While various multi-agent reinforcement learning methods have been proposed in cooperative settings, few works investigate how self-interested learning agents achieve mutual coordination in decentralized general-sum games and generalize pre-trained policies to non-cooperative opponents during execution. In this paper, we present Generalizable Risk-Sensitive Policy (GRSP). GRSP learns the distributions over agent's return and estimate a dynamic risk-seeking bonus to discover risky coordination strategies. Furthermore, to avoid overfitting to training opponents, GRSP learns an auxiliary opponent modeling task to infer opponents' types and dynamically alter corresponding strategies during execution. Empirically, agents trained via GRSP can achieve mutual coordination during training stably and avoid being exploited by non-cooperative opponents during execution. To the best of our knowledge, it is the first method to learn coordination strategies between agents both in iterated prisoner's dilemma (IPD) and iterated stag hunt (ISH) without shaping opponents or rewards, and firstly consider generalization during execution. Furthermore, we show that GRSP can be scaled to high-dimensional settings.

The offline reinforcement learning (RL) problem is often motivated by the need to learn data-driven decision policies in financial, legal and healthcare applications. However, the learned policy could retain sensitive information of individuals in the training data (e.g., treatment and outcome of patients), thus susceptible to various privacy risks. We design offline RL algorithms with differential privacy guarantees which provably prevent such risks. These algorithms also enjoy strong instance-dependent learning bounds under both tabular and linear Markov decision process (MDP) settings. Our theory and simulation suggest that the privacy guarantee comes at (almost) no drop in utility comparing to the non-private counterpart for a medium-size dataset.

Audio-visual navigation combines sight and hearing to navigate to a sound-emitting source in an unmapped environment. While recent approaches have demonstrated the benefits of audio input to detect and find the goal, they focus on clean and static sound sources and struggle to generalize to unheard sounds. In this work, we propose the novel dynamic audio-visual navigation benchmark which requires catching a moving sound source in an environment with noisy and distracting sounds, posing a range of new challenges. We introduce a reinforcement learning approach that learns a robust navigation policy for these complex settings. To achieve this, we propose an architecture that fuses audio-visual information in the spatial feature space to learn correlations of geometric information inherent in both local maps and audio signals. We demonstrate that our approach consistently outperforms the current state-of-the-art by a large margin across all tasks of moving sounds, unheard sounds, and noisy environments, on two challenging 3D scanned real-world environments, namely Matterport3D and Replica. The benchmark is available at http://dav-nav.cs.uni-freiburg.de.

This document is a concise summary of many key results in single-agent reinforcement learning (RL). The intended audience are those who already have some familiarity with RL and are looking to review, reference and/or remind themselves of important ideas in the field.

Precision medicine focuses on providing personalized treatment to patients by taking their personal information into consideration (see e.g., Kosorok and Laber, 2019; Tsiatis et al., 2019). It has found various applications in numerous studies, ranging from the cardiovascular disease study to cancer treatment and gene therapy (Jameson and Longo, 2015). A dynamic treatment regime (DTR) consists of a sequence of treatment decisions rules tailored to each individual patient's status at each time, mathematically formulating the idea behind precision medicine. One of the major objectives in precision medicine is to identify the optimal dynamic treatment regime that yields the most favorable outcome on average. With the rapidly development of mobile health (mHealth) technology, it becomes feasible to collect rich longitudinal data through mobile apps in medical studies.

In this paper, hypernetworks are trained to generate behaviors across a range of unseen task conditions, via a novel TD-based training objective and data from a set of near-optimal RL solutions for training tasks. This work relates to meta RL, contextual RL, and transfer learning, with a particular focus on zero-shot performance at test time, enabled by knowledge of the task parameters (also known as context). Our technical approach is based upon viewing each RL algorithm as a mapping from the MDP specifics to the near-optimal value function and policy and seek to approximate it with a hypernetwork that can generate near-optimal value functions and policies, given the parameters of the MDP. We show that, under certain conditions, this mapping can be considered as a supervised learning problem. We empirically evaluate the effectiveness of our method for zero-shot transfer to new reward and transition dynamics on a series of continuous control tasks from DeepMind Control Suite. Our method demonstrates significant improvements over baselines from multitask and meta RL approaches.

General nonlinear sieve learnings are classes of nonlinear sieves that can approximate nonlinear functions of high dimensional variables much more flexibly than various linear sieves (or series). This paper considers general nonlinear sieve quasi-likelihood ratio (GN-QLR) based inference on expectation functionals of time series data, where the functionals of interest are based on some nonparametric function that satisfy conditional moment restrictions and are learned using multilayer neural networks. While the asymptotic normality of the estimated functionals depends on some unknown Riesz representer of the functional space, we show that the optimally weighted GN-QLR statistic is asymptotically Chi-square distributed, regardless whether the expectation functional is regular (root-$n$ estimable) or not. This holds when the data are weakly dependent beta-mixing condition. We apply our method to the off-policy evaluation in reinforcement learning, by formulating the Bellman equation into the conditional moment restriction framework, so that we can make inference about the state-specific value functional using the proposed GN-QLR method with time series data. In addition, estimating the averaged partial means and averaged partial derivatives of nonparametric instrumental variables and quantile IV models are also presented as leading examples. Finally, a Monte Carlo study shows the finite sample performance of the procedure

Reinforcement learning (RL)-based driver assistance systems seek to improve fuel consumption via continual improvement of powertrain control actions considering experiential data from the field. However, the need to explore diverse experiences in order to learn optimal policies often limits the application of RL techniques in safety-critical systems like vehicle control. In this paper, an exponential control barrier function (ECBF) is derived and utilized to filter unsafe actions proposed by an RL-based driver assistance system. The RL agent freely explores and optimizes the performance objectives while unsafe actions are projected to the closest actions in the safe domain. The reward is structured so that driver's acceleration requests are met in a manner that boosts fuel economy and doesn't compromise comfort. The optimal gear and traction torque control actions that maximize the cumulative reward are computed via the Maximum a Posteriori Policy Optimization (MPO) algorithm configured for a hybrid action space. The proposed safe-RL scheme is trained and evaluated in car following scenarios where it is shown that it effectively avoids collision both during training and evaluation while delivering on the expected fuel economy improvements for the driver assistance system.

We consider infinite horizon Markov decision processes (MDPs) with fast-slow structure, meaning that certain parts of the state space move "fast" (and in a sense, are more influential) while other parts transition more "slowly." Such structure is common in real-world problems where sequential decisions need to be made at high frequencies, yet information that varies at a slower timescale also influences the optimal policy. Examples include: (1) service allocation for a multi-class queue with (slowly varying) stochastic costs, (2) a restless multi-armed bandit with an environmental state, and (3) energy demand response, where both day-ahead and real-time prices play a role in the firm's revenue. Models that fully capture these problems often result in MDPs with large state spaces and large effective time horizons (due to frequent decisions), rendering them computationally intractable. We propose an approximate dynamic programming algorithmic framework based on the idea of "freezing" the slow states, solving a set of simpler finite-horizon MDPs (the lower-level MDPs), and applying value iteration (VI) to an auxiliary MDP that transitions on a slower timescale (the upper-level MDP). We also extend the technique to a function approximation setting, where a feature-based linear architecture is used. On the theoretical side, we analyze the regret incurred by each variant of our frozen-state approach. Finally, we give empirical evidence that the frozen-state approach generates effective policies using just a fraction of the computational cost, while illustrating that simply omitting slow states from the decision modeling is often not a viable heuristic.

We study the problem of preserving privacy while still providing high utility in sequential decision making scenarios in a changing environment. We consider abruptly changing environment: the environment remains constant during periods and it changes at unknown time instants. To formulate this problem, we propose a variant of multi-armed bandits called non-stationary stochastic corrupt bandits. We construct an algorithm called SW-KLUCB-CF and prove an upper bound on its utility using the performance measure of regret. The proven regret upper bound for SW-KLUCB-CF is near-optimal in the number of time steps and matches the best known bound for analogous problems in terms of the number of time steps and the number of changes. Moreover, we present a provably optimal mechanism which can guarantee the desired level of local differential privacy while providing high utility.