Federated Multi-armed Bandits

This is the package of codes and datasets used in paper ''Federated Multi-armed Bandits'', which is accepted to AAAI 2021.

The files ''Fed1_UCB_CR.py'', ''Fed2_UCB_CR.py'' and ''Fed2_UCB_CR_short.py'' are for the simulations of cognitive radio systems with the synthetic datasets. ''Fed1_UCB_RS.py'' and ''Fed2_UCB_RS.py'' are for the simulations of recommender systems with the MovieLens datasets. The synthetic datasets are generated in the corresponding codes and the preprocessed MovieLens datasets are in the file ''movielens_norm_100.npy''. The original MovieLens datasets can be downloaded here and the preprocessing steps are specified in the paper.

Dependencies

The original codes are written with Python 3.7, and the needed packages are ''numpy 1.18.1'' and ''matplotlib 3.1.3''.

Results

The performance of Fed1-UCB algorithm with the synthetic datasets as shown in Fig. 3 can be get by directly running the file ''Fed1_UCB_CR.py''. The default setting is for $f(p)=\lceil10\log(T)\rceil$ and $M=5$ with communication loss. To ignore the communication loss, comment out the line of computing it, which is labelled in the code. Results with other choices of $M$ and $f(p)$ under different bandit environments can also be get by changing the corresponding parameters in the code.

The performance of Fed2-UCB algorithm with the synthetic datasets as shown in Fig. 4 can be get by directly running the file ''Fed2_UCB_CR.py''. The default setting is for $f(p)=100$ and $g(p)=2^p$ with communication loss. Similar changes can be made as above to get the other results.

The performance of Fed2-UCB algorithm with the synthetic datasets and a reduced horizon as shown in Fig. 5 can be get by directly running the file ''Fed2_UCB_CR_short.py''. The default setting is for $f(p)=50$ and $g(p)=2^p$ with communication loss. Similar changes can be made as above to get the other results.

The performance of Fed1-UCB algorithm with the real-world datasets as shown in Fig. 6 can be get by directly running the file ''Fed1_UCB_RS.py''. The default setting for Fed1-UCB is $f(p)=\lceil10\log(T)\rceil$ with all available clients, which is for the curve labelled as ''Fed1-UCB, full'' in Fig. 5. The curve labelled as ''Fed1-UCB, $M=200$'' can be get by changing $M=200$ in the code. The results for Fed2-UCB can be get by directly running the file ''Fed2_UCB_RS.py'', which is set for $f(p)=200$ and $g(p)=2^p$ by default.