Deep Q-Learning Project to play 2048, inspired from georgwiese/2048-rl. See this presentation for an introduction.
The game class is stored in cloudmercato/2048-game.
pip install -r requirements.txt pip install https://github.com/cloudmercato/2048-game/archive/refs/heads/master.zip python setup.py install
The main way to use this repository is its CLI 2048-rl:
$ 2048-rl --help
usage: 2048-rl [-h] [--runs RUNS] [--games-per-cycle GAMES_PER_CYCLE]
[--batch-size BATCH_SIZE] [--mem-size MEM_SIZE] [--lr LR]
[--gamma GAMMA] [--gamma1 GAMMA1] [--gamma2 GAMMA2]
[--gamma3 GAMMA3] [--epsilon EPSILON]
[--epsilon-min EPSILON_MIN] [--epsilon-dec EPSILON_DEC]
[--model-path MODEL_PATH] [--model-load-file MODEL_LOAD_FILE]
[--model-save-file MODEL_SAVE_FILE]
[--training-epochs TRAINING_EPOCHS] [--model-disable-autosave]
[--disable-collect-random-data] [--refill-episode-db]
[--log-dir LOG_DIR] [--tf-log-device] [--tf-dump-debug-info]
[--tf-profiler-port TF_PROFILER_PORT] [--verbose VERBOSE]
[--tf-verbose TF_VERBOSE]
[{train,infer}]
positional arguments:
{train,infer}
optional arguments:
-h, --help show this help message and exit
--runs RUNS Number of runs
--games-per-cycle GAMES_PER_CYCLE
Number of games per model learn run/cycle
--batch-size BATCH_SIZE
Training batch selection size (in number of episodes
--mem-size MEM_SIZE Learning episode DB size (in number of episodes
--lr LR Learning rate
--gamma GAMMA Gamma
--gamma1 GAMMA1 Gamma1
--gamma2 GAMMA2 Gamma2
--gamma3 GAMMA3 Gamma3
--epsilon EPSILON Epsilon
--epsilon-min EPSILON_MIN
Epsilon - start value
--epsilon-dec EPSILON_DEC
Epsilon - step value
--model-path MODEL_PATH
Python path to the model to compile
--model-load-file MODEL_LOAD_FILE
Model load file path (h5)
--model-save-file MODEL_SAVE_FILE
Model save file path (h5)
--training-epochs TRAINING_EPOCHS
Number of epoch rns for every model training run
--model-disable-autosave
--disable-collect-random-data
--refill-episode-db
--log-dir LOG_DIR Tensorboard log directory
--tf-log-device Determines whether TF compute device info is
displayed.
--tf-dump-debug-info
--tf-profiler-port TF_PROFILER_PORT
--verbose VERBOSE, -v VERBOSE
--tf-verbose TF_VERBOSE, -tfv TF_VERBOSE
Basically you can:
- Train (train) a model, record it, reload an existing one
- Solve (infer) game with an existing model