CoinDICE: Off-Policy Confidence Interval Estimation

Authors: Bo Dai, Ofir Nachum, Yinlam Chow, Lihong Li, Csaba Szepesvari, Dale Schuurmans

NeurIPS 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We now evaluate the empirical performance of Coin DICE, comparing it to a number of existing confidence interval estimators for OPE based on concentration inequalities.
Researcher Affiliation Collaboration 1Google Research, Brain Team 2University of Alberta 3Deep Mind
Pseudocode No The provided text does not contain any structured pseudocode or algorithm blocks, although it refers to "Algorithm 1" in an appendix.
Open Source Code Yes Open-source code for Coin DICE is available at https://github.com/google-research/dice_rl.
Open Datasets Yes We use Frozen Lake (Brockman et al., 2016), a highly stochastic gridworld environment, and Taxi (Dietterich, 1998), an environment with a moderate state space of 2 000 elements. ... Lastly, we evaluate Coin DICE on Reacher (Brockman et al., 2016; Todorov et al., 2012), a continuous control environment.
Dataset Splits No The paper mentions collecting a static dataset and sampling from it, but does not specify explicit training/validation/test splits, percentages, or sample counts.
Hardware Specification No The paper does not provide specific hardware details such as GPU/CPU models, processor types, or memory amounts used for running experiments.
Software Dependencies No The paper does not provide specific ancillary software details with version numbers (e.g., library or solver names with versions).
Experiment Setup Yes The discount factor is γ = 0.99. The target policy is taken to be a near-optimal one, while the behavior policy is highly suboptimal. The behavior policy in Frozen Lake is the optimal policy with 0.2 white noise... in this setting, we use a one-hidden-layer neural network with ReLU activations.