The remarkable success of reinforcement learning (RL) heavily relies on observing the reward of every visited state-action pair. In many real world applications, however, an agent can observe only a score that represents the quality of the whole trajectory, which is referred to as the {\em trajectory-wise reward}. In such a situation, it is difficult for standard RL methods to well utilize trajectory-wise reward, and large bias and variance errors can be incurred in policy evaluation. In this work, we propose a novel offline RL algorithm, called Pessimistic vAlue iteRaTion with rEward Decomposition (PARTED), which decomposes the trajectory return into per-step proxy rewards via least-squares-based reward redistribution, and then performs pessimistic value iteration based on the learned proxy reward. To ensure the value functions constructed by PARTED are always pessimistic with respect to the optimal ones, we design a new penalty term to offset the uncertainty of the proxy reward. For general episodic MDPs with large state space, we show that PARTED with overparameterized neural network function approximation achieves an $\tilde{\mathcal{O}}(D_{\text{eff}}H^2/\sqrt{N})$ suboptimality, where $H$ is the length of episode, $N$ is the total number of samples, and $D_{\text{eff}}$ is the effective dimension of the neural tangent kernel matrix. To further illustrate the result, we show that PARTED achieves an $\tilde{\mathcal{O}}(dH^3/\sqrt{N})$ suboptimality with linear MDPs, where $d$ is the feature dimension, which matches with that with neural network function approximation, when $D_{\text{eff}}=dH$. To the best of our knowledge, PARTED is the first offline RL algorithm that is provably efficient in general MDP with trajectory-wise reward.

We propose an end-to-end learning framework based on hierarchical reinforcement learning, called H-TSP, for addressing the large-scale Travelling Salesman Problem (TSP). The proposed H-TSP constructs a solution of a TSP instance starting from the scratch relying on two components: the upper-level policy chooses a small subset of nodes (up to 200 in our experiment) from all nodes that are to be traversed, while the lower-level policy takes the chosen nodes as input and outputs a tour connecting them to the existing partial route (initially only containing the depot). After jointly training the upper-level and lower-level policies, our approach can directly generate solutions for the given TSP instances without relying on any time-consuming search procedures. To demonstrate effectiveness of the proposed approach, we have conducted extensive experiments on randomly generated TSP instances with different numbers of nodes. We show that H-TSP can achieve comparable results (gap 3.42% vs. 7.32%) as SOTA search-based approaches, and more importantly, we reduce the time consumption up to two orders of magnitude (3.32s vs. 395.85s). To the best of our knowledge, H-TSP is the first end-to-end deep reinforcement learning approach that can scale to TSP instances of up to 10000 nodes. Although there are still gaps to SOTA results with respect to solution quality, we believe that H-TSP will be useful for practical applications, particularly those that are time-sensitive e.g., on-call routing and ride hailing service.

An appropriate reward function is of paramount importance in specifying a task in reinforcement learning (RL). Yet, it is known to be extremely challenging in practice to design a correct reward function for even simple tasks. Human-in-the-loop (HiL) RL allows humans to communicate complex goals to the RL agent by providing various types of feedback. However, despite achieving great empirical successes, HiL RL usually requires too much feedback from a human teacher and also suffers from insufficient theoretical understanding. In this paper, we focus on addressing this issue from a theoretical perspective, aiming to provide provably feedback-efficient algorithmic frameworks that take human-in-the-loop to specify rewards of given tasks. We provide an active-learning-based RL algorithm that first explores the environment without specifying a reward function and then asks a human teacher for only a few queries about the rewards of a task at some state-action pairs. After that, the algorithm guarantees to provide a nearly optimal policy for the task with high probability. We show that, even with the presence of random noise in the feedback, the algorithm only takes $\widetilde{O}(H{{\dim_{R}^2}})$ queries on the reward function to provide an $\epsilon$-optimal policy for any $\epsilon > 0$. Here $H$ is the horizon of the RL environment, and $\dim_{R}$ specifies the complexity of the function class representing the reward function. In contrast, standard RL algorithms require to query the reward function for at least $\Omega(\operatorname{poly}(d, 1/\epsilon))$ state-action pairs where $d$ depends on the complexity of the environmental transition.

One of the key challenges of Reinforcement Learning (RL) is the ability of agents to generalise their learned policy to unseen settings. Moreover, training RL agents requires large numbers of interactions with the environment. Motivated by the recent success of Offline RL and Imitation Learning (IL), we conduct a study to investigate whether agents can leverage offline data in the form of trajectories to improve the sample-efficiency in procedurally generated environments. We consider two settings of using IL from offline data for RL: (1) pre-training a policy before online RL training and (2) concurrently training a policy with online RL and IL from offline data. We analyse the impact of the quality (optimality of trajectories) and diversity (number of trajectories and covered level) of available offline trajectories on the effectiveness of both approaches. Across four well-known sparse reward tasks in the MiniGrid environment, we find that using IL for pre-training and concurrently during online RL training both consistently improve the sample-efficiency while converging to optimal policies. Furthermore, we show that pre-training a policy from as few as two trajectories can make the difference between learning an optimal policy at the end of online training and not learning at all. Our findings motivate the widespread adoption of IL for pre-training and concurrent IL in procedurally generated environments whenever offline trajectories are available or can be generated.

The success of transformer models trained with a language modeling objective brings a promising opportunity to the reinforcement learning framework. Decision Transformer (Chen et al., 2021) is a step towards this direction, showing how to train transformers with a similar next-step prediction objective on offline data. Another important development in this area is the recent emergence of large-scale datasets collected from the internet, such as the ones composed of tutorial videos with captions where people talk about what they are doing. To take advantage of this language component, we propose a novel method for unifying language reasoning with actions in a single policy. Specifically, we augment a transformer policy with word outputs, so it can generate textual captions interleaved with actions. When tested on the most challenging task in BabyAI, with captions describing next subgoals, our reasoning policy consistently outperforms the caption-free baseline.

Safe reinforcement learning (RL) aims to solve an optimal control problem under safety constraints. Existing $\textit{direct}$ safe RL methods use the original constraint throughout the learning process. They either lack theoretical guarantees of the policy during iteration or suffer from infeasibility problems. To address this issue, we propose an $\textit{indirect}$ safe RL method called feasible policy iteration (FPI) that iteratively uses the feasible region of the last policy to constrain the current policy. The feasible region is represented by a feasibility function called constraint decay function (CDF). The core of FPI is a region-wise policy update rule called feasible policy improvement, which maximizes the return under the constraint of the CDF inside the feasible region and minimizes the CDF outside the feasible region. This update rule is always feasible and ensures that the feasible region monotonically expands and the state-value function monotonically increases inside the feasible region. Using the feasible Bellman equation, we prove that FPI converges to the maximum feasible region and the optimal state-value function. Experiments on classic control tasks and Safety Gym show that our algorithms achieve lower constraint violations and comparable or higher performance than the baselines.

We apply Bayesian optimization and reinforcement learning to a problem in topology: the question of when a knot bounds a ribbon disk. This question is relevant in an approach to disproving the four-dimensional smooth Poincar\'e conjecture; using our programs, we rule out many potential counterexamples to the conjecture. We also show that the programs are successful in detecting many ribbon knots in the range of up to 70 crossings.

The utilization of broad datasets has proven to be crucial for generalization for a wide range of fields. However, how to effectively make use of diverse multi-task data for novel downstream tasks still remains a grand challenge in robotics. To tackle this challenge, we introduce a framework that acquires goal-conditioned policies for unseen temporally extended tasks via offline reinforcement learning on broad data, in combination with online fine-tuning guided by subgoals in learned lossy representation space. When faced with a novel task goal, the framework uses an affordance model to plan a sequence of lossy representations as subgoals that decomposes the original task into easier problems. Learned from the broad data, the lossy representation emphasizes task-relevant information about states and goals while abstracting away redundant contexts that hinder generalization. It thus enables subgoal planning for unseen tasks, provides a compact input to the policy, and facilitates reward shaping during fine-tuning. We show that our framework can be pre-trained on large-scale datasets of robot experiences from prior work and efficiently fine-tuned for novel tasks, entirely from visual inputs without any manual reward engineering.

General-purpose robots require diverse repertoires of behaviors to complete challenging tasks in real-world unstructured environments. To address this issue, goal-conditioned reinforcement learning aims to acquire policies that can reach configurable goals for a wide range of tasks on command. However, such goal-conditioned policies are notoriously difficult and time-consuming to train from scratch. In this paper, we propose Planning to Practice (PTP), a method that makes it practical to train goal-conditioned policies for long-horizon tasks that require multiple distinct types of interactions to solve. Our approach is based on two key ideas. First, we decompose the goal-reaching problem hierarchically, with a high-level planner that sets intermediate subgoals using conditional subgoal generators in the latent space for a low-level model-free policy. Second, we propose a hybrid approach which first pre-trains both the conditional subgoal generator and the policy on previously collected data through offline reinforcement learning, and then fine-tunes the policy via online exploration. This fine-tuning process is itself facilitated by the planned subgoals, which breaks down the original target task into short-horizon goal-reaching tasks that are significantly easier to learn. We conduct experiments in both the simulation and real world, in which the policy is pre-trained on demonstrations of short primitive behaviors and fine-tuned for temporally extended tasks that are unseen in the offline data. Our experimental results show that PTP can generate feasible sequences of subgoals that enable the policy to efficiently solve the target tasks.

In recent years, reinforcement learning (RL) has acquired a prominent position in the space of health-related sequential decision-making, becoming an increasingly popular tool for delivering adaptive interventions (AIs). However, despite potential benefits, its real-life application is still limited, partly due to a poor synergy between the methodological and the applied communities. In this work, we provide the first unified survey on RL methods for learning AIs, using the common methodological umbrella of RL to bridge the two AI areas of dynamic treatment regimes and just-in-time adaptive interventions in mobile health. We outline similarities and differences between these two AI domains and discuss their implications for using RL. Finally, we leverage our experience in designing case studies in both areas to illustrate the tremendous collaboration opportunities between statistical, RL, and healthcare researchers in the space of AIs.