Bayesian optimization (BO) conventionally relies on handcrafted acquisition functions (AFs) to sequentially determine the sample points. However, it has been widely observed in practice that the best-performing AF in terms of regret can vary significantly under different types of black-box functions. It has remained a challenge to design one AF that can attain the best performance over a wide variety of black-box functions. This paper aims to attack this challenge through the perspective of reinforced few-shot AF learning (FSAF). Specifically, we first connect the notion of AFs with Q-functions and view a deep Q-network (DQN) as a surrogate differentiable AF. While it serves as a natural idea to combine DQN and an existing few-shot learning method, we identify that such a direct combination does not perform well due to severe overfitting, which is particularly critical in BO due to the need of a versatile sampling policy. To address this, we present a Bayesian variant of DQN with the following three features: (i) It learns a distribution of Q-networks as AFs based on the Kullback-Leibler regularization framework. This inherently provides the uncertainty required in sampling for BO and mitigates overfitting.

Incontrast,our work is concerned with an overall limit on the total amount of information an agent may acquire fromtheenvironment and,inturn,howthattranslates intoitsselection ofafeasible learning target.

Inthiswork,weconsider thescenario where agent limitations may entirely preclude identifying an exactly value-equivalent model, immediately giving rise to a trade-off between identifying a model that is simple enough to learn while only incurring bounded sub-optimality.

We study the off-policy evaluation (OPE) problem in reinforcement learning with linear function approximation, which aims to estimate the value function of a target policy based on the offline data collected by a behavior policy. We propose to incorporate the variance information of the value function to improve the sample efficiency of OPE. More specifically, for time-inhomogeneous episodic linear Markov decision processes (MDPs), we propose an algorithm, VA-OPE, which uses the estimated variance of the value function to reweight the Bellman residual in Fitted Q-Iteration. We show that our algorithm achieves a tighter error bound than the best-known result. We also provide a fine-grained characterization of the distribution shift between the behavior policy and the target policy. Extensive numerical experiments corroborate our theory.

Proposition 1. Suppose Qφj(s,a) = Q(s,a) and Qφj(s,) is locally linear in the neighborhood of a for all j [N]. Let λmin and wmin be the smallest eigenvalue and the corresponding normalized eigenvector of the matrix Var aQφj(s,a) and > 0 be the value such that mini6=j aQφi(s,a), aQφj(s,a) = 1 . We first prove that the smallest eigenvalue λmin of Var aQφj(s,a) is upper-bounded by some constant multiple of . By Lemma 1, the total variance of the matrix is less or equal to N 1N. Note that, using the fact that the Q-values coincide at the action a and the local linearity of the Q-functions, we have derived Var(Qφj(s,a+ kw)) = k2w|Var aQφj(s,a) w. (2) Plugging w = wmin in Equation (2) and using Equation (1), we have Var(Qφj(s,a+ kwmin)) = k2w|minVar aQφj(s,a) wmin = k2λmin A.2 Relationship between maximizing the total variance and maximizing the smallest eigenvalue As we have shown in Section 4, maximizing the total variance of the matrix Var ( aQφi(s,a)) is equivalent to minimizing the cosine similarity of all distinct pairs of the gradients aQφi(s,a), 2 which makes the gradients uniformly distributed on the unit sphere S|A| 1.

Offline reinforcement learning (offline RL), which aims to find an optimal policy from a previously collected static dataset, bears algorithmic difficulties due to function approximation errors from out-of-distribution (OOD) data points. To this end, offline RL algorithms adopt either a constraint or a penalty term that explicitly guides the policy to stay close to the given dataset. However, prior methods typically require accurate estimation of the behavior policy or sampling from OOD data points, which themselves can be a non-trivial problem. Moreover, these methods under-utilize the generalization ability of deep neural networks and often fall into suboptimal solutions too close to the given dataset. In this work, we propose an uncertainty-based offline RL method that takes into account the confidence of the Q-value prediction and does not require any estimation or sampling of the data distribution. We show that the clipped Q-learning, a technique widely used in online RL, can be leveraged to successfully penalize OOD data points with high prediction uncertainties. Surprisingly, we find that it is possible to substantially outperform existing offline RL methods on various tasks by simply increasing the number of Q-networks along with the clipped Q-learning. Based on this observation, we propose an ensemble-diversified actor-critic algorithm that reduces the number of required ensemble networks down to a tenth compared to the naive ensemble while achieving state-of-the-art performance on most of the D4RL benchmarks considered.