Designing efficient algorithms to learn and plan in the sequential decision-making environment modeled by a Markov decision process (MDP) is one of the main tasks in reinforcement learning (RL). However, traditional tabular RL algorithms suffer from the curse-of-dimensionality due to the large size of the state and action spaces in practice. To enable learning in high-dimensional state and action spaces, using a predefined function class to approximate the underlying transition dynamic or the value function is a common approach. Most existing works for RL with function approximation focus on simple linear function classes such as the linear mixture MDP (Modi et al., 2020; Ayoub et al., 2020; Zhou et al., 2021b), which can replace the size of the state and action spaces with the dimension of the linear function class. However, these assumptions are often too restrictive to hold in practice. Recently, a line of works (Russo and Van Roy, 2013; Du et al., 2021; Jin et al., 2021) emerged that studies RL with general function approximation, introducing new complexity measures for the general function class and proposing new algorithms with regret bounds or PAC guarantees in terms of the complexity of the general function class. All existing results of RL with a general function class are limited to either regret bounds or PAC sample complexity, both of which cannot ensure convergence to the optimal policy up to arbitrary accuracy.

Brachiation is a dynamic, coordinated swinging maneuver of body and arms used by monkeys and apes to move between branches. As a unique underactuated mode of locomotion, it is interesting to study from a robotics perspective since it can broaden the deployment scenarios for humanoids and animaloids. While several brachiating robots of varying complexity have been proposed in the past, this paper presents the simplest possible prototype of a brachiation robot, using only a single actuator and unactuated grippers. The novel passive gripper design allows it to snap on and release from monkey bars, while guaranteeing well defined start and end poses of the swing. The brachiation behavior is realized in three different ways, using trajectory optimization via direct collocation and stabilization by a model-based time-varying linear quadratic regulator (TVLQR) or model-free proportional derivative (PD) control, as well as by a reinforcement learning (RL) based control policy. The three control schemes are compared in terms of robustness to disturbances, mass uncertainty, and energy consumption. The system design and controllers have been open-sourced. Due to its minimal and open design, the system can serve as a canonical underactuated platform for education and research.

This paper presents an approach for data-driven policy refinement in reinforcement learning, specifically designed for safety-critical applications. Our methodology leverages the strengths of data-driven optimization and reinforcement learning to enhance policy safety and optimality through iterative refinement. Our principal contribution lies in the mathematical formulation of this data-driven policy refinement concept. This framework systematically improves reinforcement learning policies by learning from counterexamples identified during data-driven verification. Furthermore, we present a series of theorems elucidating key theoretical properties of our approach, including convergence, robustness bounds, generalization error, and resilience to model mismatch. These results not only validate the effectiveness of our methodology but also contribute to a deeper understanding of its behavior in different environments and scenarios.

Optimization problems over dynamic networks have been extensively studied and widely used in the past decades to formulate numerous real-world problems. However, (1) traditional optimization-based approaches do not scale to large networks, and (2) the design of good heuristics or approximation algorithms often requires significant manual trial-and-error. In this work, we argue that data-driven strategies can automate this process and learn efficient algorithms without compromising optimality. To do so, we present network control problems through the lens of reinforcement learning and propose a graph network-based framework to handle a broad class of problems. Instead of naively computing actions over high-dimensional graph elements, e.g., edges, we propose a bi-level formulation where we (1) specify a desired next state via RL, and (2) solve a convex program to best achieve it, leading to drastically improved scalability and performance. We further highlight a collection of desirable features to system designers, investigate design decisions, and present experiments on real-world control problems showing the utility, scalability, and flexibility of our framework.

This paper proposes a spatial-temporal recurrent neural network architecture for deep $Q$-networks that can be used to steer an autonomous ship. The network design makes it possible to handle an arbitrary number of surrounding target ships while offering robustness to partial observability. Furthermore, a state-of-the-art collision risk metric is proposed to enable an easier assessment of different situations by the agent. The COLREG rules of maritime traffic are explicitly considered in the design of the reward function. The final policy is validated on a custom set of newly created single-ship encounters called `Around the Clock' problems and the commonly used Imazu (1987) problems, which include 18 multi-ship scenarios. Performance comparisons with artificial potential field and velocity obstacle methods demonstrate the potential of the proposed approach for maritime path planning. Furthermore, the new architecture exhibits robustness when it is deployed in multi-agent scenarios and it is compatible with other deep reinforcement learning algorithms, including actor-critic frameworks.

Traditionally, approximate dynamic programming is employed in dialogue generation with greedy policy improvement through action sampling, as the natural language action space is vast. However, this practice is inefficient for reinforcement learning (RL) due to the sparsity of eligible responses with high action values, which leads to weak improvement sustained by random sampling. This paper presents theoretical analysis and experiments that reveal the performance of the dialogue policy is positively correlated with the sampling size. To overcome this limitation, we introduce a novel dual-granularity Q-function that explores the most promising response category to intervene in the sampling process. Our approach extracts actions based on a grained hierarchy, thereby achieving the optimum with fewer policy iterations. Additionally, we use offline RL and learn from multiple reward functions designed to capture emotional nuances in human interactions. Empirical studies demonstrate that our algorithm outperforms baselines across automatic metrics and human evaluations. Further testing reveals that our algorithm exhibits both explainability and controllability and generates responses with higher expected rewards.

Recent studies have shown that episodic reinforcement learning (RL) is no harder than bandits when the total reward is bounded by $1$, and proved regret bounds that have a polylogarithmic dependence on the planning horizon $H$. However, it remains an open question that if such results can be carried over to adversarial RL, where the reward is adversarially chosen at each episode. In this paper, we answer this question affirmatively by proposing the first horizon-free policy search algorithm. To tackle the challenges caused by exploration and adversarially chosen reward, our algorithm employs (1) a variance-uncertainty-aware weighted least square estimator for the transition kernel; and (2) an occupancy measure-based technique for the online search of a \emph{stochastic} policy. We show that our algorithm achieves an $\tilde{O}\big((d+\log (|\mathcal{S}|^2 |\mathcal{A}|))\sqrt{K}\big)$ regret with full-information feedback, where $d$ is the dimension of a known feature mapping linearly parametrizing the unknown transition kernel of the MDP, $K$ is the number of episodes, $|\mathcal{S}|$ and $|\mathcal{A}|$ are the cardinalities of the state and action spaces. We also provide hardness results and regret lower bounds to justify the near optimality of our algorithm and the unavoidability of $\log|\mathcal{S}|$ and $\log|\mathcal{A}|$ in the regret bound.

This paper considers multi-agent reinforcement learning (MARL) where the rewards are received after delays and the delay time varies across agents and across time steps. Based on the V-learning framework, this paper proposes MARL algorithms that efficiently deal with reward delays. When the delays are finite, our algorithm reaches a coarse correlated equilibrium (CCE) with rate $\tilde{\mathcal{O}}(\frac{H^3\sqrt{S\mathcal{T}_K}}{K}+\frac{H^3\sqrt{SA}}{\sqrt{K}})$ where $K$ is the number of episodes, $H$ is the planning horizon, $S$ is the size of the state space, $A$ is the size of the largest action space, and $\mathcal{T}_K$ is the measure of total delay formally defined in the paper. Moreover, our algorithm is extended to cases with infinite delays through a reward skipping scheme. It achieves convergence rate similar to the finite delay case.

The Adaptive Large Neighborhood Search (ALNS) algorithm has shown considerable success in solving complex combinatorial optimization problems (COPs). ALNS selects various heuristics adaptively during the search process, leveraging their strengths to find good solutions for optimization problems. However, the effectiveness of ALNS depends on the proper configuration of its selection and acceptance parameters. To address this limitation, we propose a Deep Reinforcement Learning (DRL) approach that selects heuristics, adjusts parameters, and controls the acceptance criteria during the search process. The proposed method aims to learn, based on the state of the search, how to configure the next iteration of the ALNS to obtain good solutions to the underlying optimization problem. We evaluate the proposed method on a time-dependent orienteering problem with stochastic weights and time windows, used in an IJCAI competition. The results show that our approach outperforms vanilla ALNS and ALNS tuned with Bayesian Optimization. In addition, it obtained better solutions than two state-of-the-art DRL approaches, which are the winning methods of the competition, with much fewer observations required for training. The implementation of our approach will be made publicly available.

Apprenticeship learning (AL) is a process of inducing effective decision-making policies via observing and imitating experts' demonstrations. Most existing AL approaches, however, are not designed to cope with the evolving reward functions commonly found in human-centric tasks such as healthcare, where offline learning is required. In this paper, we propose an offline Time-aware Hierarchical EM Energy-based Sub-trajectory (THEMES) AL framework to tackle the evolving reward functions in such tasks. The effectiveness of THEMES is evaluated via a challenging task -- sepsis treatment. The experimental results demonstrate that THEMES can significantly outperform competitive state-of-the-art baselines.