Jacob Beck

Jacob Beck, Risto Vuorio, Zheng Xiong et al.

Deep reinforcement learning (RL) is notoriously impractical to deploy due to sample inefficiency. Meta-RL directly addresses this sample inefficiency by learning to perform few-shot learning when a distribution of related tasks is available for meta-training. While many specialized meta-RL methods have been proposed, recent work suggests that end-to-end learning in conjunction with an off-the-shelf sequential model, such as a recurrent network, is a surprisingly strong baseline. However, such claims have been controversial due to limited supporting evidence, particularly in the face of prior work establishing precisely the opposite. In this paper, we conduct an empirical investigation. While we likewise find that a recurrent network can achieve strong performance, we demonstrate that the use of hypernetworks is crucial to maximizing their potential. Surprisingly, when combined with hypernetworks, the recurrent baselines that are far simpler than existing specialized methods actually achieve the strongest performance of all methods evaluated. We provide code at https://github.com/jacooba/hyper.

Jacob Beck, Risto Vuorio, Evan Zheran Liu et al.

While deep reinforcement learning (RL) has fueled multiple high-profile successes in machine learning, it is held back from more widespread adoption by its often poor data efficiency and the limited generality of the policies it produces. A promising approach for alleviating these limitations is to cast the development of better RL algorithms as a machine learning problem itself in a process called meta-RL. Meta-RL is most commonly studied in a problem setting where, given a distribution of tasks, the goal is to learn a policy that is capable of adapting to any new task from the task distribution with as little data as possible. In this survey, we describe the meta-RL problem setting in detail as well as its major variations. We discuss how, at a high level, meta-RL research can be clustered based on the presence of a task distribution and the learning budget available for each individual task. Using these clusters, we then survey meta-RL algorithms and applications. We conclude by presenting the open problems on the path to making meta-RL part of the standard toolbox for a deep RL practitioner.

Jacob Beck, Matthew Thomas Jackson, Risto Vuorio et al.

Training a reinforcement learning (RL) agent on a real-world robotics task remains generally impractical due to sample inefficiency. Multi-task RL and meta-RL aim to improve sample efficiency by generalizing over a distribution of related tasks. However, doing so is difficult in practice: In multi-task RL, state of the art methods often fail to outperform a degenerate solution that simply learns each task separately. Hypernetworks are a promising path forward since they replicate the separate policies of the degenerate solution while also allowing for generalization across tasks, and are applicable to meta-RL. However, evidence from supervised learning suggests hypernetwork performance is highly sensitive to the initialization. In this paper, we 1) show that hypernetwork initialization is also a critical factor in meta-RL, and that naive initializations yield poor performance; 2) propose a novel hypernetwork initialization scheme that matches or exceeds the performance of a state-of-the-art approach proposed for supervised settings, as well as being simpler and more general; and 3) use this method to show that hypernetworks can improve performance in meta-RL by evaluating on multiple simulated robotics benchmarks.

Zheng Xiong, Jacob Beck, Shimon Whiteson

Learning a universal policy across different robot morphologies can significantly improve learning efficiency and generalization in continuous control. However, it poses a challenging multi-task reinforcement learning problem, as the optimal policy may be quite different across robots and critically depend on the morphology. Existing methods utilize graph neural networks or transformers to handle heterogeneous state and action spaces across different morphologies, but pay little attention to the dependency of a robot's control policy on its morphology context. In this paper, we propose a hierarchical architecture to better model this dependency via contextual modulation, which includes two key submodules: (1) Instead of enforcing hard parameter sharing across robots, we use hypernetworks to generate morphology-dependent control parameters; (2) We propose a fixed attention mechanism that solely depends on the morphology to modulate the interactions between different limbs in a robot. Experimental results show that our method not only improves learning performance on a diverse set of training robots, but also generalizes better to unseen morphologies in a zero-shot fashion.

Risto Vuorio, Jacob Beck, Shimon Whiteson et al.

Meta-gradients provide a general approach for optimizing the meta-parameters of reinforcement learning (RL) algorithms. Estimation of meta-gradients is central to the performance of these meta-algorithms, and has been studied in the setting of MAML-style short-horizon meta-RL problems. In this context, prior work has investigated the estimation of the Hessian of the RL objective, as well as tackling the problem of credit assignment to pre-adaptation behavior by making a sampling correction. However, we show that Hessian estimation, implemented for example by DiCE and its variants, always adds bias and can also add variance to meta-gradient estimation. Meanwhile, meta-gradient estimation has been studied less in the important long-horizon setting, where backpropagation through the full inner optimization trajectories is not feasible. We study the bias and variance tradeoff arising from truncated backpropagation and sampling correction, and additionally compare to evolution strategies, which is a recently popular alternative strategy to long-horizon meta-learning. While prior work implicitly chooses points in this bias-variance space, we disentangle the sources of bias and variance and present an empirical study that relates existing estimators to each other.

Christoph Kern, Stephanie Eckman, Jacob Beck et al.

When training data are collected from human annotators, the design of the annotation instrument, the instructions given to annotators, the characteristics of the annotators, and their interactions can impact training data. This study demonstrates that design choices made when creating an annotation instrument also impact the models trained on the resulting annotations. We introduce the term annotation sensitivity to refer to the impact of annotation data collection methods on the annotations themselves and on downstream model performance and predictions. We collect annotations of hate speech and offensive language in five experimental conditions of an annotation instrument, randomly assigning annotators to conditions. We then fine-tune BERT models on each of the five resulting datasets and evaluate model performance on a holdout portion of each condition. We find considerable differences between the conditions for 1) the share of hate speech/offensive language annotations, 2) model performance, 3) model predictions, and 4) model learning curves. Our results emphasize the crucial role played by the annotation instrument which has received little attention in the machine learning literature. We call for additional research into how and why the instrument impacts the annotations to inform the development of best practices in instrument design.

Jacob Beck, Matthew Jackson, Risto Vuorio et al.

A core ambition of reinforcement learning (RL) is the creation of agents capable of rapid learning in novel tasks. Meta-RL aims to achieve this by directly learning such agents. Black box methods do so by training off-the-shelf sequence models end-to-end. By contrast, task inference methods explicitly infer a posterior distribution over the unknown task, typically using distinct objectives and sequence models designed to enable task inference. Recent work has shown that task inference methods are not necessary for strong performance. However, it remains unclear whether task inference sequence models are beneficial even when task inference objectives are not. In this paper, we present evidence that task inference sequence models are indeed still beneficial. In particular, we investigate sequence models with permutation invariant aggregation, which exploit the fact that, due to the Markov property, the task posterior does not depend on the order of data. We empirically confirm the advantage of permutation invariant sequence models without the use of task inference objectives. However, we also find, surprisingly, that there are multiple conditions under which permutation variance remains useful. Therefore, we propose SplAgger, which uses both permutation variant and invariant components to achieve the best of both worlds, outperforming all baselines evaluated on continuous control and memory environments. Code is provided at https://github.com/jacooba/hyper.

Amir Moeini, Jiuqi Wang, Jacob Beck et al.

Reinforcement learning (RL) agents typically optimize their policies by performing expensive backward passes to update their network parameters. However, some agents can solve new tasks without updating any parameters by simply conditioning on additional context such as their action-observation histories. This paper surveys work on such behavior, known as in-context reinforcement learning.

Zheng Xiong, Risto Vuorio, Jacob Beck et al.

Learning a universal policy across different robot morphologies can significantly improve learning efficiency and enable zero-shot generalization to unseen morphologies. However, learning a highly performant universal policy requires sophisticated architectures like transformers (TF) that have larger memory and computational cost than simpler multi-layer perceptrons (MLP). To achieve both good performance like TF and high efficiency like MLP at inference time, we propose HyperDistill, which consists of: (1) A morphology-conditioned hypernetwork (HN) that generates robot-wise MLP policies, and (2) A policy distillation approach that is essential for successful training. We show that on UNIMAL, a benchmark with hundreds of diverse morphologies, HyperDistill performs as well as a universal TF teacher policy on both training and unseen test robots, but reduces model size by 6-14 times, and computational cost by 67-160 times in different environments. Our analysis attributes the efficiency advantage of HyperDistill at inference time to knowledge decoupling, i.e., the ability to decouple inter-task and intra-task knowledge, a general principle that could also be applied to improve inference efficiency in other domains.

Jacob Beck

While internet-scale image and textual data have enabled strong generalization in Vision-Language Models (VLMs), the absence of internet-scale control data has impeded the development of similar generalization in standard reinforcement learning (RL) agents. Although VLMs are fundamentally limited in their ability to solve control tasks due to their lack of action-conditioned training data, their capacity for image understanding allows them to provide valuable feedback in RL tasks by recognizing successful outcomes. A key challenge in Reinforcement Learning from AI Feedback (RLAIF) is determining how best to integrate VLM-derived signals into the learning process. We explore this question in the context of offline RL and introduce a class of methods called Sub-Trajectory Filtered Optimization (SFO). We identify three key insights. First, trajectory length plays a crucial role in offline RL, as full-trajectory preference learning exacerbates the stitching problem, necessitating the use of sub-trajectories. Second, even in Markovian environments, a non-Markovian reward signal from a sequence of images is required to assess trajectory improvement, as VLMs do not interpret control actions and must rely on visual cues over time. Third, a simple yet effective approach--filtered and weighted behavior cloning--consistently outperforms more complex RLHF-based methods. We propose Sub-Trajectory Filtered Behavior Cloning (SFBC), a method that leverages VLM feedback on sub-trajectories while incorporating a retrospective filtering mechanism that removes sub-trajectories preceding failures to improve robustness and prevent turbulence. Please enjoy our airport puns.

Mingfei Sun, Sam Devlin, Jacob Beck et al.

We present trust region bounds for optimizing decentralized policies in cooperative Multi-Agent Reinforcement Learning (MARL), which holds even when the transition dynamics are non-stationary. This new analysis provides a theoretical understanding of the strong performance of two recent actor-critic methods for MARL, which both rely on independent ratios, i.e., computing probability ratios separately for each agent's policy. We show that, despite the non-stationarity that independent ratios cause, a monotonic improvement guarantee still arises as a result of enforcing the trust region constraint over all decentralized policies. We also show this trust region constraint can be effectively enforced in a principled way by bounding independent ratios based on the number of agents in training, providing a theoretical foundation for proximal ratio clipping. Finally, our empirical results support the hypothesis that the strong performance of IPPO and MAPPO is a direct result of enforcing such a trust region constraint via clipping in centralized training, and tuning the hyperparameters with regards to the number of agents, as predicted by our theoretical analysis.

Zheng Xiong, Luisa Zintgraf, Jacob Beck et al.

Consistency is the theoretical property of a meta learning algorithm that ensures that, under certain assumptions, it can adapt to any task at test time. An open question is whether and how theoretical consistency translates into practice, in comparison to inconsistent algorithms. In this paper, we empirically investigate this question on a set of representative meta-RL algorithms. We find that theoretically consistent algorithms can indeed usually adapt to out-of-distribution (OOD) tasks, while inconsistent ones cannot, although they can still fail in practice for reasons like poor exploration. We further find that theoretically inconsistent algorithms can be made consistent by continuing to update all agent components on the OOD tasks, and adapt as well or better than originally consistent ones. We conclude that theoretical consistency is indeed a desirable property, and inconsistent meta-RL algorithms can easily be made consistent to enjoy the same benefits.

Matt Cooper, Jun Ki Lee, Jacob Beck et al.

Mutually beneficial behavior in repeated games can be enforced via the threat of punishment, as enshrined in game theory's well-known "folk theorem." There is a cost, however, to a player for generating these disincentives. In this work, we seek to minimize this cost by computing a "Stackelberg punishment," in which the player selects a behavior that sufficiently punishes the other player while maximizing its own score under the assumption that the other player will adopt a best response. This idea generalizes the concept of a Stackelberg equilibrium. Known efficient algorithms for computing a Stackelberg equilibrium can be adapted to efficiently produce a Stackelberg punishment. We demonstrate an application of this idea in an experiment involving a virtual autonomous vehicle and human participants. We find that a self-driving car with a Stackelberg punishment policy discourages human drivers from bullying in a driving scenario requiring social negotiation.

Jacob Beck, Zoe Papakipos, Michael Littman

In autonomous vehicle (AV) control, allowing mistakes can be quite dangerous and costly in the real world. For this reason we investigate methods of training an AV without allowing the agent to explore and instead having a human explorer collect the data. Supervised learning has been explored for AV control, but it encounters the issue of the covariate shift. That is, training data collected from an optimal demonstration consists only of the states induced by the optimal control policy, but at runtime, the trained agent may encounter a vastly different state distribution with little relevant training data. To mitigate this issue, we have our human explorer make sub-optimal decisions. In order to have our agent not replicate these sub-optimal decisions, supervised learning requires that we either erase these actions, or replace these action with the correct action. Erasing is wasteful and replacing is difficult, since it is not easy to know the correct action without driving. We propose an alternate framework that includes continuous scalar feedback for each action, marking which actions we should replicate, which we should avoid, and how sure we are. Our framework learns continuous control from sub-optimal demonstration and evaluative feedback collected before training. We find that a human demonstrator can explore sub-optimal states in a safe manner, while still getting enough gradation to benefit learning. The collection method for data and feedback we call "Backseat Driver." We call the more general learning framework ReNeg, since it learns a regression from states to actions given negative as well as positive examples. We empirically validate several models in the ReNeg framework, testing on lane-following with limited data. We find that the best solution is a generalization of mean-squared error and outperforms supervised learning on the positive examples alone.

Jacob Beck, Zoe Papakipos

Neural networks are based on a simplified model of the brain. In this project, we wanted to relax the simplifying assumptions of a traditional neural network by making a model that more closely emulates the low level interactions of neurons. Like in an RNN, our model has a state that persists between time steps, so that the energies of neurons persist. However, unlike an RNN, our state consists of a 2 dimensional matrix, rather than a 1 dimensional vector, thereby introducing a concept of distance to other neurons within the state. In our model, neurons can only fire to adjacent neurons, as in the brain. Like in the brain, we only allow neurons to fire in a time step if they contain enough energy, or excitement. We also enforce a notion of conservation of energy, so that a neuron cannot excite its neighbors more than the excitement it already contained at that time step. Taken together, these two features allow signals in the form of activations to flow around in our network over time, making our neural mesh more closely model signals traveling through the brain the brain. Although our main goal is to design an architecture to more closely emulate the brain in the hope of having a correct internal representation of information by the time we know how to properly train a general intelligence, we did benchmark our neural mash on a specific task. We found that by increasing the runtime of the mesh, we were able to increase its accuracy without increasing the number of parameters.