This work proposes a procedure for designing algorithms for specific adaptive data collection tasks like active learning and pure-exploration multi-armed bandits. Unlike the design of traditional adaptive algorithms that rely on concentration of measure and careful analysis to justify the correctness and sample complexity of the procedure, our adaptive algorithm is learned via adversarial training over equivalence classes of problems derived from information theoretic lower bounds. In particular, a single adaptive learning algorithm is learned that competes with the best adaptive algorithm learned for each equivalence class. Our procedure takes as input just the available queries, set of hypotheses, loss function, and total query budget. This is in contrast to existing meta-learning work that learns an adaptive algorithm relative to an explicit, user-defined subset or prior distribution over problems which can be challenging to define and be mismatched to the instance encountered at test time. This work is particularly focused on the regime when the total query budget is very small, such as a few dozen, which is much smaller than those budgets typically considered by theoretically derived algorithms. We perform synthetic experiments to justify the stability and effectiveness of the training procedure, and then evaluate the method on tasks derived from real data including a noisy 20 Questions game and a joke recommendation task.

Reinforcement learning agents are fundamentally limited by the quality of the reward functions they learn from, yet reward design is often overlooked under the assumption that a well-defined reward is readily available. However, in practice, designing rewards is difficult, and even when specified, evaluating their correctness is equally problematic: how do we know if a reward function is correctly specified? In our work, we address these challenges by focusing on reward alignment -- assessing whether a reward function accurately encodes the preferences of a human stakeholder. As a concrete measure of reward alignment, we introduce the Trajectory Alignment Coefficient to quantify the similarity between a human stakeholder's ranking of trajectory distributions and those induced by a given reward function. We show that the Trajectory Alignment Coefficient exhibits desirable properties, such as not requiring access to a ground truth reward, invariance to potential-based reward shaping, and applicability to online RL. Additionally, in an 11 -- person user study of RL practitioners, we found that access to the Trajectory Alignment Coefficient during reward selection led to statistically significant improvements. Compared to relying only on reward functions, our metric reduced cognitive workload by 1.5x, was preferred by 82% of users and increased the success rate of selecting reward functions that produced performant policies by 41%.

In this paper, we study offline preference-based reinforcement learning (PbRL), where learning is based on pre-collected preference feedback over pairs of trajectories. While offline PbRL has demonstrated remarkable empirical success, existing theoretical approaches face challenges in ensuring conservatism under uncertainty, requiring computationally intractable confidence set constructions. We address this limitation by proposing Adversarial Preference-based Policy Optimization ( APPO), a computationally efficient algorithm for offline PbRL that guarantees sample complexity bounds without relying on explicit confidence sets. By framing PbRL as a two-player game between a policy and a model, our approach enforces conservatism in a tractable manner. Using standard assumptions on function approximation and bounded trajectory concentrability, we derive a sample complexity bound. To our knowledge, APPO is the first offline PbRL algorithm to offer both statistical efficiency and practical applicability. Experimental results on continuous control tasks demonstrate that APPO effectively learns from complex datasets, showing comparable performance with existing state-of-the-art methods. While Reinforcement Learning (RL) has achieved remarkable success in real-world applications (Mnih, 2013; Silver et al., 2017; Kalashnikov et al., 2018; Brohan et al., 2022), its performance heavily depends on the design of the reward function (Wirth et al., 2017), which can be challenging in practice. To address this issue, preference-based reinforcement learning (PbRL), also known as reinforcement learning with human feedback, has gained increasing attention as an alternative to manually designed rewards. In PbRL, a reward model is learned from preference feedback provided by human experts, who compare pairs of trajectories (Christiano et al., 2017). This approach enables the learning process to align better with human intentions. However, collecting preference feedback can be costly, especially when real-time feedback from human experts is required. In such cases, learning from pre-collected data is preferred over online learning. This approach is referred to as offline PbRL, where the learning process relies solely on pre-collected trajectories and preference feedback.

This research considers Bayesian decision-analytic approaches toward the traversal of an uncertain graph. Namely, a traveler progresses over a graph in which rewards are gained upon a node's first visit and costs are incurred for every edge traversal. The traveler knows the graph's adjacency matrix and his starting position but does not know the rewards and costs. The traveler is a Bayesian who encodes his beliefs about these values using a Gaussian process prior and who seeks to maximize his expected utility over these beliefs. Adopting a decision-analytic perspective, we develop sequential decision-making solution strategies for this coupled information-collection and network-routing problem. We show that the problem is NP-Hard and derive properties of the optimal walk. These properties provide heuristics for the traveler's problem that balance exploration and exploitation. We provide a practical case study focused on the use of unmanned aerial systems for public safety and empirically study policy performance in myriad Erdos-Renyi settings.

Many reinforcement learning (RL) algorithms require large amounts of data, prohibiting their use in applications where frequent interactions with operational systems are infeasible, or high-fidelity simulations are expensive or unavailable. Meanwhile, low-fidelity simulators--such as reduced-order models, heuristic reward functions, or generative world models--can cheaply provide useful data for RL training, even if they are too coarse for direct sim-to-real transfer. We propose multi-fidelity policy gradients (MFPGs), an RL framework that mixes a small amount of data from the target environment with a large volume of low-fidelity simulation data to form unbiased, reduced-variance estimators (control variates) for on-policy policy gradients. We instantiate the framework by developing multi-fidelity variants of two policy gradient algorithms: REINFORCE and proximal policy optimization. Experimental results across a suite of simulated robotics benchmark problems demonstrate that when target-environment samples are limited, MFPG achieves up to 3.9x higher reward and improves training stability when compared to baselines that only use high-fidelity data. Moreover, even when the baselines are given more high-fidelity samples--up to 10x as many interactions with the target environment--MFPG continues to match or outperform them. Finally, we observe that MFPG is capable of training effective policies even when the low-fidelity environment is drastically different from the target environment. MFPG thus not only offers a novel paradigm for efficient sim-to-real transfer but also provides a principled approach to managing the trade-off between policy performance and data collection costs.

Data scarcity has long been an issue in the robot learning community. Particularly, in safety-critical domains like surgical applications, obtaining high-quality data can be especially difficult. It poses challenges to researchers seeking to exploit recent advancements in reinforcement learning and imitation learning, which have greatly improved generalizability and enabled robots to conduct tasks autonomously. We introduce dARt Vinci, a scalable data collection platform for robot learning in surgical settings. The system uses Augmented Reality (AR) hand tracking and a high-fidelity physics engine to capture subtle maneuvers in primitive surgical tasks: By eliminating the need for a physical robot setup and providing flexibility in terms of time, space, and hardware resources-such as multiview sensors and actuators-specialized simulation is a viable alternative. At the same time, AR allows the robot data collection to be more egocentric, supported by its body tracking and content overlaying capabilities. Our user study confirms the proposed system's efficiency and usability, where we use widely-used primitive tasks for training teleoperation with da Vinci surgical robots. Data throughput improves across all tasks compared to real robot settings by 41% on average. The total experiment time is reduced by an average of 10%. The temporal demand in the task load survey is improved. These gains are statistically significant. Additionally, the collected data is over 400 times smaller in size, requiring far less storage while achieving double the frequency.

As image-based deep reinforcement learning tackles more challenging tasks, increasing model size has become an important factor in improving performance. Recent studies achieved this by focusing on the parameter efficiency of scaled networks, typically using Impala-CNN, a 15-layer ResNet-inspired network, as the image encoder. However, while Impala-CNN evidently outperforms older CNN architectures, potential advancements in network design for deep reinforcement learning-specific image encoders remain largely unexplored. We find that replacing the flattening of output feature maps in Impala-CNN with global average pooling leads to a notable performance improvement. This approach outperforms larger and more complex models in the Procgen Benchmark, particularly in terms of generalization. We call our proposed encoder model Impoola-CNN. A decrease in the network's translation sensitivity may be central to this improvement, as we observe the most significant gains in games without agent-centered observations. Our results demonstrate that network scaling is not just about increasing model size - efficient network design is also an essential factor.

Research on human-AI collaboration often prioritizes objective performance. However, understanding human subjective preferences is essential to improving human-AI complementarity and human experiences. We investigate human preferences for controllability in a shared workspace task with AI partners using Behavior Shaping (BS), a reinforcement learning algorithm that allows humans explicit control over AI behavior. In one experiment, we validate the robustness of BS in producing effective AI policies relative to self-play policies, when controls are hidden. In another experiment, we enable human control, showing that participants perceive AI partners as more effective and enjoyable when they can directly dictate AI behavior. Our findings highlight the need to design AI that prioritizes both task performance and subjective human preferences. By aligning AI behavior with human preferences, we demonstrate how human-AI complementarity can extend beyond objective outcomes to include subjective preferences.

RL-based post-training of language models is almost exclusively done using on-policy methods such as PPO. These methods cannot learn from arbitrary sequences such as those produced earlier in training, in earlier runs, by human experts or other policies, or by decoding and exploration methods. This results in severe sample inefficiency and exploration difficulties, as well as a potential loss of diversity in the policy responses. Moreover, asynchronous PPO implementations require frequent and costly model transfers, and typically use value models which require a large amount of memory. In this paper we introduce Soft Policy Optimization (SPO), a simple, scalable and principled Soft RL method for sequence model policies that can learn from arbitrary online and offline trajectories and does not require a separate value model. In experiments on code contests, we shows that SPO outperforms PPO on pass@10, is significantly faster and more memory efficient, is able to benefit from off-policy data, enjoys improved stability, and learns more diverse (i.e. soft) policies.

An issue concerning the use of deep reinforcement learning (RL) agents is whether they can be trusted to perform reliably when deployed, as training environments may not reflect real-life environments. Anticipating instances outside their training scope, learning-enabled systems are often equipped with out-of-distribution (OOD) detectors that alert when a trained system encounters a state it does not recognize or in which it exhibits uncertainty. There exists limited work conducted on the problem of OOD detection within RL, with prior studies being unable to achieve a consensus on the definition of OOD execution within the context of RL. By framing our problem using a Markov Decision Process, we assume there is a transition distribution mapping each state-action pair to another state with some probability. Based on this, we consider the following definition of OOD execution within RL: A transition is OOD if its probability during real-life deployment differs from the transition distribution encountered during training. As such, we utilize conditional variational autoencoders (CVAE) to approximate the transition dynamics of the training environment and implement a conformity-based detector using reconstruction loss that is able to guarantee OOD detection with a pre-determined confidence level. We evaluate our detector by adapting existing benchmarks and compare it with existing OOD detection models for RL.