improved-funicular

Python 3 implementation of the 2048 game.

Performance

Here's how various 2048 strategies perform (tested on a 2012 macbook air with a 1.8GHz i5 and 4GB of 1600MHz DDR3 RAM and run using python do_stats.py nick 100 [strategy_name] ):

Strategy Name	Max tile	Max Score	Mean Score	Score Standard Dev	Mean steps per game	Mean sec per game	Git Hash
only_go_right	16	140	13	18	6	.00005 sec	b51ed7d
random	256	3064	1058	504	138	.006 sec	b51ed7d
down_left	512	7228	2332	1131	207	.05 sec	b51ed7d
fixed_action_order	512	4864	2492	1345	221	.05 sec	b51ed7d
greedy	512	5404	2122	930	188	.05 sec	b51ed7d
greedy_fixed_order	1024	11376	3028	1580	256	.06 sec	b51ed7d
down_left_greedy	512	7084	2107	1076	192	.05 sec	b51ed7d
max_space_then_greedy	1024	12320	3157	1477	266	.06 sec	b51ed7d
lookahead_1	1024	11940	3008	1864	252	.1 sec	b51ed7d
lookahead_2	1024	15428	7446	3157	491	.7 sec	b51ed7d
lookahead_3	2048	25868	11659	3958	724	5.1 sec	b51ed7d
lookahead_4	2048	35500	16757	7443	945	33.1 sec	02b412d
lookahead_with_rollout1	4096	75832	36670	14124	1853	758 sec	c3bb225

1. Lookahead 3 that switches to 150 random rollouts per move when board has <= 7 empty spaces

Setup

Developed under Python 3.7 on OS X Catalina with pip and virtualenv installed on the system. Look here to setup Python 3.7.

• git clone https://github.com/nickjalbert/improved-funicular.git
• cd improved-funicular
• virtualenv improved-funicular
• echo `pwd`/ > improved-funicular/lib/python3.7/site-packages/curr_dir.pth
• source improved-funicular/bin/activate
• pip install -r requirements.txt
• python play_in_console.py nick or python play_in_console.py andy

Dev and Testing

To automatically format your code, run:

• black [filename.py]

To lint, run:

• ./scripts/lint

To test, run:

• pytest

Example strategy trials

Use do_stats.py to try different 2048 game strategies. Run:

python do_stats.py

for detailed usage. An example running 100 trials with Nick's implementation of a random strategy:

python do_stats.py nick 100 random

Manually play 2048

Use python play_in_console.py nick to manually play 2048 in the console.

Use python play_in_console.py lookahead to play with suggested moves from the lookahead strategies.

Using MLflow

Some of the algorithms (e.g. reinforce.py) use MLflow to track experiment runs. By default, MLflow writes experiment metadata to the local filesystem. To visualize the results in the MLflow UI, just run mlflow ui in the terminal and then visit localhost:5000 in your browser.

There you can see the history of runs of the algorithms, including parameters and metrics associated with each run.