Finite-Time Frequentist Regret Bounds of Multi-Agent Thompson Sampling on Sparse Hypergraphs (ϵ-MATS)

[AAAI 2024 Oral]

Tianyuan Jin^* · Hao-Lun Hsu^† · William Chang^‡ · Pan Xu^†

^* National University of Singapore · ^† Duke University · ^‡ University of California, Los Angles

Official implementation of the paper "Finite-Time Frequentist Regret Bounds of Multi-Agent Thompson Sampling on Sparse Hypergraphs (ϵ-MATS)" which combines the MATS exploration with probability ε and greedy exploitation with probability 1 − ε.

Installation instructions

Dependencies

• python==3.6
• scipy >=1.2.1
• matplotlib >= 3.0.2
• pandas >= 0.25.3
• numpy >= 1.17.0

Example

# Enter the anaconda virtual environment
source activate epsilon_mats
# Train on Bernoulli0101 using random exploration on 10 agents
python main.py --algo rd --env_name bernoulli --iter 2000 --seed 0 --n_agents 10

# Train on Poisson0101 using mats (including different epsilon) on 20 agents
python main.py --algo all --env_name poisson --iter 2000 --seed 0 --n_agents 20

Citation

@inproceedings{Jin2024MATS,
  title={Finite-Time Frequentist Regret Bounds of Multi-Agent Thompson Sampling on Sparse Hypergraphs},
  author={Jin, Tianyuan and Hsu, Hao-Lun and Chang, William and Xu, Pan},
  booktitle={Annual AAAI Conference on Artificial Intelligence (AAAI)},
  volume={38},
  number={11},
  pages={12956--12964},
  year={2024}
}