In reinforcement learning (RL) with experience replay, experiences stored in a replay buffer influence the RL agent's performance. Information about the influence of these experiences is valuable for various purposes, such as identifying experiences that negatively influence poorly performing RL agents. One method for estimating the influence of experiences is the leave-one-out (LOO) method. However, this method is usually computationally prohibitive. In this paper, we present Policy Iteration with Turn-over Dropout (PIToD), which efficiently estimates the influence of experiences. We evaluate how accurately PIToD estimates the influence of experiences and its efficiency compared to LOO. We then apply PIToD to amend poorly performing RL agents, i.e., we use PIToD to estimate negatively influential experiences for the RL agents and to delete the influence of these experiences. We show that RL agents' performance is significantly improved via amendments with PIToD.

Reinforcement Learning (RL) is a powerful method for controlling dynamic systems, but its learning mechanism can lead to unpredictable actions that undermine the safety of critical systems. Here, we propose RL with Adaptive Control Regularization (RL-ACR), an algorithm that enables safe RL exploration by combining the RL policy with a policy regularizer that hard-codes safety constraints. We perform policy combination via a "focus network," which determines the appropriate combination depending on the state -- relying more on the safe policy regularizer for less-exploited states while allowing unbiased convergence for well-exploited states. In a series of critical control applications, we demonstrate that RL-ACR ensures safety during training while achieving the performance standards of model-free RL approaches that disregard safety.

Human-in-the-loop reinforcement learning (HRL) allows the training of agents through various interfaces, even for non-expert humans. Recently, preference-based methods (PBRL), where the human has to give his preference over two trajectories, increased in popularity since they allow training in domains where more direct feedback is hard to formulate. However, the current PBRL methods have limitations and do not provide humans with an expressive interface for giving feedback. With this work, we propose a new preference-based learning method that provides humans with a more expressive interface to provide their preference over trajectories and a factual explanation (or annotation of why they have this preference). These explanations allow the human to explain what parts of the trajectory are most relevant for the preference. We allow the expression of the explanations over individual trajectory steps. We evaluate our method in various simulations using a simulated human oracle (with realistic restrictions), and our results show that our extended feedback can improve the speed of learning. Code & data: github.com/under-rewiev

Humans use social context to specify preferences over behaviors, i.e. their reward functions. Yet, algorithms for inferring reward models from preference data do not take this social learning view into account. Inspired by pragmatic human communication, we study how to extract fine-grained data regarding why an example is preferred that is useful for learning more accurate reward models. We propose to enrich binary preference queries to ask both (1) which features of a given example are preferable in addition to (2) comparisons between examples themselves. We derive an approach for learning from these feature-level preferences, both for cases where users specify which features are reward-relevant, and when users do not. We evaluate our approach on linear bandit settings in both vision- and language-based domains. Results support the efficiency of our approach in quickly converging to accurate rewards with fewer comparisons vs. example-only labels. Finally, we validate the real-world applicability with a behavioral experiment on a mushroom foraging task. Our findings suggest that incorporating pragmatic feature preferences is a promising approach for more efficient user-aligned reward learning.

In this paper, we consider the supervised pretrained transformer for a class of sequential decision-making problems. The class of considered problems is a subset of the general formulation of reinforcement learning in that there is no transition probability matrix, and the class of problems covers bandits, dynamic pricing, and newsvendor problems as special cases. Such a structure enables the use of optimal actions/decisions in the pretraining phase, and the usage also provides new insights for the training and generalization of the pretrained transformer. We first note that the training of the transformer model can be viewed as a performative prediction problem, and the existing methods and theories largely ignore or cannot resolve the arisen out-of-distribution issue. We propose a natural solution that includes the transformer-generated action sequences in the training procedure, and it enjoys better properties both numerically and theoretically. The availability of the optimal actions in the considered tasks also allows us to analyze the properties of the pretrained transformer as an algorithm and explains why it may lack exploration and how this can be automatically resolved. Numerically, we categorize the advantages of the pretrained transformer over the structured algorithms such as UCB and Thompson sampling into three cases: (i) it better utilizes the prior knowledge in the pretraining data; (ii) it can elegantly handle the misspecification issue suffered by the structured algorithms; (iii) for short time horizon such as $T\le50$, it behaves more greedy and enjoys much better regret than the structured algorithms which are designed for asymptotic optimality.

We focus on offline imitation learning (IL), which aims to mimic an expert's behavior using demonstrations without any interaction with the environment. One of the main challenges in offline IL is the limited support of expert demonstrations, which typically cover only a small fraction of the state-action space. While it may not be feasible to obtain numerous expert demonstrations, it is often possible to gather a larger set of sub-optimal demonstrations. For example, in treatment optimization problems, there are varying levels of doctor treatments available for different chronic conditions. These range from treatment specialists and experienced general practitioners to less experienced general practitioners. Similarly, when robots are trained to imitate humans in routine tasks, they might learn from individuals with different levels of expertise and efficiency. In this paper, we propose an offline IL approach that leverages the larger set of sub-optimal demonstrations while effectively mimicking expert trajectories. Existing offline IL methods based on behavior cloning or distribution matching often face issues such as overfitting to the limited set of expert demonstrations or inadvertently imitating sub-optimal trajectories from the larger dataset. Our approach, which is based on inverse soft-Q learning, learns from both expert and sub-optimal demonstrations. It assigns higher importance (through learned weights) to aligning with expert demonstrations and lower importance to aligning with sub-optimal ones. A key contribution of our approach, called SPRINQL, is transforming the offline IL problem into a convex optimization over the space of Q functions. Through comprehensive experimental evaluations, we demonstrate that the SPRINQL algorithm achieves state-of-the-art (SOTA) performance on offline IL benchmarks. Code is available at https://github.com/hmhuy2000/SPRINQL.

Deep reinforcement learning is used in various domains, but usually under the assumption that the environment has stationary conditions like transitions and state distributions. When this assumption is not met, performance suffers. For this reason, tracking continuous environmental changes and adapting to unpredictable conditions is challenging yet crucial because it ensures that systems remain reliable and flexible in practical scenarios. Our research introduces Behavior-Aware Detection and Adaptation (BADA), an innovative framework that merges environmental change detection with behavior adaptation. The key inspiration behind our method is that policies exhibit different global behaviors in changing environments. Specifically, environmental changes are identified by analyzing variations between behaviors using Wasserstein distances without manually set thresholds. The model adapts to the new environment through behavior regularization based on the extent of changes. The results of a series of experiments demonstrate better performance relative to several current algorithms. This research also indicates significant potential for tackling this long-standing challenge.

Traditional offline reinforcement learning methods predominantly operate in a batch-constrained setting. This confines the algorithms to a specific state-action distribution present in the dataset, reducing the effects of distributional shift but restricting the algorithm greatly. In this paper, we alleviate this limitation by introducing a novel framework named \emph{state-constrained} offline reinforcement learning. By exclusively focusing on the dataset's state distribution, our framework significantly enhances learning potential and reduces previous limitations. The proposed setting not only broadens the learning horizon but also improves the ability to combine different trajectories from the dataset effectively, a desirable property inherent in offline reinforcement learning. Our research is underpinned by solid theoretical findings that pave the way for subsequent advancements in this domain. Additionally, we introduce StaCQ, a deep learning algorithm that is both performance-driven on the D4RL benchmark datasets and closely aligned with our theoretical propositions. StaCQ establishes a strong baseline for forthcoming explorations in state-constrained offline reinforcement learning.

We resolve the open problem of designing a computationally efficient algorithm for infinite-horizon average-reward linear Markov Decision Processes (MDPs) with $\widetilde{O}(\sqrt{T})$ regret. Previous approaches with $\widetilde{O}(\sqrt{T})$ regret either suffer from computational inefficiency or require strong assumptions on dynamics, such as ergodicity. In this paper, we approximate the average-reward setting by the discounted setting and show that running an optimistic value iteration-based algorithm for learning the discounted setting achieves $\widetilde{O}(\sqrt{T})$ regret when the discounting factor $\gamma$ is tuned appropriately. The challenge in the approximation approach is to get a regret bound with a sharp dependency on the effective horizon $1 / (1 - \gamma)$. We use a computationally efficient clipping operator that constrains the span of the optimistic state value function estimate to achieve a sharp regret bound in terms of the effective horizon, which leads to $\widetilde{O}(\sqrt{T})$ regret.

Deep reinforcement learning (DRL) is a powerful machine learning paradigm for generating agents that control autonomous systems. However, the "black box" nature of DRL agents limits their deployment in real-world safety-critical applications. A promising approach for providing strong guarantees on an agent's behavior is to use Neural Lyapunov Barrier (NLB) certificates, which are learned functions over the system whose properties indirectly imply that an agent behaves as desired. However, NLB-based certificates are typically difficult to learn and even more difficult to verify, especially for complex systems. In this work, we present a novel method for training and verifying NLB-based certificates for discrete-time systems. Specifically, we introduce a technique for certificate composition, which simplifies the verification of highly-complex systems by strategically designing a sequence of certificates. When jointly verified with neural network verification engines, these certificates provide a formal guarantee that a DRL agent both achieves its goals and avoids unsafe behavior. Furthermore, we introduce a technique for certificate filtering, which significantly simplifies the process of producing formally verified certificates. We demonstrate the merits of our approach with a case study on providing safety and liveness guarantees for a DRL-controlled spacecraft.