Minimax Optimal Algorithms for Fixed-Budget Best Arm Identification

Authors: Junpei Komiyama, Taira Tsuchiya, Junya Honda

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This section numerically tests the performance of the TNN algorithm. We compared the performance of TNN (Algorithm 4) with two algorithms: Uniform algorithm, which samples each arm in a round-robin fashion, and Successive Rejects (SR, Audibert et al., 2010), where the entire trial is divided into segments before the game starts, and one arm with the smallest estimated mean reward is removed for each segment.
Researcher Affiliation Academia Junpei Komiyama New York University New York, NY, United States junpei@komiyama.info Taira Tsuchiya Kyoto University, Kyoto, Japan RIKEN AIP, Tokyo, Japan tsuchiya@sys.i.kyoto-u.ac.jp Junya Honda Kyoto University, Kyoto, Japan RIKEN AIP, Tokyo, Japan honda@i.kyoto-u.ac.jp
Pseudocode Yes Algorithm 1: Rgo-Tracking; Algorithm 2: Delayed optimal tracking (DOT); Algorithm 3: Gradient descent method for θ; Algorithm 4: Rgo-Tracking by Neural Network (TNN)
Open Source Code Yes The source code of the simulation is available at https://github.com/tsuchhiii/fixedbudget-bai .
Open Datasets No The paper describes generating its own data instances (e.g., 'Bernoulli bandits with K = 3 arms') rather than using an established public dataset with access information or citations.
Dataset Splits No The paper describes experiments run for a fixed number of rounds (T=2000) and repeated many times, but it does not define dataset splits for training, validation, or testing.
Hardware Specification No The paper does not provide specific hardware details such as GPU or CPU models used for running the experiments. It mentions 'See appendix' regarding compute resources, but this information is not present in the provided text.
Software Dependencies No The paper mentions using 'Adam W' optimizer but does not specify software dependencies with version numbers, such as Python, PyTorch, or TensorFlow versions.
Experiment Setup Yes We used the neural network with four layers (including the input layer and output layer), where we used the Re LU for the activation functions and introduced the skip-connection (He et al., 2016) between each hidden layer to make training the network easier. The number of nodes in the hidden layers was fixed to K 3. We used Adam W (Loshchilov and Hutter, 2019) with a learning rate 10 3 and weight decay 10 7 to update the parameters. For training the neural network, we ran Algorithm 3 with N true = 32 and N emp = 90.