Simple Q-learning agent for pong.
src
├── game.py # game logic (using pygame)
├── main.py # parse args and run
└── world.py # main module: Q-learning + game interactions
python src/main.py \
--canvas_size 32 24 \
--paddle_length 7 \
--velocity 1 \
--epsilon 0.03 \
--learning_rate 0.7 \
--discount 0.99 \
--train_episodes 50000 \
--eval_episodes 10 \
--eval_every 200 \
--max_iter 1000 \
--agent_strategy eps_greedy \
--opponent_strategy almost_perfect --alpha 0.8 \
--load --filename q-values/32_24_7_003_07_50k_eg_ap_08.pTo train with custom parameters, omit the --load --filename <file> args.
You can also use --plot_scores to plot train and eval scores, after training.
Using --filename <file> (without --load) will save final Q-values to the specified file,
as a pickle dictionary.
Pickle file naming convention:
<w>_<h>_<paddle_len>_<eps>_<lr>_<train_episodes>_<agent_strategy>_<opponent_strategy>
