This paper introduces the Shepherd Test, a new conceptual test for assessing the moral and relational dimensions of superintelligent artificial agents. The test is inspired by human interactions with animals, where ethical considerations about care, manipulation, and consumption arise in contexts of asymmetric power and self-preservation. We argue that AI crosses an important, and potentially dangerous, threshold of intelligence when it exhibits the ability to manipulate, nurture, and instrumentally use less intelligent agents, while also managing its own survival and expansion goals. This includes the ability to weigh moral trade-offs between self-interest and the well-being of subordinate agents. The Shepherd Test thus challenges traditional AI evaluation paradigms by emphasizing moral agency, hierarchical behavior, and complex decision-making under existential stakes. We argue that this shift is critical for advancing AI governance, particularly as AI systems become increasingly integrated into multi-agent environments. We conclude by identifying key research directions, including the development of simulation environments for testing moral behavior in AI, and the formalization of ethical manipulation within multi-agent systems.

Recent advances in Large Language Models (LLMs) have yielded impressive successes on many language tasks. However, efficient processing of long contexts using LLMs remains a significant challenge. We introduce EpMAN – a method for processing long contexts in an episodic memory module while holistically attending to semantically-relevant context chunks. Output from episodic attention is then used to reweigh the decoder’s self-attention to the stored KV cache of the context during training and generation. When an LLM decoder is trained using EpMAN, its performance on multiple challenging single-hop long-context recall and question-answering benchmarks is found to be stronger and more robust across the range from 16k to 256k tokens than baseline decoders trained with self-attention, and popular retrieval-augmented generation frameworks.

Developing value-aligned agents is a complex undertaking and an ongoing challenge in the field of AI. Indeed, designing Large Language Models (LLMs) that can balance multiple possibly conflicting moral values based on the context is a problem of paramount importance. In this paper, we propose a system that performs contextual value alignment based on contextual aggregation of possible responses. This aggregation is achieved by integrating a subset of possible LLM responses that are best suited to a user's input while taking into account features extracted about the user's moral preferences. The proposed system trained using the Moral Integrity Corpus displays better alignment to human values than state-of-the-art baselines.

Real-time reinforcement learning (RL) introduces several challenges. First, policies are constrained to a fixed number of actions per second due to hardware limitations. Second, the environment may change while the network is still computing an action, leading to observational delay. The first issue can partly be addressed with pipelining, leading to higher throughput and potentially better policies. However, the second issue remains: if each neuron operates in parallel with an execution time of $\tau$, an $N$-layer feed-forward network experiences observation delay of $\tau N$. Reducing the number of layers can decrease this delay, but at the cost of the network's expressivity. In this work, we explore the trade-off between minimizing delay and network's expressivity. We present a theoretically motivated solution that leverages temporal skip connections combined with history-augmented observations. We evaluate several architectures and show that those incorporating temporal skip connections achieve strong performance across various neuron execution times, reinforcement learning algorithms, and environments, including four Mujoco tasks and all MinAtar games. Moreover, we demonstrate parallel neuron computation can accelerate inference by 6-350\% on standard hardware. Our investigation into temporal skip connections and parallel computations paves the way for more efficient RL agents in real-time setting.

Realtime environments change even as agents perform action inference and learning, thus requiring high interaction frequencies to effectively minimize regret. However, recent advances in machine learning involve larger neural networks with longer inference times, raising questions about their applicability in realtime systems where reaction time is crucial. We present an analysis of lower bounds on regret in realtime reinforcement learning (RL) environments to show that minimizing long-term regret is generally impossible within the typical sequential interaction and learning paradigm, but often becomes possible when sufficient asynchronous compute is available. We propose novel algorithms for staggering asynchronous inference processes to ensure that actions are taken at consistent time intervals, and demonstrate that use of models with high action inference times is only constrained by the environment's effective stochasticity over the inference horizon, and not by action frequency. Our analysis shows that the number of inference processes needed scales linearly with increasing inference times while enabling use of models that are multiple orders of magnitude larger than existing approaches when learning from a realtime simulation of Game Boy games such as Pokemon and Tetris.

Megh Thakkar, Quentin Fournier, Matthew Riemer, Pin-Yu Chen, Amal Zouaq, Payel Das, Sarath Chandar. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024.

Due to the recent remarkable advances in artificial intelligence, researchers have begun to consider challenging learning problems such as learning to generalize behavior from large offline datasets or learning online in non-Markovian environments. Meanwhile, recent advances in both of these areas have increasingly relied on conditioning policies on large context lengths. A natural question is if there is a limit to the performance benefits of increasing the context length if the computation needed is available. In this work, we establish a novel theoretical result that links the context length of a policy to the time needed to reliably evaluate its performance (i.e., its mixing time) in large scale partially observable reinforcement learning environments that exhibit latent sub-task structure. This analysis underscores a key tradeoff: when we extend the context length, our policy can more effectively model non-Markovian dependencies, but this comes at the cost of potentially slower policy evaluation and as a result slower downstream learning. Moreover, our empirical results highlight the relevance of this analysis when leveraging Transformer based neural networks. This perspective will become increasingly pertinent as the field scales towards larger and more realistic environments, opening up a number of potential future directions for improving the way we design learning agents.

Hierarchical reinforcement learning has focused on discovering temporally extended actions, such as options, that can provide benefits in problems requiring extensive exploration. One promising approach that learns these options end-to-end is the option-critic (OC) framework. We examine and show in this paper that OC does not decompose a problem into simpler sub-problems, but instead increases the size of the search over policy space with each option considering the entire state space during learning. This issue can result in practical limitations of this method, including sample inefficient learning. To address this problem, we introduce Context-Specific Representation Abstraction for Deep Option Learning (CRADOL), a new framework that considers both temporal abstraction and context-specific representation abstraction to effectively reduce the size of the search over policy space. Specifically, our method learns a factored belief state representation that enables each option to learn a policy over only a subsection of the state space. We test our method against hierarchical, non-hierarchical, and modular recurrent neural network baselines, demonstrating significant sample efficiency improvements in challenging partially observable environments.

The main challenge of multiagent reinforcement learning is the difficulty of learning useful policies in the presence of other simultaneously learning agents whose changing behaviors jointly affect the environment's transition and reward dynamics. An effective approach that has recently emerged for addressing this non-stationarity is for each agent to anticipate the learning of other agents and influence the evolution of future policies towards desirable behavior for its own benefit. Unfortunately, previous approaches for achieving this suffer from myopic evaluation, considering only a finite number of policy updates. As such, these methods can only influence transient future policies rather than achieving the promise of scalable equilibrium selection approaches that influence the behavior at convergence. In this paper, we propose a principled framework for considering the limiting policies of other agents as time approaches infinity. Specifically, we develop a new optimization objective that maximizes each agent's average reward by directly accounting for the impact of its behavior on the limiting set of policies that other agents will converge to. Our paper characterizes desirable solution concepts within this problem setting and provides practical approaches for optimizing over possible outcomes. As a result of our farsighted objective, we demonstrate better long-term performance than state-of-the-art baselines across a suite of diverse multiagent benchmark domains.

The mixing time of the Markov chain induced by a policy limits performance in real-world continual learning scenarios. Yet, the effect of mixing times on learning in continual reinforcement learning (RL) remains underexplored. In this paper, we characterize problems that are of long-term interest to the development of continual RL, which we call scalable MDPs, through the lens of mixing times. In particular, we theoretically establish that scalable MDPs have mixing times that scale polynomially with the size of the problem. We go on to demonstrate that polynomial mixing times present significant difficulties for existing approaches that suffer from myopic bias and stale bootstrapped estimates. To validate the proposed theory, we study the empirical scaling behavior of mixing times with respect to the number of tasks and task switching frequency for pretrained high performing policies on seven Atari games. Our analysis demonstrates both that polynomial mixing times do emerge in practice and how their existence may lead to unstable learning behavior like catastrophic forgetting in continual learning settings.

In this article, we aim to provide a literature review of different formulations and approaches to continual reinforcement learning (RL), also known as lifelong or non-stationary RL. We begin by discussing our perspective on why RL is a natural fit for studying continual learning. We then provide a taxonomy of different continual RL formulations by mathematically characterizing two key properties of non-stationarity, namely, the scope and driver non-stationarity. This offers a unified view of various formulations. Next, we review and present a taxonomy of continual RL approaches. We go on to discuss evaluation of continual RL agents, providing an overview of benchmarks used in the literature and important metrics for understanding agent performance. Finally, we highlight open problems and challenges in bridging the gap between the current state of continual RL and findings in neuroscience. While still in its early days, the study of continual RL has the promise to develop better incremental reinforcement learners that can function in increasingly realistic applications where non-stationarity plays a vital role. These include applications such as those in the fields of healthcare, education, logistics, and robotics.

Training agents to learn efficiently in multi-agent environments can benefit from the explicit modelling of other agent's beliefs, especially in complex limited-information games such as the Hanabi card game. However, generalization is also highly relevant to performance in these games, though model comparisons at large training timescales can be difficult. In this work, we address this by introducing a novel model trained using a sleep metaphor on a reduced complexity version of the Hanabi game. This sleep metaphor consists an altered training regiment, as well as an information-theoretic constraint on the agent's policy. Results from experimentation demonstrate improved performance through this sleep-metaphor method, and provide a promising motivation for using similar techniques in more complex methods that incorporate explicit models of other agent's beliefs.

This paper explores information-theoretic constraints on methods for multi-agent reinforcement learning (MARL) in mixed cooperative and competitive games. Within this domain, decentralized training has been employed to increase learning sample efficiency. However, these approaches do not explicitly discourage complex policies, which can lead to overfitting. To address this, we apply an information theoretic constraint onto agents' policies that discourages overly complex behaviour when it is not associated with a significant increase in reward. A second challenge in MARL is the non-stationarity of the environment introduced by other agents' changing policies. Previous methods in MARL have sought to reduce the impact of non-stationarity by inferring other agents' policies, but this can lead to over-fitting to previously observed behaviour. To avoid this, a similar information-theoretic constraint is applied onto the inference of other agents' policies, resulting in a more robust estimate. We evaluate the effects of these information-theoretic constraints on a test suite of multi-agent games, and report an overall improvement in performance, with greater improvements found in competitive domains compared to cooperative games.

A fundamental challenge in multiagent reinforcement learning is to learn beneficial behaviors in a shared environment with other simultaneously learning agents. In particular, each agent perceives the environment as effectively non-stationary due to the changing policies of other agents. Moreover, each agent is itself constantly learning, leading to natural non-stationarity in the distribution of experiences encountered. In this paper, we propose a novel meta-multiagent policy gradient theorem that directly accounts for the non-stationary policy dynamics inherent to multiagent learning settings. This is achieved by modeling our gradient updates to consider both an agent’s own non-stationary policy dynamics and the non-stationary policy dynamics of other agents in the environment. We show that our theoretically grounded approach provides a general solution to the multiagent learning problem, which inherently comprises all key aspects of previous state of the art approaches on this topic. We test our method on a diverse suite of multiagent benchmarks and demonstrate a more efficient ability to adapt to new agents as they learn than baseline methods across the full spectrum of mixed incentive, competitive, and cooperative domains.

The difficulty of deterministic planning increases exponentially with search-tree depth. Black-box planning presents an even greater challenge, since planners must operate without an explicit model of the domain. Heuristics can make search more efficient, but goal-aware heuristics for black-box planning usually rely on goal counting, which is often quite uninformative. In this work, we show how to overcome this limitation by discovering macro-actions that make the goal-count heuristic more accurate. Our approach searches for macro-actions with focused effects (i.e. macros that modify only a small number of state variables), which align well with the assumptions made by the goal-count heuristic. Focused macros dramatically improve black-box planning efficiency across a wide range of planning domains, sometimes beating even state-of-the-art planners with access to a full domain model.

Option-critic learning is a general-purpose reinforcement learning (RL) framework that aims to address the issue of long term credit assignment by leveraging temporal abstractions. However, when dealing with extended timescales, discounting future rewards can lead to incorrect credit assignments. In this work, we address this issue by extending the hierarchical option-critic policy gradient theorem for the average reward criterion. Our proposed framework aims to maximize the long-term reward obtained in the steady-state of the Markov chain defined by the agent's policy. Furthermore, we use an ordinary differential equation based approach for our convergence analysis and prove that the parameters of the intra-option policies, termination functions, and value functions, converge to their corresponding optimal values, with probability one. Finally, we illustrate the competitive advantage of learning options, in the average reward setting, on a grid-world environment with sparse rewards.

The options framework is a popular approach for building temporally extended actions in reinforcement learning. In particular, the option-critic architecture provides general purpose policy gradient theorems for learning actions from scratch that are extended in time. However, past work makes the key assumption that each of the components of option-critic has independent parameters. In this work we note that while this key assumption of the policy gradient theorems of option-critic holds in the tabular case, it is always violated in practice for the deep function approximation setting. We thus reconsider this assumption and consider more general extensions of option-critic and hierarchical option-critic training that optimize for the full architecture with each update. It turns out that not assuming parameter independence challenges a belief in prior work that training the policy over options can be disentangled from the dynamics of the underlying options. In fact, learning can be sped up by focusing the policy over options on states where options are actually likely to terminate. We put our new algorithms to the test in application to sample efficient learning of Atari games, and demonstrate significantly improved stability and faster convergence when learning long options. 1

Collective human knowledge has clearly benefited from the fact that innovations by individuals are taught to others through communication. Similar to human social groups, agents in distributed learning systems would likely benefit from communication to share knowledge and teach skills. The problem of teaching to improve agent learning has been investigated by prior works, but these approaches make assumptions that prevent application of teaching to general multiagent problems, or require domain expertise for problems they can apply to. This learning to teach problem has inherent complexities related to measuring long-term impacts of teaching that compound the standard multiagent coordination challenges. In contrast to existing works, this paper presents the first general framework and algorithm for intelligent agents to learn to teach in a multiagent environment. Our algorithm, Learning to Coordinate and Teach Reinforcement (LeCTR), addresses peer-to-peer teaching in cooperative multiagent reinforcement learning. Each agent in our approach learns both when and what to advise, then uses the received advice to improve local learning. Importantly, these roles are not fixed; these agents learn to assume the role of student and/or teacher at the appropriate moments, requesting and providing advice in order to improve teamwide performance and learning. Empirical comparisons against state-of-the-art teaching methods show that our teaching agents not only learn significantly faster, but also learn to coordinate in tasks where existing methods fail.

Given the recent success of Deep Learning applied to a variety of single tasks, it is natural to consider more human-realistic settings. Perhaps the most difficult of these settings is that of continual lifelong learning, where the model must learn online over a continuous stream of non-stationary data. A successful continual lifelong learning system must have three key capabilities: it must learn and adapt over time, it must not forget what it has learned, and it must be efficient in both training time and memory. Recent techniques have focused their efforts primarily on the first two capabilities while questions of efficiency remain largely unexplored. In this paper, we consider the problem of efficient and effective storage of experiences over very large time-frames. In particular we consider the case where typical experiences are O(n) bits and memories are limited to O(k) bits for k

Building systems that autonomously create temporal abstractions from data is a key challenge in scaling learning and planning in reinforcement learning. One popular approach for addressing this challenge is the options framework (Sutton et al., 1999). However, only recently in (Bacon et al., 2017) was a policy gradient theorem derived for online learning of general purpose options in an end to end fashion. In this work, we extend previous work on this topic that only focuses on learning a two-level hierarchy including options and primitive actions to enable learning simultaneously at multiple resolutions in time. We achieve this by considering an arbitrarily deep hierarchy of options where high level temporally extended options are composed of lower level options with finer resolutions in time. We extend results from (Bacon et al., 2017) and derive policy gradient theorems for a deep hierarchy of options. Our proposed hierarchical option-critic architecture is capable of learning internal policies, termination conditions, and hierarchical compositions over options without the need for any intrinsic rewards or subgoals. Our empirical results in both discrete and continuous environments demonstrate the efficiency of our framework.

Automatic forecasting of time series data is a challenging problem in many industries. Current forecast models adopted by businesses do not provide adequate means for including data representing external factors that may have a significant impact on the time series, such as weather, national events, local events, social media trends, promotions, etc. This paper introduces a novel neural network attention mechanism that naturally incorporates data from multiple external sources without the feature engineering needed to get other techniques to work. We demonstrate empirically that the proposed model achieves superior performance for predicting the demand of 20 commodities across 107 stores of one of America’s largest retailers when compared to other baseline models, including neural networks, linear models, certain kernel methods, Bayesian regression, and decision trees. Our method ultimately accounts for a 23.9% relative improvement as a result of the incorporation of external data sources, and provides an unprecedented level of descriptive ability for a neural network forecasting model.