gym-contin

Gym-style API

The domain features a continuos state and a dicrete action space.

The environment initializes:

• cross-sectional dataset with variables X_a, X_s, Y and N observations;
• logit model fitted on the dataset, retrieving parameters \theta_0, \theta_1, \theta_2;

The agent:

• sees a patient (sample observation);
• predict his risk of admission \rho, using initialized parameters
• if \rho < 1/2:
  • do not intervene on X_a, which stays the same
• else:
  • sample an action a in [0,1]
  • compute g(a, X_a) = newX_a
  • intervene on X_a by updating it to newX_a
• give reward equal to average risk of admission, using predicted Y, initial parameters and sampled values

(shouldn't I fit a new logit-link? parameters are now diff?)

To install

• git clone https://github.com/claudia-viaro/gym-contin.git

• cd gym-contin

• !pip install gym-contin

• import gym

• import gym_contin

• env =gym.make('contin-v0')

To change version

• change version to, e.g., 1.0.7 from setup.py file
• git clone https://github.com/claudia-viaro/gym-contin.git
• cd gym-contin
• python setup.py sdist bdist_wheel
• twine check dist/*
• twine upload --repository-url https://upload.pypi.org/legacy/ dist/*