A common setting in multitask reinforcement learning (RL) demands that an agent rapidly adapt to various stationary reward functions randomly sampled from a fixed distribution. In such situations, the successor representation (SR) is a popular framework which supports rapid policy evaluation by decoupling a policy's expected discounted, cumulative state occupancies from a specific reward function. However, in the natural world, sequential tasks are rarely independent, and instead reflect shifting priorities based on the availability and subjective perception of rewarding stimuli. Reflecting this disjunction, in this paper we study the phenomenon of diminishing marginal utility and introduce a novel state representation, the $\lambda$ representation ($\lambda$R) which, surprisingly, is required for policy evaluation in this setting and which generalizes the SR as well as several other state representations from the literature. We establish the $\lambda$R's formal properties and examine its normative advantages in the context of machine learning, as well as its usefulness for studying natural behaviors, particularly foraging.

Understanding the interactions of agents trained with deep reinforcement learning is crucial for deploying agents in games or the real world. In the former, unreasonable actions confuse players. In the latter, that effect is even more significant, as unexpected behavior cause accidents with potentially grave and long-lasting consequences for the involved individuals. In this work, we propose using program synthesis to imitate reinforcement learning policies after seeing a trajectory of the action sequence. Programs have the advantage that they are inherently interpretable and verifiable for correctness. We adapt the state-of-the-art program synthesis system DreamCoder for learning concepts in grid-based environments, specifically, a navigation task and two miniature versions of Atari games, Space Invaders and Asterix. By inspecting the generated libraries, we can make inferences about the concepts the black-box agent has learned and better understand the agent's behavior. We achieve the same by visualizing the agent's decision-making process for the imitated sequences. We evaluate our approach with different types of program synthesizers based on a search-only method, a neural-guided search, and a language model fine-tuned on code.

Coverage path planning (CPP) is a critical problem in robotics, where the goal is to find an efficient path that covers every point in an area of interest. This work addresses the power-constrained CPP problem with recharge for battery-limited unmanned aerial vehicles (UAVs). In this problem, a notable challenge emerges from integrating recharge journeys into the overall coverage strategy, highlighting the intricate task of making strategic, long-term decisions. We propose a novel proximal policy optimization (PPO)-based deep reinforcement learning (DRL) approach with map-based observations, utilizing action masking and discount factor scheduling to optimize coverage trajectories over the entire mission horizon. We further provide the agent with a position history to handle emergent state loops caused by the recharge capability. Our approach outperforms a baseline heuristic, generalizes to different target zones and maps, with limited generalization to unseen maps. We offer valuable insights into DRL algorithm design for long-horizon problems and provide a publicly available software framework for the CPP problem.

Reinforcement learning (RL, Sutton and Barto, 2018) aims to optimize an agent's long-term reward by learning an optimal policy that determines the best action to take under every circumstance. RL is closely related to the dynamic treatment regimens (DTR) or adaptive treatment strategies in statistical research for precision medicine (Murphy, 2003; Robins, 2004; Qian and Murphy, 2011; Kosorok and Moodie, 2015; Tsiatis et al., 2019; Qi et al., 2020; Zhou et al., 2022a), which seeks to obtain the optimal treatment policy in finite horizon settings with a few treatment stages that maximizes patients' expected outcome. Nevertheless, statistical methods for DTR mentioned above normally cannot handle large or infinite horizon settings. They require the number of trajectories to tend to infinity to achieve estimation consistency, unlike RL, which works even with finite number of trajectories under certain conditions. In addition to precision medicine, RL has been applied to various fields, such as games (Silver et al., 2016), ridesharing (Xu et al., 2018), mobile health (Liao et al., 2021) and robotics (Levine et al., 2020).

Multi-agent Deep Reinforcement Learning (MADRL) based traffic signal control becomes a popular research topic in recent years. To alleviate the scalability issue of completely centralized RL techniques and the non-stationarity issue of completely decentralized RL techniques on large-scale traffic networks, some literature utilizes a regional control approach where the whole network is firstly partitioned into multiple disjoint regions, followed by applying the centralized RL approach to each region. However, the existing partitioning rules either have no constraints on the topology of regions or require the same topology for all regions. Meanwhile, no existing regional control approach explores the performance of optimal joint action in an exponentially growing regional action space when intersections are controlled by 4-phase traffic signals (EW, EWL, NS, NSL). In this paper, we propose a novel RL training framework named RegionLight to tackle the above limitations. Specifically, the topology of regions is firstly constrained to a star network which comprises one center and an arbitrary number of leaves. Next, the network partitioning problem is modeled as an optimization problem to minimize the number of regions. Then, an Adaptive Branching Dueling Q-Network (ABDQ) model is proposed to decompose the regional control task into several joint signal control sub-tasks corresponding to particular intersections. Subsequently, these sub-tasks maximize the regional benefits cooperatively. Finally, the global control strategy for the whole network is obtained by concatenating the optimal joint actions of all regions. Experimental results demonstrate the superiority of our proposed framework over all baselines under both real and synthetic datasets in all evaluation metrics.

Modern quadrupeds are skillful in traversing or even sprinting on uneven terrains in a remote uncontrolled environment. However, survival in the wild requires not only maneuverability, but also the ability to handle potential critical hardware failures. How to grant such ability to quadrupeds is rarely investigated. In this paper, we propose a novel methodology to train and test hardware fault-tolerant controllers for quadruped locomotion, both in the simulation and physical world. We adopt the teacher-student reinforcement learning framework to train the controller with close-to-reality joint-locking failure in the simulation, which can be zero-shot transferred to the physical robot without any fine-tuning. Extensive experiments show that our fault-tolerant controller can efficiently lead a quadruped stably when it faces joint failures during locomotion.

Reinforcement learning algorithms often require finiteness of state and action spaces in Markov decision processes (MDPs) (also called controlled Markov chains) and various efforts have been made in the literature towards the applicability of such algorithms for continuous state and action spaces. In this paper, we show that under very mild regularity conditions (in particular, involving only weak continuity of the transition kernel of an MDP), Q-learning for standard Borel MDPs via quantization of states and actions (called Quantized Q-Learning) converges to a limit, and furthermore this limit satisfies an optimality equation which leads to near optimality with either explicit performance bounds or which are guaranteed to be asymptotically optimal. Our approach builds on (i) viewing quantization as a measurement kernel and thus a quantized MDP as a partially observed Markov decision process (POMDP), (ii) utilizing near optimality and convergence results of Q-learning for POMDPs, and (iii) finally, near-optimality of finite state model approximations for MDPs with weakly continuous kernels which we show to correspond to the fixed point of the constructed POMDP. Thus, our paper presents a very general convergence and approximation result for the applicability of Q-learning for continuous MDPs.

For many decades, experimental solid mechanics has played a crucial role in characterizing and understanding the mechanical properties of natural and novel materials. Recent advances in machine learning (ML) provide new opportunities for the field, including experimental design, data analysis, uncertainty quantification, and inverse problems. As the number of papers published in recent years in this emerging field is exploding, it is timely to conduct a comprehensive and up-to-date review of recent ML applications in experimental solid mechanics. Here, we first provide an overview of common ML algorithms and terminologies that are pertinent to this review, with emphasis placed on physics-informed and physics-based ML methods. Then, we provide thorough coverage of recent ML applications in traditional and emerging areas of experimental mechanics, including fracture mechanics, biomechanics, nano- and micro-mechanics, architected materials, and 2D material. Finally, we highlight some current challenges of applying ML to multi-modality and multi-fidelity experimental datasets and propose several future research directions. This review aims to provide valuable insights into the use of ML methods as well as a variety of examples for researchers in solid mechanics to integrate into their experiments.

Decisions made by machine learning models may have lasting impacts over time, making long-term fairness a crucial consideration. It has been shown that when ignoring the long-term effect, naively imposing fairness criterion in static settings can actually exacerbate bias over time. To explicitly address biases in sequential decision-making, recent works formulate long-term fairness notions in Markov Decision Process (MDP) framework. They define the long-term bias to be the sum of static bias over each time step. However, we demonstrate that naively summing up the step-wise bias can cause a false sense of fairness since it fails to consider the importance difference of different time steps during transition. In this work, we introduce a long-term fairness notion called Equal Long-term Benefit Rate (ELBERT), which explicitly considers varying temporal importance and adapts static fairness principles to the sequential setting. Moreover, we show that the policy gradient of Long-term Benefit Rate can be analytically reduced to standard policy gradient. This makes standard policy optimization methods applicable for reducing the bias, leading to our proposed bias mitigation method ELBERT-PO. Experiments on three sequential decision making environments show that ELBERT-PO significantly reduces bias and maintains high utility. Code is available at https://github.com/Yuancheng-Xu/ELBERT.

The utility of reinforcement learning is limited by the alignment of reward functions with the interests of human stakeholders. One promising method for alignment is to learn the reward function from human-generated preferences between pairs of trajectory segments, a type of reinforcement learning from human feedback (RLHF). These human preferences are typically assumed to be informed solely by partial return, the sum of rewards along each segment. We find this assumption to be flawed and propose modeling human preferences instead as informed by each segment's regret, a measure of a segment's deviation from optimal decision-making. Given infinitely many preferences generated according to regret, we prove that we can identify a reward function equivalent to the reward function that generated those preferences, and we prove that the previous partial return model lacks this identifiability property in multiple contexts. We empirically show that our proposed regret preference model outperforms the partial return preference model with finite training data in otherwise the same setting. Additionally, we find that our proposed regret preference model better predicts real human preferences and also learns reward functions from these preferences that lead to policies that are better human-aligned. Overall, this work establishes that the choice of preference model is impactful, and our proposed regret preference model provides an improvement upon a core assumption of recent research. We have open sourced our experimental code, the human preferences dataset we gathered, and our training and preference elicitation interfaces for gathering a such a dataset.