Reinforcement learning (RL) has drawn increasing interests in recent years due to its tremendous success in various applications. However, standard RL algorithms can only be applied for single reward function, and cannot adapt to an unseen reward function quickly. In this paper, we advocate a general operator view of reinforcement learning, which enables us to directly approximate the operator that maps from reward function to value function. The benefit of learning the operator is that we can incorporate any new reward function as input and attain its corresponding value function in a zero-shot manner. To approximate this special type of operator, we design a number of novel operator neural network architectures based on its theoretical properties. Our design of operator networks outperform the existing methods and the standard design of general purpose operator network, and we demonstrate the benefit of our operator deep Q-learning framework in several tasks including reward transferring for offline policy evaluation (OPE) and reward transferring for offline policy optimization in a range of tasks.

Study of reinforcement learning algorithms for sequential learning-based decision-making in unknown linear systems has become increasingly popular in the recent years. In the canonical version of the problem, the true dynamics matrices of the plant are unknown, and the goal consists of adaptive design of the control input for minimizing deviations from optimal policy. Still, the control actions must be diverse enough to lead to accurate identification of the unknown parameters [1]. The existing literature is notably rich, including adaptive policies based on optimistic approximations of the dynamics matrices over a confidence region [2, 3], as well as plugin estimates of unknown parameters after leveraging a dither signal [4, 5, 6], Bayesian approaches [7, 8, 9], and statistical bootstrap [10]. An important problem in in different areas of control theory is that of stabilization.

In this paper, we use concepts from supervisory control theory of discrete event systems to propose a method to learn optimal control policies for a finite-state Markov Decision Process (MDP) in which (only) certain sequences of actions are deemed unsafe (respectively safe). We assume that the set of action sequences that are deemed unsafe and/or safe are given in terms of a finite-state automaton; and propose a supervisor that disables a subset of actions at every state of the MDP so that the constraints on action sequence are satisfied. Then we present a version of the Q-learning algorithm for learning optimal policies in the presence of non-Markovian action-sequence and state constraints, where we use the development of reward machines to handle the state constraints. We illustrate the method using an example that captures the utility of automata-based methods for non-Markovian state and action specifications for reinforcement learning and show the results of simulations in this setting.

Stochastic and soft optimal policies resulting from entropy-regularized Markov decision processes (ER-MDP) are desirable for exploration and imitation learning applications. Motivated by the fact that such policies are sensitive with respect to the state transition probabilities, and the estimation of these probabilities may be inaccurate, we study a robust version of the ER-MDP model, where the stochastic optimal policies are required to be robust with respect to the ambiguity in the underlying transition probabilities. Our work is at the crossroads of two important schemes in reinforcement learning (RL), namely, robust MDP and entropy regularized MDP. We show that essential properties that hold for the non-robust ER-MDP and robust unregularized MDP models also hold in our settings, making the robust ER-MDP problem tractable. We show how our framework and results can be integrated into different algorithmic schemes including value or (modified) policy iteration, which would lead to new robust RL and inverse RL algorithms to handle uncertainties. Analyses on computational complexity and error propagation under conventional uncertainty settings are also provided.

We hypothesize that empirically studying the sample complexity of offline reinforcement learning (RL) is crucial for the practical applications of RL in the real world. Several recent works have demonstrated the ability to learn policies directly from offline data. In this work, we ask the question of the dependency on the number of samples for learning from offline data. Our objective is to emphasize that studying sample complexity for offline RL is important, and is an indicator of the usefulness of existing offline algorithms. We propose an evaluation approach for sample complexity analysis of offline RL.

In recent years, gradient based Meta-RL (GMRL) methods have achieved remarkable successes in either discovering effective online hyperparameter for one single task (Xu et al., 2018) or learning good initialisation for multi-task transfer learning (Finn et al., 2017). Despite the empirical successes, it is often neglected that computing meta gradients via vanilla backpropagation is ill-defined. In this paper, we argue that the stochastic meta-gradient estimation adopted by many existing MGRL methods are in fact biased; the bias comes from two sources: 1) the compositional bias that is inborn in the structure of compositional optimisation problems and 2) the bias of multi-step Hessian estimation caused by direct automatic differentiation. To better understand the meta gradient biases, we perform the first of its kind study to quantify the amount for each of them. We start by providing a unifying derivation for existing GMRL algorithms, and then theoretically analyse both the bias and the variance of existing gradient estimation methods. On understanding the underlying principles of bias, we propose two mitigation solutions based on off-policy correction and multi-step Hessian estimation techniques. Comprehensive ablation studies have been conducted and results reveals: (1) The existence of these two biases and how they influence the meta-gradient estimation when combined with different estimator/sample size/step and learning rate. (2) The effectiveness of these mitigation approaches for meta-gradient estimation and thereby the final return on two practical Meta-RL algorithms: LOLA-DiCE and Meta-gradient Reinforcement Learning.

Shared autonomy refers to approaches for enabling an autonomous agent to collaborate with a human with the aim of improving human performance. However, besides improving performance, it may often also be beneficial that the agent concurrently accounts for preserving the user's experience or satisfaction of collaboration. In order to address this additional goal, we examine approaches for improving the user experience by constraining the number of interventions by the autonomous agent. We propose two model-free reinforcement learning methods that can account for both hard and soft constraints on the number of interventions. We show that not only does our method outperform the existing baseline, but also eliminates the need to manually tune a black-box hyperparameter for controlling the level of assistance. We also provide an in-depth analysis of intervention scenarios in order to further illuminate system understanding.

Inferring reward functions from demonstrations and pairwise preferences are auspicious approaches for aligning Reinforcement Learning (RL) agents with human intentions. However, state-of-the art methods typically focus on learning a single reward model, thus rendering it difficult to trade off different reward functions from multiple experts. We propose Multi-Objective Reinforced Active Learning (MORAL), a novel method for combining diverse demonstrations of social norms into a Pareto-optimal policy. Through maintaining a distribution over scalarization weights, our approach is able to interactively tune a deep RL agent towards a variety of preferences, while eliminating the need for computing multiple policies. We empirically demonstrate the effectiveness of MORAL in two scenarios, which model a delivery and an emergency task that require an agent to act in the presence of normative conflicts. Overall, we consider our research a step towards multi-objective RL with learned rewards, bridging the gap between current reward learning and machine ethics literature.

There is a general trend of applying reinforcement learning (RL) techniques for traffic signal control (TSC). Recently, most studies pay attention to the neural network design and rarely concentrate on the state representation. Does the design of state representation has a good impact on TSC? In this paper, we (1) propose an effective state representation as queue length of vehicles with intensive knowledge; (2) present a TSC method called MaxQueue based on our state representation approach; (3) develop a general RL-based TSC template called QL-XLight with queue length as state and reward and generate QL-FRAP, QL-CoLight, and QL-DQN by our QL-XLight template based on traditional and latest RL models.Through comprehensive experiments on multiple real-world datasets, we demonstrate that: (1) our MaxQueue method outperforms the latest RL based methods; (2) QL-FRAP and QL-CoLight achieves a new state-of-the-art (SOTA). In general, state representation with intensive knowledge is also essential for TSC methods. Our code is released on Github.

This work explores how to learn robust and generalizable state representation from image-based observations with deep reinforcement learning methods. Addressing the computational complexity, stringent assumptions, and representation collapse challenges in the existing work of bisimulation metric, we devise Simple State Representation (SimSR) operator, which achieves equivalent functionality while reducing the complexity by an order in comparison with bisimulation metric. SimSR enables us to design a stochastic-approximation-based method that can practically learn the mapping functions (encoders) from observations to latent representation space. Besides the theoretical analysis, we experimented and compared our work with recent state-of-the-art solutions in visual MuJoCo tasks. The results show that our model generally achieves better performance and has better robustness and good generalization.