Offline policy evaluation (OPE) allows us to evaluate and estimate a new sequential decision-making policy's performance by leveraging historical interaction data collected from other policies. Evaluating a new policy online without a confident estimate of its performance can lead to costly, unsafe, or hazardous outcomes, especially in education and healthcare. Several OPE estimators have been proposed in the last decade, many of which have hyperparameters and require training. Unfortunately, choosing the best OPE algorithm for each task and domain is still unclear. In this paper, we propose a new algorithm that adaptively blends a set of OPE estimators given a dataset without relying on an explicit selection using a statistical procedure. We prove that our estimator is consistent and satisfies several desirable properties for policy evaluation. Additionally, we demonstrate that when compared to alternative approaches, our estimator can be used to select higher-performing policies in healthcare and robotics. Our work contributes to improving ease of use for a general-purpose, estimator-agnostic, off-policy evaluation framework for offline RL.

Sequential decision-making domains such as recommender systems, healthcare and education often have unobserved heterogeneity in the population that can be modeled using latent bandits $-$ a framework where an unobserved latent state determines the model for a trajectory. While the latent bandit framework is compelling, the extent of its generality is unclear. We first address this by establishing a de Finetti theorem for decision processes, and show that $\textit{every}$ exchangeable and coherent stateless decision process is a latent bandit. The latent bandit framework lends itself particularly well to online learning with offline datasets, a problem of growing interest in sequential decision-making. One can leverage offline latent bandit data to learn a complex model for each latent state, so that an agent can simply learn the latent state online to act optimally. We focus on a linear model for a latent bandit with $d_A$-dimensional actions, where the latent states lie in an unknown $d_K$-dimensional subspace for $d_K \ll d_A$. We present SOLD, a novel principled method to learn this subspace from short offline trajectories with guarantees. We then provide two methods to leverage this subspace online: LOCAL-UCB and ProBALL-UCB. We demonstrate that LOCAL-UCB enjoys $\tilde O(\min(d_A\sqrt{T}, d_K\sqrt{T}(1+\sqrt{d_AT/d_KN})))$ regret guarantees, where the effective dimension is lower when the size $N$ of the offline dataset is larger. ProBALL-UCB enjoys a slightly weaker guarantee, but is more practical and computationally efficient. Finally, we establish the efficacy of our methods using experiments on both synthetic data and real-life movie recommendation data from MovieLens.

We study continuous-time reinforcement learning (RL) for stochastic control in which system dynamics are governed by jump-diffusion processes. We formulate an entropy-regularized exploratory control problem with stochastic policies to capture the exploration--exploitation balance essential for RL. Unlike the pure diffusion case initially studied by Wang et al. (2020), the derivation of the exploratory dynamics under jump-diffusions calls for a careful formulation of the jump part. Through a theoretical analysis, we find that one can simply use the same policy evaluation and q-learning algorithms in Jia and Zhou (2022a, 2023), originally developed for controlled diffusions, without needing to check a priori whether the underlying data come from a pure diffusion or a jump-diffusion. However, we show that the presence of jumps ought to affect parameterizations of actors and critics in general. Finally, we investigate as an application the mean-variance portfolio selection problem with stock price modelled as a jump-diffusion, and show that both RL algorithms and parameterizations are invariant with respect to jumps.

We present Causal Amortized Active Structure Learning (CAASL), an active intervention design policy that can select interventions that are adaptive, real-time and that does not require access to the likelihood. This policy, an amortized network based on the transformer, is trained with reinforcement learning on a simulator of the design environment, and a reward function that measures how close the true causal graph is to a causal graph posterior inferred from the gathered data. On synthetic data and a single-cell gene expression simulator, we demonstrate empirically that the data acquired through our policy results in a better estimate of the underlying causal graph than alternative strategies. Our design policy successfully achieves amortized intervention design on the distribution of the training environment while also generalizing well to distribution shifts in test-time design environments. Further, our policy also demonstrates excellent zero-shot generalization to design environments with dimensionality higher than that during training, and to intervention types that it has not been trained on.

Weakly-supervised segmentation (WSS) methods, reliant on image-level labels indicating object presence, lack explicit correspondence between labels and regions of interest (ROIs), posing a significant challenge. Despite this, WSS methods have attracted attention due to their much lower annotation costs compared to fully-supervised segmentation. Leveraging reinforcement learning (RL) self-play, we propose a novel WSS method that gamifies image segmentation of a ROI. We formulate segmentation as a competition between two agents that compete to select ROI-containing patches until exhaustion of all such patches. The score at each time-step, used to compute the reward for agent training, represents likelihood of object presence within the selection, determined by an object presence detector pre-trained using only image-level binary classification labels of object presence. Additionally, we propose a game termination condition that can be called by either side upon exhaustion of all ROI-containing patches, followed by the selection of a final patch from each. Upon termination, the agent is incentivised if ROI-containing patches are exhausted or disincentivised if an ROI-containing patch is found by the competitor. This competitive setup ensures minimisation of over- or under-segmentation, a common problem with WSS methods. Extensive experimentation across four datasets demonstrates significant performance improvements over recent state-of-the-art methods. Code: https://github.com/s-sd/spurl/tree/main/wss

Unsupervised and self-supervised objectives, such as next token prediction, have enabled pre-training generalist models from large amounts of unlabeled data. In reinforcement learning (RL), however, finding a truly general and scalable unsupervised pre-training objective for generalist policies from offline data remains a major open question. While a number of methods have been proposed to enable generic self-supervised RL, based on principles such as goal-conditioned RL, behavioral cloning, and unsupervised skill learning, such methods remain limited in terms of either the diversity of the discovered behaviors, the need for high-quality demonstration data, or the lack of a clear adaptation mechanism for downstream tasks. In this work, we propose a novel unsupervised framework to pre-train generalist policies that capture diverse, optimal, long-horizon behaviors from unlabeled offline data such that they can be quickly adapted to any arbitrary new tasks in a zero-shot manner. Our key insight is to learn a structured representation that preserves the temporal structure of the underlying environment, and then to span this learned latent space with directional movements, which enables various zero-shot policy "prompting" schemes for downstream tasks. Through our experiments on simulated robotic locomotion and manipulation benchmarks, we show that our unsupervised policies can solve goal-conditioned and general RL tasks in a zero-shot fashion, even often outperforming prior methods designed specifically for each setting. Our code and videos are available at https://seohong.me/projects/hilp/.

Meta-reinforcement learning has widely been used as a learning-to-learn framework to solve unseen tasks with limited experience. However, the aspect of constraint violations has not been adequately addressed in the existing works, making their application restricted in real-world settings. In this paper, we study the problem of meta-safe reinforcement learning (Meta-SRL) through the CMDP-within-online framework to establish the first provable guarantees in this important setting. We obtain task-averaged regret bounds for the reward maximization (optimality gap) and constraint violations using gradient-based meta-learning and show that the task-averaged optimality gap and constraint satisfaction improve with task-similarity in a static environment or task-relatedness in a dynamic environment. Several technical challenges arise when making this framework practical. To this end, we propose a meta-algorithm that performs inexact online learning on the upper bounds of within-task optimality gap and constraint violations estimated by off-policy stationary distribution corrections. Furthermore, we enable the learning rates to be adapted for every task and extend our approach to settings with a competing dynamically changing oracle. Finally, experiments are conducted to demonstrate the effectiveness of our approach.

In human-AI interaction, a prominent goal is to attain human's desirable outcome with the assistance of AI agents, which can be ideally delineated as a problem of seeking the optimal Nash Equilibrium that matches the human's desirable outcome. However, reaching the outcome is usually challenging due to the existence of multiple Nash Equilibria that are related to the assisting task but do not correspond to the human's desirable outcome. To tackle this issue, we employ a theoretical framework called structural causal game (SCG) to formalize the human-AI interactive process. Furthermore, we introduce a strategy referred to as pre-policy intervention on the SCG to steer AI agents towards attaining the human's desirable outcome. In more detail, a pre-policy is learned as a generalized intervention to guide the agents' policy selection, under a transparent and interpretable procedure determined by the SCG. To make the framework practical, we propose a reinforcement learning-like algorithm to search out this pre-policy. The proposed algorithm is tested in both gridworld environments and realistic dialogue scenarios with large language models, demonstrating its adaptability in a broader class of problems and potential effectiveness in real-world situations.

Adversarial Imitation Learning (AIL) faces challenges with sample inefficiency because of its reliance on sufficient on-policy data to evaluate the performance of the current policy during reward function updates. In this work, we study the convergence properties and sample complexity of off-policy AIL algorithms. We show that, even in the absence of importance sampling correction, reusing samples generated by the $o(\sqrt{K})$ most recent policies, where $K$ is the number of iterations of policy updates and reward updates, does not undermine the convergence guarantees of this class of algorithms. Furthermore, our results indicate that the distribution shift error induced by off-policy updates is dominated by the benefits of having more data available. This result provides theoretical support for the sample efficiency of off-policy AIL algorithms. To the best of our knowledge, this is the first work that provides theoretical guarantees for off-policy AIL algorithms.

A major challenge in decision making domains with large state spaces is to effectively select actions which maximize utility. In recent years, approaches such as reinforcement learning (RL) and search algorithms have been successful to tackle this issue, despite their differences. RL defines a learning framework that an agent explores and interacts with. Search algorithms provide a formalism to search for a solution. However, it is often difficult to evaluate the performances of such approaches in a practical way. Motivated by this problem, we focus on one game domain, i.e., Connect-4, and develop a novel evolutionary framework to evaluate three classes of algorithms: RL, Minimax and Monte Carlo tree search (MCTS). The contribution of this paper is threefold: i) we implement advanced versions of these algorithms and provide a systematic comparison with their standard counterpart, ii) we develop a novel evaluation framework, which we call the Evolutionary Tournament, and iii) we conduct an extensive evaluation of the relative performance of each algorithm to compare our findings. We evaluate different metrics and show that MCTS achieves the best results in terms of win percentage, whereas Minimax and Q-Learning are ranked in second and third place, respectively, although the latter is shown to be the fastest to make a decision.