Sarah Keren

Matthias Gerstgrasser, Tom Danino, Sarah Keren

We present a novel multi-agent RL approach, Selective Multi-Agent Prioritized Experience Relay, in which agents share with other agents a limited number of transitions they observe during training. The intuition behind this is that even a small number of relevant experiences from other agents could help each agent learn. Unlike many other multi-agent RL algorithms, this approach allows for largely decentralized training, requiring only a limited communication channel between agents. We show that our approach outperforms baseline no-sharing decentralized training and state-of-the art multi-agent RL algorithms. Further, sharing only a small number of highly relevant experiences outperforms sharing all experiences between agents, and the performance uplift from selective experience sharing is robust across a range of hyperparameters and DQN variants. A reference implementation of our algorithm is available at https://github.com/mgerstgrasser/super.

Mira Finkelstein, Lucy Liu, Nitsan Levy Schlot et al.

Understanding emerging behaviors of reinforcement learning (RL) agents may be difficult since such agents are often trained in complex environments using highly complex decision making procedures. This has given rise to a variety of approaches to explainability in RL that aim to reconcile discrepancies that may arise between the behavior of an agent and the behavior that is anticipated by an observer. Most recent approaches have relied either on domain knowledge that may not always be available, on an analysis of the agent's policy, or on an analysis of specific elements of the underlying environment, typically modeled as a Markov Decision Process (MDP). Our key claim is that even if the underlying model is not fully known (e.g., the transition probabilities have not been accurately learned) or is not maintained by the agent (i.e., when using model-free methods), the model can nevertheless be exploited to automatically generate explanations. For this purpose, we suggest using formal MDP abstractions and transforms, previously used in the literature for expediting the search for optimal policies, to automatically produce explanations. Since such transforms are typically based on a symbolic representation of the environment, they can provide meaningful explanations for gaps between the anticipated and actual agent behavior. We formally define the explainability problem, suggest a class of transforms that can be used for explaining emergent behaviors, and suggest methods that enable efficient search for an explanation. We demonstrate the approach on a set of standard benchmarks.

Guy Azran, Mohamad H. Danesh, Stefano V. Albrecht et al.

Recent studies show that deep reinforcement learning (DRL) agents tend to overfit to the task on which they were trained and fail to adapt to minor environment changes. To expedite learning when transferring to unseen tasks, we propose a novel approach to representing the current task using reward machines (RMs), state machine abstractions that induce subtasks based on the current task's rewards and dynamics. Our method provides agents with symbolic representations of optimal transitions from their current abstract state and rewards them for achieving these transitions. These representations are shared across tasks, allowing agents to exploit knowledge of previously encountered symbols and transitions, thus enhancing transfer. Empirical results show that our representations improve sample efficiency and few-shot transfer in a variety of domains.

Adi Amuzig, David Dovrat, Sarah Keren

Mobile robotic agents often suffer from localization uncertainty which grows with time and with the agents' movement. This can hinder their ability to accomplish their task. In some settings, it may be possible to perform assistive actions that reduce uncertainty about a robot's location. For example, in a collaborative multi-robot system, a wheeled robot can request assistance from a drone that can fly to its estimated location and reveal its exact location on the map or accompany it to its intended location. Since assistance may be costly and limited, and may be requested by different members of a team, there is a need for principled ways to support the decision of which assistance to provide to an agent and when, as well as to decide which agent to help within a team. For this purpose, we propose Value of Assistance (VOA) to represent the expected cost reduction that assistance will yield at a given point of execution. We offer ways to compute VOA based on estimations of the robot's future uncertainty, modeled as a Gaussian process. We specify conditions under which our VOA measures are valid and empirically demonstrate the ability of our measures to predict the agent's average cost reduction when receiving assistance in both simulated and real-world robotic settings.

Eliezer Shlomi, Ido Levy, Eilam Shapira et al.

Planning with LLMs is bottlenecked by token-by-token generation and repeated full forward passes, making multi-step lookahead and rollout-based search expensive in latency and compute. We propose EmbedPlan, which replaces autoregressive next-state generation with a lightweight transition model operating in a frozen language embedding space. EmbedPlan encodes natural language state and action descriptions into vectors, predicts the next-state embedding, and retrieves the next state by nearest-neighbor similarity, enabling fast planning computation without fine-tuning the encoder. We evaluate next-state prediction across nine classical planning domains using six evaluation protocols of increasing difficulty: interpolation, plan-variant, extrapolation, multi-domain, cross-domain, and leave-one-out. Results show near-perfect interpolation performance but a sharp degradation when generalization requires transfer to unseen problems or unseen domains; plan-variant evaluation indicates generalization to alternative plans rather than memorizing seen trajectories. Overall, frozen embeddings support within-domain dynamics learning after observing a domain's transitions, while transfer across domain boundaries remains a bottleneck.

Mohammad Masarwy, Yuval Goshen, David Dovrat et al.

In multiple realistic settings, a robot is tasked with grasping an object without knowing its exact pose and relies on a probabilistic estimation of the pose to decide how to attempt the grasp. We support settings in which it is possible to provide the robot with an observation of the object before a grasp is attempted but this possibility is limited and there is a need to decide which sensing action would be most beneficial. We support this decision by offering a novel Value of Assistance (VOA) measure for assessing the expected effect a specific observation will have on the robot's ability to complete its task. We evaluate our suggested measure in simulated and real-world collaborative grasping settings.

Kelsey Sikes, Sarah Keren, Sarath Sreedharan

One of the most difficult challenges in creating successful human-AI collaborations is aligning a robot's behavior with a human user's expectations. When this fails to occur, a robot may misinterpret their specified goals, prompting it to perform actions with unanticipated, potentially dangerous side effects. To avoid this, we propose a new metric we call Goal State Divergence $\mathcal{(GSD)}$, which represents the difference between a robot's final goal state and the one a human user expected. In cases where $\mathcal{GSD}$ cannot be directly calculated, we show how it can be approximated using maximal and minimal bounds. We then input the $\mathcal{GSD}$ value into our novel human-robot goal alignment (HRGA) design problem, which identifies a minimal set of environment modifications that can prevent mismatches like this. To show the effectiveness of $\mathcal{GSD}$ for reducing differences between human-robot goal states, we empirically evaluate our approach on several standard benchmarks.

Robert Kasumba, Guanghui Yu, Chien-Ju Ho et al.

Goal recognition design (GRD) aims to make limited modifications to decision-making environments to make it easier to infer the goals of agents acting within those environments. Although various research efforts have been made in goal recognition design, existing approaches are computationally demanding and often assume that agents are (near-)optimal in their decision-making. To address these limitations, we leverage machine learning methods for goal recognition design that can both improve run-time efficiency and account for agents with general behavioral models. Following existing literature, we use worst-case distinctiveness (wcd) as a measure of the difficulty in inferring the goal of an agent in a decision-making environment. Our approach begins by training a machine learning model to predict the wcd for a given environment and the agent behavior model. We then propose a gradient-based optimization framework that accommodates various constraints to optimize decision-making environments for enhanced goal recognition. Through extensive simulations, we demonstrate that our approach outperforms existing methods in reducing wcd and enhances runtime efficiency. Moreover, our approach also adapts to settings in which existing approaches do not apply, such as those involving flexible budget constraints, more complex environments, and suboptimal agent behavior. Finally, we conducted human-subject experiments that demonstrate that our method creates environments that facilitate efficient goal recognition from human decision-makers.

Sarah Keren, Chaimaa Essayeh, Stefano V. Albrecht et al.

The rapidly changing architecture and functionality of electrical networks and the increasing penetration of renewable and distributed energy resources have resulted in various technological and managerial challenges. These have rendered traditional centralized energy-market paradigms insufficient due to their inability to support the dynamic and evolving nature of the network. This survey explores how multi-agent reinforcement learning (MARL) can support the decentralization and decarbonization of energy networks and mitigate the associated challenges. This is achieved by specifying key computational challenges in managing energy networks, reviewing recent research progress on addressing them, and highlighting open challenges that may be addressed using MARL.

Elinor Ginzburg, Itay Segev, Yoash Levron et al.

We aim to better understand the tradeoffs between traditional and reinforcement learning (RL) approaches for energy storage management. More specifically, we wish to better understand the performance loss incurred when using a generative RL policy instead of using a traditional approach to find optimal control policies for specific instances. Our comparison is based on a simplified micro-grid model, that includes a load component, a photovoltaic source, and a storage device. Based on this model, we examine three use cases of increasing complexity: ideal storage with convex cost functions, lossy storage devices, and lossy storage devices with convex transmission losses. With the aim of promoting the principled use RL based methods in this challenging and important domain, we provide a detailed formulation of each use case and a detailed description of the optimization challenges. We then compare the performance of traditional and RL methods, discuss settings in which it is beneficial to use each method, and suggest avenues for future investigation.

Ilai Shraga, Guy Azran, Matthias Gerstgrasser et al.

To effectively operate in various dynamic scenarios, RL agents must be resilient to unexpected changes in their environment. Previous work on this form of resilience has focused on single-agent settings. In this work, we introduce and formalize a multi-agent variant of resilience, which we term group resilience. We further hypothesize that collaboration with other agents is key to achieving group resilience; collaborating agents adapt better to environmental perturbations in multi-agent reinforcement learning (MARL) settings. We test our hypothesis empirically by evaluating different collaboration protocols and examining their effect on group resilience. Our experiments show that all the examined collaborative approaches achieve higher group resilience than their non-collaborative counterparts.

Anagha Kulkarni, Sarath Sreedharan, Sarah Keren et al.

Designing robots capable of generating interpretable behavior is a prerequisite for achieving effective human-robot collaboration. This means that the robots need to be capable of generating behavior that aligns with human expectations and, when required, provide explanations to the humans in the loop. However, exhibiting such behavior in arbitrary environments could be quite expensive for robots, and in some cases, the robot may not even be able to exhibit the expected behavior. Given structured environments (like warehouses and restaurants), it may be possible to design the environment so as to boost the interpretability of the robot's behavior or to shape the human's expectations of the robot's behavior. In this paper, we investigate the opportunities and limitations of environment design as a tool to promote a type of interpretable behavior -- known in the literature as explicable behavior. We formulate a novel environment design framework that considers design over multiple tasks and over a time horizon. In addition, we explore the longitudinal aspect of explicable behavior and the trade-off that arises between the cost of design and the cost of generating explicable behavior over a time horizon.