Truncated Importance Sampling

By Edward L. Ionides. 2008

252 citations recorded by the source corpus.

Subfields: Artificial Intelligence, Statistics and Probability

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Statistics and Probability

Citations in the corpus

Balanced Policy Evaluation and Learning (2017, neutral): “Another stopgap measure is to clip the propensities [14, 20] resulting in the clipped IPW (CIPW) estimator: ⌧̂M -CIPW(⇡) = ⌧̂WM -CIPW(⇡), W M -CIPW i (⇡) = ⇡Ti(Xi)/max{M, '̂Ti(Xi)}.”
End-to-End Offline Goal-Oriented Dialog Policy Learning via Policy Gradient (2017, background): “To deal with this problem, one can clip this value, or simply set it to a constant value if πθ(·|st ) is close to q(·|st ) [7, 12].”
Batch learning from logged bandit feedback through counterfactual risk minimization (2015, background): “This can be ﬁxed by “clipping” the importance sampling weights (Ionides, 2008; Strehl et al., 2010; Bottou et al., 2013; Cortes et al., 2010) (2) This objective captures the essence of previous oﬄine policy optimization approaches (Bot-tou et al., 2013; Strehl et al., 2010).”
Bayesian Counterfactual Risk Minimization (2018, background): “For instance, we could truncate the ratio of the policy and the logging policy, min{π(ai |xi)/pi, τ−1} (Ionides, 2008; Swaminathan and Joachims, 2015a).”
Performances of LOO and WAIC as IRT Model Selection Methods (2017, methodology): “Ionides, E. L. (2008). Truncated importance sampling. Journal of Computational and Graphical Statistics, 17(2), 295-311. Kang, T., & Cohen, A. S. (2007). IRT model selection methods for dichotomous items. Applied Psychological Measurement, 31(4), 331-358. Kang, T., Cohen, A. S., & Sung, H. J. (2009)”
Reinforcement Learning based Recommender Systems: A Survey (2021, methodology): “Truncated importance sampling [189] is used to alleviate the bias in the offline data.”
Constructing effective personalized policies using counterfactual inference from biased data sets with many features (2016, methodology): “To avoid the ﬁrst problem, we follow [Ionides(2008)] and modify the estimated rewards by truncating the importance sampling weights.”
Stochastic Structured Prediction under Bandit Feedback (2016, background): “We deal with this problem by clipping too small sampling probabilities to p̂wt(ỹt|xt) = max{pwt(ỹt|xt), k} for a constant k [9].”
Rethinking the Effective Sample Size (2018, result): “We could also non-linearly transform the unnormalised weights before normalising them as in other studies (Ionides, 2008; Koblents & Míguez, 2013; Vehtari & Gelman, 2015) (refer to a review in Martino et al.”
Bandit Algorithms (2020, methodology): “A comprehensive examination of various generalisations is given by Joulani et al. [2017]. For follow-the-regularisedleader, convexity of F suffices, as you will show using directional derivatives in Exercise 28.”
Online Evaluation for Information Retrieval (2016, methodology): “This technique, known as truncated importance sampling [Ionides, 2008], often leads to a lower mean squared estimation error:”
Off-Policy Evaluation and Learning from Logged Bandit Feedback: Error Reduction via Surrogate Policy (2018, methodology): “To reduce the mean squared error of IPS, several thresholding approaches have been proposed and studied in the literature (e.g., [1, 2, 9, 16]).”
Distributed Structured Actor-Critic Reinforcement Learning for Universal Dialogue Management (2020, background): “for V , and ρk def = min(ρ, π(ak|bk) μ(ak |bk) ) and cd def = min(c, π(ad|bd) μ(ad|bd) ) are truncated importance sampling weights [34].”
Pathfinder: Parallel quasi-Newton variational inference (2021, unknown): “Ionides (2008) showed that truncating the importance weights improves the efficiency of the resulting Monte Carlo estimator by reducing its variance.”
Pareto Smoothed Importance Sampling (2015, methodology): “We can then write the PSIS estimator as a Winsorized version of the truncated estimator of Ionides (2008) plus an extra bias correction term: The following theorem, which is proved in Appendix C, shows that our idealized version of PSIS converges under mild conditions.”
A Bayesian view of doubly robust causal inference (2016, background): “In order to control this, a standard approach would be to truncate the weights (e.g. Xiao et al., 2013), which would also be possible in the importance sampling context (Ionides, 2008).”