Getting started with CI code

[mollyow]

Getting started with GitHub

• [x] Clone repository locally. Consider using the free version of fork for git management.
• [x] Create a new branch. Make a minor change on this branch. Push it to the repository. Create a Pull Request, and ask for review from @mollyow.

Getting started with bandits code

• [x] Run through TS_experiment.R script to run a simulated experiment, and start assessing results. Discuss with @mollyow.
• [x] Next steps will be to work on testing the code, discussed in more detail in #7. @YinghuiZhouu will work on the data generation functions; @AmirWilliams will work on testing.

[AmirWilliams]

Variables:

(xs):Covariates. All the variables that predict outcomes. Features of individuals. Associates with ys. There will be a matrix as long as A and as wide as p. xs affect ys

• ex. Age, diet , gender, polical party
• Contextual experiments include in xs in the argument. Non-contextual experiments do not include xs.
• xs is a matrix of covariates

muxs: Output if there were no noise. Shape A by k. Xs and Ys are related. If everyone were the genetically identical, then they would have the same muxs, but not necessarily the same ys.

underlying truth

(ys): Potential outcomes. Shape A by k. Muxs plus random noise.

(yobs) This is the output that researchers ultimately evaluate. More noise

(A): Number of subjects (1000, usually)

(P): Number of covariates (5, in current version of code)

(K): Number of variables. In this case treatments (4, in current version of code)

[mollyow]

Break this into arguments that go into line 54, and the sub-components of the results object.

[AmirWilliams]

Assessing the arguments and outputs of: results <- run_experiment(ys, floor_start, floor_decay, batch_sizes, xs)

• Arguments:
  • run_experiments: function that contains 5 arguments
  • ys: Potential outcomes. Shape A by k. muxs plus random noise.
  • floor_start: 5, jointly with floor_decay, describes how we set probability floors
  • floor_decay: 0.9, jointly with floor_decay, describes how probability floors decay.
  • batch_sizes: 100, 200, 300, 400 for a total of 1000. Batches represent number of individuals being tested (sum to A)
  • xs: Matrix of covariates as long as A and as wide as p. xs predicts ys. Its inclusion indicates that this is a contextual experiment as factors that vary are considered among individuals.
• Sub-components of results:
  • yobs: The outcomes that researchers ultimately evaluate. More noise.
  • ws: treatments assigned
  • xs: Matrix of covariates as long as A and as wide as p. xs predicts ys. Its inclusion indicates that this is a contextual experiment as factors that vary are considered among individuals.
  • ys: Potential outcomes. Shape A by k. muxs plus random noise.
  • probs: matrix A x A x K, representing time x context x treatment condition. Probabilities of treatment assignment
  • fitted_bandit_model: the model used in the algorithm
  • num_mc: Internal parameter: number of "monte-carlo" simulations. Number of repeated simulations used for calculating probabilities.
  • K: Number of treatments. In this case treatments (4, in current version of code)
  • p: Number of covariates (5, in current version of code)
  • floor_start: same as above
  • floor_decay: same as above
  • y: list of length K; Represents outcomes of subjects assigned to each treatment
  • ps: list of length K; Represents probabilities of subjects being assigned the treatment they received, under each different version of treatment
  • mu: matrix K rows by (p+1) columns; means used by model
  • V: (p+1) sets of data, each with (p+1) column by K rows; variance used by model
  • X: list of length K; Represents covariates of subjects assigned to each treatment
• Warning Message
  • 1: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per fold
  • 2: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per fold
  • 3: Option grouped=FALSE enforced in cv.glmnet, since < 3 observations per fold
• Additional
  • mu_hat
  • balwts