Skip to content

qpwoeirut/2048-solver

Repository files navigation

2048 Solver

Demo

Website

To watch the AI yourself, go to this website or check out this repository.

Selenium Remote Controls

You can also run this with Selenium, although that's not recommended. The instructions for that are here.

Summary

This project is a set of AI strategies that play the popular 2048 tile-merging game. See the roadmap for details on the project.

Although most of the ideas I'll use are taken directly from others' projects, I want to try writing most of the source code myself. There are some implementation ideas that I'm taking as well (such as the bitwise board transposition code), but most of the code should be my own.

Structure

The code for simulating a game is in the GameSimulator class under game.hpp. util.hpp stores helpful utilities for the heuristic and player functions.

Each strategy is in the strategy directory. All strategies implement a function which provides a move when given a board. Some strategies have parameters and heuristic functions or secondary strategies that are passed in. All heuristics are in heuristics.hpp.

tester.cpp simulates games for each solver and write the results into the results directory as a CSV file. Giving each solver its own file means that I don't have to rerun every solver simulation if I only need to test one solver. These CSV files are then combined by collate.py. In the future I might write a program to plot/visualize the data, especially for comparing different parameters on certain solvers. For now, the best visual I have is a Google Sheet with conditional formatting.

Testing

Each test runs at least 500 games in order to try and minimize random variance. In general, the number of games run is increased until it takes a least a half minute to complete. As a result, some of the faster solvers (such as the random strategy) run hundreds of thousands of games.

All game tests are run in parallel using C++'s std::async.

For Stages 1 and 2, games were run on an AWS EC2 Amazon Linux c6g.large instance. From Stage 3 onwards, games were run on an AWS EC2 Ubuntu c6g.xlarge instance.

Results

The best strategy right now is the expectimax strategy with the corner building heuristic. On the most recent set of tests, it reaches 4096 99.4% of the time, gets the 8192 tile with a 91.8% success rate, and reaches 16384 in 34.6% of its games.

The results file has the latest statistics for all tested strategies.

Caveats

The Google sparsehash repository needs to be installed.

About

A set of AIs for the 2048 tile-merging game. Includes an expectimax strategy that reaches 16384 with 34.6% success and an ML model trained with temporal difference learning.

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published