Code for the paper Jointly Efficient and Optimal Algorithms for Logistic Bandits, by Louis Faury, Marc Abeille, Clément Calauzènes and Kwang-Sun Jun.

Install

Clone the repository and run:

$ pip install .

Usage

This code implements the adaECOLog algorithms (OFU and TS variants) - both from the aforedmentioned paper, along with several baselines (oldest to newest):

• GLM-UCB from Filippi et al. 2010,
• OL2M from Zhang et al. 2016,
• GLOC from Jun et al. 2017,
• LogUCB1 from Faury et al. 2020,
• OFULog-r from Abeille et al. 2021.

Experiments can be ran for several Logistic Bandit (i.e structured Bernoulli feedback) environments, such as static and time-varying finite arm-sets, or inifinite arm-sets (e.g. unit ball).

Single Experiment

Single experiments (one algorithm for one environment) can be ran thanks to scripts/run_example.py. The script instantiate the algorithm and environment indicated in the file scripts/configs/example_config.py and plots the regret.

Benchmark

Benchmarks can be obtained thanks to scripts/run_all.py. This script runs experiments for any config file in scripts/configs/generated_configs/ and stores the result in scripts/logs/.

Plot results

You can use scripts/plot_regret.py to plot regret curves. This scripts plot regret curves for all logs in scripts/logs/ that match the indicated dimension and parameter norm.

usage: plot_regret.py [-h] [-d [D]] [-pn [PN]]

Plot regret curves (by default for dimension=2 and parameter norm=3)

optional arguments:
  -h, --help  show this help message and exit
  -d [D]      Dimension (default: 2)
  -pn [PN]    Parameter norm (default: 4.0)

Generating configs

You can automatically generate config files thanks to scripts/generate_configs.py.

usage: generate_configs.py [-h] [-dims DIMS [DIMS ...]] [-pn PN [PN ...]] [-algos ALGOS [ALGOS ...]] [-r [R]] [-hz [HZ]] [-ast [AST]] [-ass [ASS]] [-fl [FL]]

Automatically creates configs, stored in configs/generated_configs/

optional arguments:
  -h, --help            show this help message and exit
  -dims DIMS [DIMS ...]
                        Dimension (default: None)
  -pn PN [PN ...]       Parameter norm (||theta_star||) (default: None)
  -algos ALGOS [ALGOS ...]
                        Algorithms. Possibilities include GLM-UCB, LogUCB1, OFULog-r, OL2M, GLOC or adaECOLog (default: None)
  -r [R]                # of independent runs (default: 20)
  -hz [HZ]              Horizon, normalized (later multiplied by sqrt(dim)) (default: 1000)
  -ast [AST]            Arm set type. Must be either fixed_discrete, tv_discrete or ball (default: fixed_discrete)
  -ass [ASS]            Arm set size, normalized (later multiplied by dim) (default: 10)
  -fl [FL]              Failure level, must be in (0,1) (default: 0.05)

For instance running python generate_configs.py -dims 2 -pn 3 4 5 -algos GLM-UCB GLOC OL2M adaECOLog generates configs in dimension 2 for GLM-UCB, GLOC, OL2M and adaECOLog, for environments (set as defaults) of ground-truth norm 3, 4 and 5.