8,096, with 16,192 being a rare guest.

Options

The code supports multiple evaluation and rollout strategies. By default t uses the following:

test(new UCTStrategy(1000, true,
                new SumMeasure(),
                new GreedyStrategy(new EnsambleMeasure()
                        .addMeasure(-1, new SmoothMeasure())
                        .addMeasure(1, new FreesMeasure()))), 1);

Which means it does 1000 roll-outs before each move to find the most promising. Each roll-out is done using a greedy strategy, which combines various herustics, by default optimsing for smoothness and free squares. The eventual node is evaluated using the SumMeasure, which means taking the sum of all tiles. Alternatively one could use the maximum tile or the score achived.

Another quite good strategy is the CyclicStrategy, which simply repeats a certain sequence of moves:

test(new CyclicStrategy(Board.DOWN, Board.LEFT, Board.DOWN, Board.RIGHT), 1000);

It is also possible to use the CyclicStrategy for rollouts ini UCT:

test(new UCTStrategy(1000, true,
        new SumMeasure(),
        new CyclicStrategy(Board.DOWN, Board.LEFT, Board.DOWN, Board.RIGHT), 1);

Inspiration

The idea of using bitboards for faster move generation is inspired by https://github.com/nneonneo/2048-ai which uses min-max search.

For some state of the art papers, using n-tuple networks, see https://link.springer.com/chapter/10.1007/978-3-319-50935-8_8 and https://arxiv.org/pdf/1604.05085.pdf