Generalization is a central challenge for the deployment of reinforcement learning (RL) systems in the real world. In this paper, we show that the sequential structure of the RL problem necessitates new approaches to generalization beyond the well-studied techniques used in supervised learning. While supervised learning methods can generalize effectively without explicitly accounting for epistemic uncertainty, we show that, perhaps surprisingly, this is not the case in RL. We show that generalization to unseen test conditions from a limited number of training conditions induces implicit partial observability, effectively turning even fully-observed MDPs into POMDPs. Informed by this observation, we recast the problem of generalization in RL as solving the induced partially observed Markov decision process, which we call the epistemic POMDP. We demonstrate the failure modes of algorithms that do not appropriately handle this partial observability, and suggest a simple ensemble-based technique for approximately solving the partially observed problem. Empirically, we demonstrate that our simple algorithm derived from the epistemic POMDP achieves significant gains in generalization over current methods on the Procgen benchmark suite.

Reinforcement learning (RL) in partially observable, fully cooperative multi-agent settings (Dec-POMDPs) can in principle be used to address many real-world challenges such as controlling a swarm of rescue robots or a synchronous team of quadcopters. However, Dec-POMDPs are significantly harder to solve than single-agent problems, with the former being NEXP-complete and the latter, MDPs, being just P-complete. Hence, current RL algorithms for Dec-POMDPs suffer from poor sample complexity, thereby reducing their applicability to practical problems where environment interaction is costly. Our key insight is that using just a polynomial number of samples, one can learn a centralized model that generalizes across different policies. We can then optimize the policy within the learned model instead of the true system, reducing the number of environment interactions. We also learn a centralized exploration policy within our model that learns to collect additional data in state-action regions with high model uncertainty. Finally, we empirically evaluate the proposed model-based algorithm, MARCO, in three cooperative communication tasks, where it improves sample efficiency by up to 20x.

Robots learning from observations in the real world using inverse reinforcement learning (IRL) may encounter objects or agents in the environment, other than the expert, that cause nuisance observations during the demonstration. These confounding elements are typically removed in fully-controlled environments such as virtual simulations or lab settings. When complete removal is impossible the nuisance observations must be filtered out. However, identifying the source of observations when large amounts of observations are made is difficult. To address this, we present a hierarchical Bayesian model that incorporates both the expert's and the confounding elements' observations thereby explicitly modeling the diverse observations a robot may receive. We extend an existing IRL algorithm originally designed to work under partial occlusion of the expert to consider the diverse observations. In a simulated robotic sorting domain containing both occlusion and confounding elements, we demonstrate the model's effectiveness. In particular, our technique outperforms several other comparative methods, second only to having perfect knowledge of the subject's trajectory.

Learning data representations that are useful for various downstream tasks is a cornerstone of artificial intelligence. While existing methods are typically evaluated on downstream tasks such as classification or generative image quality, we propose to assess representations through their usefulness in downstream control tasks, such as reaching or pushing objects. By training over 10,000 reinforcement learning policies, we extensively evaluate to what extent different representation properties affect out-of-distribution (OOD) generalization. Finally, we demonstrate zero-shot transfer of these policies from simulation to the real world, without any domain randomization or fine-tuning. This paper aims to establish the first systematic characterization of the usefulness of learned representations for real-world OOD downstream tasks.

Power grid operation subject to an extreme event requires decision-making by human operators under stressful condition with high cognitive load. Decision support under adverse dynamic events, specially if forecasted, can be supplemented by intelligent proactive control. Power system operation during wildfires require resiliency-driven proactive control for load shedding, line switching and resource allocation considering the dynamics of the wildfire and failure propagation. However, possible number of line- and load-switching in a large system during an event make traditional prediction-driven and stochastic approaches computationally intractable, leading operators to often use greedy algorithms. We model and solve the proactive control problem as a Markov decision process and introduce an integrated testbed for spatio-temporal wildfire propagation and proactive power-system operation. We transform the enormous wildfire-propagation observation space and utilize it as part of a heuristic for proactive de-energization of transmission assets. We integrate this heuristic with a reinforcement-learning based proactive policy for controlling the generating assets. Our approach allows this controller to provide setpoints for a part of the generation fleet, while a myopic operator can determine the setpoints for the remaining set, which results in a symbiotic action. We evaluate our approach utilizing the IEEE 24-node system mapped on a hypothetical terrain. Our results show that the proposed approach can help the operator to reduce load loss during an extreme event, reduce power flow through lines that are to be de-energized, and reduce the likelihood of infeasible power-flow solutions, which would indicate violation of short-term thermal limits of transmission lines.

Energy-based models (EBMs) are explicit generative models which work by considering Gibbs measures defined through an energy function f, with a probability density proportional to exp( βf(x)), where β is the inverse temperature. Such models originate in statistical physics [Gibbs, 2010, Ruelle, 1969], and have become a fundamental modeling tool in statistics and machine learning [Wainwright and Jordan, 2008, Ranzato et al., 2007, LeCun et al., 2006, Du and Mordatch, 2019, Song and Kingma, 2021]. Given data samples from a target distribution, the learning algorithms for EBMs attempt to estimate an energy function f to model the samples density. The resulting learned model can then be used to obtain new samples, typically through Markov Chain Monte Carlo (MCMC) techniques. The standard method to train EBMs is maximum likelihood estimation, i.e. the learned energy is the one maximizing the likelihood of the target samples, within a certain function class. One generic approach for this is to use gradient descent, where gradients may be approximated using MCMC samples from the trained model. However, this is computationally difficult for highly non-convex trained energies, which in recent years has motivated a myriad of alternative losses to learn EBM energies, such as the popular score matching; see [Song and Kingma, 2021] for a review. EBMs also have structural connections with maximum entropy (maxent) models, which have been studied for decades through Fenchel duality. Dai et al. [2019b] was the first work to leverage similar duality arguments

This paper studies the traffic monitoring problem in a road network using a team of aerial robots. The problem is challenging due to two main reasons. First, the traffic events are stochastic, both temporally and spatially. Second, the problem has a non-homogeneous structure as the traffic events arrive at different locations of the road network at different rates. Accordingly, some locations require more visits by the robots compared to other locations. To address these issues, we define an uncertainty metric for each location of the road network and formulate a path planning problem for the aerial robots to minimize the network's average uncertainty. We express this problem as a partially observable Markov decision process (POMDP) and propose a distributed and scalable algorithm based on deep reinforcement learning to solve it. We consider two different scenarios depending on the communication mode between the agents (aerial robots) and the traffic management center (TMC). The first scenario assumes that the agents continuously communicate with the TMC to send/receive real-time information about the traffic events. Hence, the agents have global and real-time knowledge of the environment. However, in the second scenario, we consider a challenging setting where the observation of the aerial robots is partial and limited to their sensing ranges. Moreover, in contrast to the first scenario, the information exchange between the aerial robots and the TMC is restricted to specific time instances. We evaluate the performance of our proposed algorithm in both scenarios for a real road network topology and demonstrate its functionality in a traffic monitoring system.

We study countably infinite Markov decision processes (MDPs) with real-valued transition rewards. Every infinite run induces the following sequences of payoffs: 1. Point payoff (the sequence of directly seen transition rewards), 2. Total payoff (the sequence of the sums of all rewards so far), and 3. Mean payoff. For each payoff type, the objective is to maximize the probability that the $\liminf$ is non-negative. We establish the complete picture of the strategy complexity of these objectives, i.e., how much memory is necessary and sufficient for $\varepsilon$-optimal (resp. optimal) strategies. Some cases can be won with memoryless deterministic strategies, while others require a step counter, a reward counter, or both.

We show that many machine-learning algorithms are specific instances of a single algorithm called the Bayesian learning rule. The rule, derived from Bayesian principles, yields a wide-range of algorithms from fields such as optimization, deep learning, and graphical models. This includes classical algorithms such as ridge regression, Newton's method, and Kalman filter, as well as modern deep-learning algorithms such as stochastic-gradient descent, RMSprop, and Dropout. The key idea in deriving such algorithms is to approximate the posterior using candidate distributions estimated by using natural gradients. Different candidate distributions result in different algorithms and further approximations to natural gradients give rise to variants of those algorithms. Our work not only unifies, generalizes, and improves existing algorithms, but also helps us design new ones.

Autonomous navigation in dense traffic scenarios remains challenging for autonomous vehicles (AVs) because the intentions of other drivers are not directly observable and AVs have to deal with a wide range of driving behaviors. To maneuver through dense traffic, AVs must be able to reason how their actions affect others (interaction model) and exploit this reasoning to navigate through dense traffic safely. This paper presents a novel framework for interaction-aware motion planning in dense traffic scenarios. We explore the connection between human driving behavior and their velocity changes when interacting. Hence, we propose to learn, via deep Reinforcement Learning (RL), an interaction-aware policy providing global guidance about the cooperativeness of other vehicles to an optimization-based planner ensuring safety and kinematic feasibility through constraint satisfaction. The learned policy can reason and guide the local optimization-based planner with interactive behavior to pro-actively merge in dense traffic while remaining safe in case the other vehicles do not yield. We present qualitative and quantitative results in highly interactive simulation environments (highway merging and unprotected left turns) against two baseline approaches, a learning-based and an optimization-based method. The presented results demonstrate that our method significantly reduces the number of collisions and increases the success rate with respect to both learning-based and optimization-based baselines.