We study online learning in episodic constrained Markov decision processes (CMDPs), where the goal of the learner is to collect as much reward as possible over the episodes, while guaranteeing that some long-term constraints are satisfied during the learning process. Rewards and constraints can be selected either stochastically or adversarially, and the transition function is not known to the learner. While online learning in classical unconstrained MDPs has received considerable attention over the last years, the setting of CMDPs is still largely unexplored. This is surprising, since in real-world applications, such as, e.g., autonomous driving, automated bidding, and recommender systems, there are usually additional constraints and specifications that an agent has to obey during the learning process. In this paper, we provide the first best-of-both-worlds algorithm for CMDPs with long-term constraints. Our algorithm is capable of handling settings in which rewards and constraints are selected either stochastically or adversarially, without requiring any knowledge of the underling process. Moreover, our algorithm matches state-of-the-art regret and constraint violation bounds for settings in which constraints are selected stochastically, while it is the first to provide guarantees in the case in which they are chosen adversarially.

A crucial challenge in reinforcement learning is to reduce the number of interactions with the environment that an agent requires to master a given task. Transfer learning proposes to address this issue by re-using knowledge from previously learned tasks. However, determining which source task qualifies as the most appropriate for knowledge extraction, as well as the choice regarding which algorithm components to transfer, represent severe obstacles to its application in reinforcement learning. The goal of this paper is to address these issues with modular multi-source transfer learning techniques. The proposed techniques automatically learn how to extract useful information from source tasks, regardless of the difference in state-action space and reward function. We support our claims with extensive and challenging cross-domain experiments for visual control.

Deep reinforcement learning (RL) has been endowed with high expectations in tackling challenging manipulation tasks in an autonomous and self-directed fashion. Despite the significant strides made in the development of reinforcement learning, the practical deployment of this paradigm is hindered by at least two barriers, namely, the engineering of a reward function and ensuring the safety guaranty of learning-based controllers. In this paper, we address these challenging limitations by proposing a framework that merges a reinforcement learning \lstinline[columns=fixed]{planner} that is trained using sparse rewards with a model predictive controller (MPC) \lstinline[columns=fixed]{actor}, thereby offering a safe policy. On the one hand, the RL \lstinline[columns=fixed]{planner} learns from sparse rewards by selecting intermediate goals that are easy to achieve in the short term and promising to lead to target goals in the long term. On the other hand, the MPC \lstinline[columns=fixed]{actor} takes the suggested intermediate goals from the RL \lstinline[columns=fixed]{planner} as the input and predicts how the robot's action will enable it to reach that goal while avoiding any obstacles over a short period of time. We evaluated our method on four challenging manipulation tasks with dynamic obstacles and the results demonstrate that, by leveraging the complementary strengths of these two components, the agent can solve manipulation tasks in complex, dynamic environments safely with a $100\%$ success rate. Videos are available at \url{https://videoviewsite.wixsite.com/mpc-hgg}.

Reinforcement learning is an attractive approach to learn good resource allocation and scheduling policies based on data when the system model is unknown. However, the cumulative regret of most RL algorithms scales as $\tilde O(\mathsf{S} \sqrt{\mathsf{A} T})$, where $\mathsf{S}$ is the size of the state space, $\mathsf{A}$ is the size of the action space, $T$ is the horizon, and the $\tilde{O}(\cdot)$ notation hides logarithmic terms. Due to the linear dependence on the size of the state space, these regret bounds are prohibitively large for resource allocation and scheduling problems. In this paper, we present a model-based RL algorithm for such problems which has scalable regret. In particular, we consider a restless bandit model, and propose a Thompson-sampling based learning algorithm which is tuned to the underlying structure of the model. We present two characterizations of the regret of the proposed algorithm with respect to the Whittle index policy. First, we show that for a restless bandit with $n$ arms and at most $m$ activations at each time, the regret scales either as $\tilde{O}(mn\sqrt{T})$ or $\tilde{O}(n^2 \sqrt{T})$ depending on the reward model. Second, under an additional technical assumption, we show that the regret scales as $\tilde{O}(n^{1.5} \sqrt{T})$ or $\tilde{O}(\max\{m\sqrt{n}, n\} \sqrt{T})$. We present numerical examples to illustrate the salient features of the algorithm.

Deep reinforcement learning has achieved impressive results in recent years; yet, it is still severely troubled by environments showcasing sparse rewards. On top of that, not all sparse-reward environments are created equal; in other words, they can differ in the presence or absence of various features, with many of them having a great impact on learning. In light of this, the present work puts together a literature compilation of such environmental features, covering particularly those that have been taken advantage of and those that continue to pose a challenge. We expect this effort to provide guidance to researchers for assessing the generality of their new proposals and to call their attention to issues that remain unresolved when dealing with sparse rewards.

In this paper, we define, evaluate, and improve the "relay-generalization" performance of reinforcement learning (RL) agents on the out-of-distribution "controllable" states. Ideally, an RL agent that generally masters a task should reach its goal starting from any controllable state of the environment instead of memorizing a small set of trajectories. For example, a self-driving system should be able to take over the control from humans in the middle of driving and must continue to drive the car safely. To practically evaluate this type of generalization, we start the test agent from the middle of other independently well-trained stranger agents' trajectories. With extensive experimental evaluation, we show the prevalence of generalization failure on controllable states from stranger agents. For example, in the Humanoid environment, we observed that a well-trained Proximal Policy Optimization (PPO) agent, with only 3.9% failure rate during regular testing, failed on 81.6% of the states generated by well-trained stranger PPO agents. To improve "relay generalization," we propose a novel method called Self-Trajectory Augmentation (STA), which will reset the environment to the agent's old states according to the Q function during training. After applying STA to the Soft Actor Critic's (SAC) training procedure, we reduced the failure rate of SAC under relay-evaluation by more than three times in most settings without impacting agent performance and increasing the needed number of environment interactions. Our code is available at https://github.com/lan-lc/STA. Generalization is critical for deploying reinforcement learning (RL) agents into real-world applications. A well-trained RL agent that can achieve high rewards under restricted settings may not be able to handle the enormous state space and complex environment variations in the real world. There are many different aspects regarding the generalization of RL agents.

Sequential decision making algorithms often struggle to leverage different sources of unstructured offline interaction data. Imitation learning (IL) methods based on supervised learning are robust, but require optimal demonstrations, which are hard to collect. Offline goal-conditioned reinforcement learning (RL) algorithms promise to learn from sub-optimal data, but face optimization challenges especially with high-dimensional data. To bridge the gap between IL and RL, we introduce Distance Weighted Supervised Learning or DWSL, a supervised method for learning goal-conditioned policies from offline data. DWSL models the entire distribution of time-steps between states in offline data with only supervised learning, and uses this distribution to approximate shortest path distances. To extract a policy, we weight actions by their reduction in distance estimates. Theoretically, DWSL converges to an optimal policy constrained to the data distribution, an attractive property for offline learning, without any bootstrapping. Across all datasets we test, DWSL empirically maintains behavior cloning as a lower bound while still exhibiting policy improvement. In high-dimensional image domains, DWSL surpasses the performance of both prior goal-conditioned IL and RL algorithms. Visualizations and code can be found at https://sites.google.com/view/dwsl/home .

Constrained multiagent reinforcement learning (C-MARL) is gaining importance as MARL algorithms find new applications in real-world systems ranging from energy systems to drone swarms. Most C-MARL algorithms use a primal-dual approach to enforce constraints through a penalty function added to the reward. In this paper, we study the structural effects of this penalty term on the MARL problem. First, we show that the standard practice of using the constraint function as the penalty leads to a weak notion of safety. However, by making simple modifications to the penalty term, we can enforce meaningful probabilistic (chance and conditional value at risk) constraints. Second, we quantify the effect of the penalty term on the value function, uncovering an improved value estimation procedure. We use these insights to propose a constrained multiagent advantage actor critic (C-MAA2C) algorithm. Simulations in a simple constrained multiagent environment affirm that our reinterpretation of the primal-dual method in terms of probabilistic constraints is effective, and that our proposed value estimate accelerates convergence to a safe joint policy.

Language models pre-trained on large self-supervised corpora, followed by task-specific fine-tuning has become the dominant paradigm in NLP. These pre-training datasets often have a one-to-many structure--e.g. in dialogue there are many valid responses for a given context. However, only some of these responses will be desirable in our downstream task. This raises the question of how we should train the model such that it can emulate the desirable behaviours, but not the undesirable ones. Current approaches train in a one-to-one setup--only a single target response is given for a single dialogue context--leading to models only learning to predict the average response, while ignoring the full range of possible responses. Using text-based games as a testbed, our approach, PASA, uses discrete latent variables to capture the range of different behaviours represented in our larger pre-training dataset. We then use knowledge distillation to distil the posterior probability distribution into a student model. This probability distribution is far richer than learning from only the hard targets of the dataset, and thus allows the student model to benefit from the richer range of actions the teacher model has learned. Results show up to 49% empirical improvement over the previous state-of-the-art model on the Jericho Walkthroughs dataset.

We investigate whether Deep Reinforcement Learning (Deep RL) is able to synthesize sophisticated and safe movement skills for a low-cost, miniature humanoid robot that can be composed into complex behavioral strategies in dynamic environments. We used Deep RL to train a humanoid robot with 20 actuated joints to play a simplified one-versus-one (1v1) soccer game. We first trained individual skills in isolation and then composed those skills end-to-end in a self-play setting. The resulting policy exhibits robust and dynamic movement skills such as rapid fall recovery, walking, turning, kicking and more; and transitions between them in a smooth, stable, and efficient manner - well beyond what is intuitively expected from the robot. The agents also developed a basic strategic understanding of the game, and learned, for instance, to anticipate ball movements and to block opponent shots. The full range of behaviors emerged from a small set of simple rewards. Our agents were trained in simulation and transferred to real robots zero-shot. We found that a combination of sufficiently high-frequency control, targeted dynamics randomization, and perturbations during training in simulation enabled good-quality transfer, despite significant unmodeled effects and variations across robot instances. Although the robots are inherently fragile, minor hardware modifications together with basic regularization of the behavior during training led the robots to learn safe and effective movements while still performing in a dynamic and agile way. Indeed, even though the agents were optimized for scoring, in experiments they walked 156% faster, took 63% less time to get up, and kicked 24% faster than a scripted baseline, while efficiently combining the skills to achieve the longer term objectives. Examples of the emergent behaviors and full 1v1 matches are available on the supplementary website.