We provide examples on shell command to run our experiment. For example, to run the CNN experiment, do:
python run.py --platform nn --nn_type cnn --alt true
The --platform nn flag tells the script that we are running neural network using PyTorch. the --nn_type cnn command tells the script that we are using CNN (using --nn_type ffn tells the script to use feaddforwad neural network, which triggers memory error. To use that, you also need to specify the --state_size to be 84 * 84 * 4). The --alt true command tells the script to use all 960K data points,
To use part of the data, use command like the following:
python run.py --platform nn --nn_type cnn --alt false --data_size 1000
To use SKlearn, use command like the following, and you need to manually define the sklearn model in run.py
python run.py --platform sklearn --data_size 1000
To introduce suboptimality, use command like the following. --suboptimal_type alter_actions tells the script that the suboptimality is random perturbation. --suboptimal_portion tells that the portion to perturb is 20%.
python run.py --platform nn --nn_type cnn --alt true --suboptimal_type alter_actions --suboptimal_portion 0.2
If the suboptimality is downsample, then do something like the following, where --downsample_action 0 tells the model to downsample action 0
python run.py --platform nn --nn_type cnn --alt true --suboptimal_type downsample --suboptimal_portion 0.2 --downsample_action 0
The other arguments in the run.py are mostly model hyperparameters, and they should be self-explanatory (e.g., bs stands for batch size). You typically do not need to modify them.