This repository is the official implementation of [Dealing With Misspecification In Fixed-Confidence Linear Top-m Identification] (to appear in the proceedings of NIPS'21).

To install requirements (Python 3.6.9):

python3 -m pip install -r requirements.txt

In order to run ExperimentXXX in the paper, do as follows

• Run command

cd experiments_scripts/
./ExperimentXXX.sh

• That starts the computation, when it is done, the following files are present in the results/ folder

  • ExperimentXXX/method=[algorithm]_[list of options = values].csv

    Contains a matrix of 3 columns ("complexity": number of sampled arms, "regret": error in identification, "linearity": 1 if the algorithm considers data as linear, 0 otherwise, "running time": time in seconds for running the iteration) and XXX rows (controlled by parameter n_simu in the command) corresponding to each iteration of the algorithm.

  • ExperimentXXX/method=[algorithm]_[list of options = values]-emp_rec.csv

    Contains a matrix of XXX columns (number of arms in the experiment, controlled by parameter K in the command), and two rows, first row being the names of the arms, and the second one being the percentage of the time a given arm was returned in the set of good arms across iterations.

  • ExperimentXXX/params.json

    Saves in a JSON file the parameters set in the call to the code.

• PNG file ExperimentXXX/boxplot.png is created in folder boxplots/

You can only run the code to plot the boxplot from a previously run ExperimentXXX

• Run command

cd experiments_scripts
./ExperimentXXX.sh boxplot

ExperimentXXX won't be run, but if the corresponding results folder is present, then it creates the boxplot in folder boxplots/ExperimentXXX

Have a look at file code/main.py to see the arguments needed.

• Add a new bandit by creating a new instance of class Misspecified in file code/misspecified.py
• Add a new dataset by adding a few lines of code to file code/data.py
• Add new types of rewards by creating a new instance of class problem in file code/problems.py
• Add new types of online learners by creating a new instance of class Learner in file code/learners.py

Please refer to the paper.

All of the code is under MIT license. Everyone is most welcome to submit pull requests.