We thank the reviewers for their insightful comments and suggestions. We hope that the rebuttal will clarify the issues. There are two reasons why Algorithm 1 should be preferred in practice. RL algorithms (e.g., MBIE or Delayed-QL), Algorithm 2 is extremely conservative, leading to very slow convergence. Algorithm 1 can be seen as a "practical" version of Algorithm 2 We will clarify this point in the final version.

In this paper, we explore the susceptibility of the independent Q-learning algorithms (a classical and widely used multi-agent reinforcement learning method) to strategic manipulation of sophisticated opponents in normal-form games played repeatedly. We quantify how much strategically sophisticated agents can exploit naive Q-learners if they know the opponents' Q-learning algorithm. To this end, we formulate the strategic actors' interactions as a stochastic game (whose state encompasses Q-function estimates of the Q-learners) as if the Q-learning algorithms are the underlying dynamical system. We also present a quantization-based approximation scheme to tackle the continuum state space and analyze its performance for two competing strategic actors and a single strategic actor both analytically and numerically.

We show that the logit-Q dynamics presented reach (near) efficient equilibrium in stochastic teams. We quantify a bound on the approximation error. We also show the rationality of the logit-Q dynamics against agents following pure stationary strategies and the convergence of the dynamics in stochastic games where the reward functions induce potential games, yet only a single agent controls the state transitions beyond stochastic teams. The key idea is to approximate the dynamics with a fictional scenario where the Q-function estimates are stationary over finite-length epochs only for analysis. We then couple the dynamics in the main and fictional scenarios to show that these two scenarios become more and more similar across epochs due to the vanishing step size.

Although distributional reinforcement learning (DRL) has been widely examined in the past few years, very few studies investigate the validity of the obtained Q-function estimator in the distributional setting. To fully understand how the approximation errors of the Q-function affect the whole training process, we do some error analysis and theoretically show how to reduce both the bias and the variance of the error terms. With this new understanding, we construct a new estimator \emph{Quantiled Expansion Mean} (QEM) and introduce a new DRL algorithm (QEMRL) from the statistical perspective. We extensively evaluate our QEMRL algorithm on a variety of Atari and Mujoco benchmark tasks and demonstrate that QEMRL achieves significant improvement over baseline algorithms in terms of sample efficiency and convergence performance.

Ensemble methods have been widely applied in Reinforcement Learning (RL) in order to enhance stability, increase convergence speed, and improve exploration. These methods typically work by employing an aggregation mechanism over actions of different RL algorithms. We show that a variety of these methods can be unified by drawing parallels from committee voting rules in Social Choice Theory. We map the problem of designing an action aggregation mechanism in an ensemble method to a voting problem which, under different voting rules, yield popular ensemble-based RL algorithms like Majority Voting Q-learning or Bootstrapped Q-learning. Our unification framework, in turn, allows us to design new ensemble-RL algorithms with better performance. For instance, we map two diversity-centered committee voting rules, namely Single Non-Transferable Voting Rule and Chamberlin-Courant Rule, into new RL algorithms that demonstrate excellent exploratory behavior in our experiments.