This is an application of temporal difference learning for the game 2048. The current repo is in need of clean up (TODO). This implementation uses linear functions of state features and neural networks to approximate the state value function for TDL. Since it uses a separate function or network for each action, it is akin to Q-learning.
Without CUDA
- Python 3.7.2
- TensorFlow 2.2.0 :
python -m pip install tensorflow==2.2.0 - NumPy 1.17.2 :
python -m pip install numpy
With CUDA
- Python 3.7.2
- TensorFlow-gpu 2.2.0 :
python -m pip install tensorflow-gpu==2.2.0 - NumPy 1.17.2 :
python -m pip install numpy - CUDA 10.1 (10.2 probably won't work!)
- CUDA Toolkit (forgot version, 10.1 probably)
- CUDA capable GPU
I have not tested it without CUDA enabled so I cannot make any guarantees about the functionality of it.
Installing with the pip commands above should work but I cannot guarantee it.
An old version with linear parameterized functions for value function approximation is found in 2048-linear_parameterized. The current root contains the neural network value function approximaters.
For both versions, to run just do python q-learn.py. Files may take some manual attunement to work how you
might expect (I did not ensure correctness of files after I finished getting the data I needed).
NOTE: Do not run the current q-learn.py expecting it to terminate. It will take on the order of days
or weeks to finish. You can still run it to see some potential scores of the agent. At any rate you can always
hit ctrl+C to terminate.