The code used for training this agent is available in a seperate repo

2048 AI

The best AI Controller for the puzzle game 2048 (as of March 2017). Its score depends on the search settings:

Search limit	Average score	32768 [%]	16384 [%]	8192 [%]	Games	Moves/s
1-ply	324710 ± 11043	19	68	90	1000	258371
2-ply	457079 ± 11112	34	91	99	1000	20524
3-ply	511759 ± 12021	50	92	99	1000	1484
5-ply	545833 ± 21500	54	97	100	300	16
1 ms	527099 ± 11486	54	95	100	1000	916
50 ms	576655 ± 20839	62	97	99	300	20
100 ms	589355 ± 20432	65	96	100	300	10
200 ms	591380 ± 21870	67	97	99	300	5
1000 ms	609104 ± 38433	69	97	98	100	1

Algorithm

The program uses expectimax with an n-tuple network for state evaluation function, which has been learned from scratch using a new variant of temporal difference learning. The learning method is described in:

Wojciech Jaśkowski, Mastering 2048 with Delayed Temporal Coherence Learning, Multi-State Weight Promotion, Redundant Encoding and Carousel Shaping, IEEE Transactions on Computational Intelligence and AI in Games (accepted) (http://arxiv.org/abs/1604.05085)

which extends our previous work:

Marcin Szubert and Wojciech Jaśkowski, Temporal Difference Learning of N-Tuple Networks for the Game 2048, Proceedings of IEEE Conference on Computational Intelligence and Games, pp. 1-8, August 2014, Dortmund, Germany, (preprint)

Authors

• Wojciech Jaśkowski (the learning algorithm)
• Adam Szczepański (this code)

Requirements

• C++11 compiler
• CMake 3.0+
• Boost 1.49.0+ (program_options, accumulators)

If you want to use the web application you will also need:

• Python 2
• Chrome or Firefox

Tested on:

• Mac OS X 10.11.1, Apple LLVM version 7.0.2
• Ubuntu 15.04 64-bit, g++ 4.9.2
• Windows 7 64-bit, Visual Studio 14 2015

Evaluation function files

The program requires a file with an evaluation function. Some basic (very poor, but small) evaluation functions are already in data/2048_strategies/ directory. However, for the state-of-the-art results download and unzip our best evaluation function.

Building and Running

Console application

Linux/OS X

1. Go to the project root directory
2. Run cmake .
3. Run make. The lib/ and bin/ directories will be created.

Windows

1. Go to the project root directory
2. Create and go to _build/ directory
3. Run cmake .. -G "Visual Studio 14 2015 Win64". You may have to set boost directories: set BOOST_ROOT=your_boost_root_directory and set BOOST_LIBRARYDIR=your_boost_library_directory first
4. Open the solution and build the release version of it

Examples

For running multiple games and testing the AI's capabilities

• 1 game, max depth 1, single thread:

./bin/main --strategy data/2048_strategies/eval-function.bin.special --uncompress false

• 10 games, max depth 3, uncompressed model (faster, but requires more RAM):

./bin/main --strategy data/2048_strategies/eval-function.bin.special --games 10 --depth 3

• 10 games, max 100ms per move, uncompressed model, multithreating in expectimax (best reasonable settings):

./bin/main --strategy data/2048_strategies/eval-function.bin.special --games 10 --depth 100 --time 100 --eval_multithreading

• 100 games, max depth 5, uncompressed model, playing 4 games in pararell:

./bin/main --strategy data/2048_strategies/eval-function.bin.special --games 100 --depth 5 --threads 4

• 10 games, min depth 1, max depth 8, 1 game thread, multithreading in expectimax evaluation, max time 50 ms per round, uncompressed model, prints boards' states to the console:

./bin/main --strategy data/2048_strategies/eval-function.bin.special --games 10 --depth 8 --time 50 --eval_multithreading  -v

Web application

For observing how the AI works on the 2048 game site

You may need to install the websocket-client first:

pip install websocket-client

Chrome

1. Close all instances of Chrome (including hangouts, etc...)
2. Run the browser with remote-debugging enabled:

   chrome --remote-debugging-port=9222
   

3. Go to 2048 game site
4. Run python script (see examples)

Firefox

1. First you need to install Remote Control plugin
2. Go to 2048 game site
3. Start remote debugging (click on the plugin's icon)
4. Run python script (see examples below)

The Web application was strongly based on the code from https://github.com/nneonneo/2048-ai.

Examples

• chrome, max depth 4, multithreading, no time limit, uncompressed model:

python 2048.py -b chrome --strategy data/2048_strategies/eval-function.bin.special --depth 4 --multithreading true

• On Windows you will need to specify WebApi library file:

python 2048.py -b chrome --lib _build/lib/Release/WebApi.dll

Usage

Console application

• --strategy arg - strategy input file (by default data/2048_strategies/2048_a_weak_player.bin.txt)
• --uncompress arg - uncompress strategy (true/false, by default true). Faster but requires more RAM.
• --seed arg - random seed (by default based on time elapsed since epoch)
• --games arg - number of games (by default 1)
• --time arg - maximum time for one round [ms], 0 means no time limit (by default 0)
• --depth arg - maximum depth for expectimax (by default 1)
• --threads arg - number of threads (each thread plays different game, by default 1)
• --eval_multithreading - enable multithreading in expectimax algorithm
• -h [ --help ] - produce help message
• -v [ --verbose ] - show board and score after each round

Web Application

• -b [ --browser ] - choose browser (Chrome or Firefox, by default Firefox)
• -p [ --port PORT ] - port number to control on (default: 32000 for Firefox, 9222 for Chrome)
• --strategy arg - strategy input file (by default data/2048_strategies/2048_a_weak_player.bin.txt)
• --uncompress arg - uncompress strategy (true/false, by default true). Faster but requires more RAM.
• --time arg - maximum time for one round [ms], 0 means no time limit (by default 0)
• --depth arg - maximum depth for expectimax (by default 1)
• --multithreading arg - enable multithreading in expectimax algorithm (true/false, by default true)