This is the test code for experimental results in the paper Best Arm Identification (BAI) with Safety Constraints (Zhenlin Wang, Andrew Wagenmaker, Kevin Jamieson, Best Arm Identification with Safety Constraints. Under Review for AISTATS, 2022. ArXiv, 2021, arXiv:2111.12151. ).
Here we outline 2 major algorithms to identify the best arm within the multi-arm bandit setting without violating a specified safety constraint.
Safe-Linear is tailored for linear reward and safety models, while Safe-Monotonic is applied on monotonic reward and safety models.
For linear case, you should specify a bandit instance by using the MAB class from CBAI, then run the Safe-Linear on this instance generated. Otherwise, you may generate any random instance using the instance_generator from utility.py
You may call any of the functions above via:
from CBAI import MAB, Algo, instance_generator
Next, you may try out monotonic models by either using our sigmoid function or your own function (make sure its monotonicity is well behaved). Construct a particular instance following the format outlined in our notebook. Then you may call
from CBAI import safe_monotonic_bai.
Feel free to play around with the code if you need it for your research purpose. A simple and fun way to learn about the code is to run the ipython notebook here.