Suppose an online platform wants to compare a treatment and control policy, e.g.,

Weaknesses: - Can we interpret the results as follows: If the TAR assumption is satisfied with positive limits, and we use MLE, then temporal interference does not cause bias. If this interpretation is correct, then it would be illuminating if the authors provide the intuitive connection between the TAR assumption and temporal interference. It is not clear if the estimations that the authors have required are feasible if the state space is large. The next natural question is how robust the results are if we use other methods for estimation. This could have been shown by providing some simulations, which is a part missing from the manuscript.

The paper studied the online experimental design problem where there are temporal dependencies between the two control policies/treatments. The novelty of the problem setup and the theoretical analysis in the paper are appreciated by all the reviewers. Although the analysis is the main contribution, the paper would be much stronger if there are meaningful experiments on toy problems to showcase the performance the online MLE-based approach vs the standard experimental design approaches.

Suppose an online platform wants to compare a treatment and control policy (e.g., two different matching algorithms in a ridesharing system, or two different inventory management algorithms in an online retail site). Standard experimental approaches to this problem are biased (due to temporal interference between the policies), and not sample efficient. We study optimal experimental design for this setting. We view testing the two policies as the problem of estimating the steady state difference in reward between two unknown Markov chains (i.e., policies). We assume estimation of the steady state reward for each chain proceeds via nonparametric maximum likelihood, and search for consistent (i.e., asymptotically unbiased) experimental designs that are efficient (i.e., asymptotically minimum variance).

We consider experimentation in the presence of non-stationarity, inter-unit (spatial) interference, and carry-over effects (temporal interference), where we wish to estimate the global average treatment effect (GATE), the difference between average outcomes having exposed all units at all times to treatment or to control. We suppose spatial interference is described by a graph, where a unit's outcome depends on its neighborhood's treatment assignments, and that temporal interference is described by a hidden Markov decision process, where the transition kernel under either treatment (action) satisfies a rapid mixing condition. We propose a clustered switchback design, where units are grouped into clusters and time steps are grouped into blocks and each whole cluster-block combination is assigned a single random treatment. Under this design, we show that for graphs that admit good clustering, a truncated exposure-mapping Horvitz-Thompson estimator achieves $\tilde O(1/NT)$ mean-squared error (MSE), matching an $\Omega(1/NT)$ lower bound up to logarithmic terms. Our results simultaneously generalize the $N=1$ setting of Hu, Wager 2022 (and improves on the MSE bound shown therein for difference-in-means estimators) as well as the $T=1$ settings of Ugander et al 2013 and Leung 2022. Simulation studies validate the favorable performance of our approach.