Martha White

Andrew Patterson, Samuel Neumann, Martha White et al. · deepmind

Empirical design in reinforcement learning is no small task. Running good experiments requires attention to detail and at times significant computational resources. While compute resources available per dollar have continued to grow rapidly, so have the scale of typical experiments in reinforcement learning. It is now common to benchmark agents with millions of parameters against dozens of tasks, each using the equivalent of 30 days of experience. The scale of these experiments often conflict with the need for proper statistical evidence, especially when comparing algorithms. Recent studies have highlighted how popular algorithms are sensitive to hyper-parameter settings and implementation details, and that common empirical practice leads to weak statistical evidence (Machado et al., 2018; Henderson et al., 2018). Here we take this one step further. This manuscript represents both a call to action, and a comprehensive resource for how to do good experiments in reinforcement learning. In particular, we cover: the statistical assumptions underlying common performance measures, how to properly characterize performance variation and stability, hypothesis testing, special considerations for comparing multiple agents, baseline and illustrative example construction, and how to deal with hyper-parameters and experimenter bias. Throughout we highlight common mistakes found in the literature and the statistical consequences of those in example experiments. The objective of this document is to provide answers on how we can use our unprecedented compute to do good science in reinforcement learning, as well as stay alert to potential pitfalls in our empirical design.

Chenjun Xiao, Han Wang, Yangchen Pan et al. · deepmind

Reinforcement learning (RL) agents can leverage batches of previously collected data to extract a reasonable control policy. An emerging issue in this offline RL setting, however, is that the bootstrapping update underlying many of our methods suffers from insufficient action-coverage: standard max operator may select a maximal action that has not been seen in the dataset. Bootstrapping from these inaccurate values can lead to overestimation and even divergence. There are a growing number of methods that attempt to approximate an \emph{in-sample} max, that only uses actions well-covered by the dataset. We highlight a simple fact: it is more straightforward to approximate an in-sample \emph{softmax} using only actions in the dataset. We show that policy iteration based on the in-sample softmax converges, and that for decreasing temperatures it approaches the in-sample max. We derive an In-Sample Actor-Critic (AC), using this in-sample softmax, and show that it is consistently better or comparable to existing offline RL methods, and is also well-suited to fine-tuning.

Han Wang, Archit Sakhadeo, Adam White et al. · deepmind

The performance of reinforcement learning (RL) agents is sensitive to the choice of hyperparameters. In real-world settings like robotics or industrial control systems, however, testing different hyperparameter configurations directly on the environment can be financially prohibitive, dangerous, or time consuming. We propose a new approach to tune hyperparameters from offline logs of data, to fully specify the hyperparameters for an RL agent that learns online in the real world. The approach is conceptually simple: we first learn a model of the environment from the offline data, which we call a calibration model, and then simulate learning in the calibration model to identify promising hyperparameters. We identify several criteria to make this strategy effective, and develop an approach that satisfies these criteria. We empirically investigate the method in a variety of settings to identify when it is effective and when it fails.

Khurram Javed, Haseeb Shah, Rich Sutton et al.

Constructing states from sequences of observations is an important component of reinforcement learning agents. One solution for state construction is to use recurrent neural networks. Back-propagation through time (BPTT), and real-time recurrent learning (RTRL) are two popular gradient-based methods for recurrent learning. BPTT requires complete trajectories of observations before it can compute the gradients and is unsuitable for online updates. RTRL can do online updates but scales poorly to large networks. In this paper, we propose two constraints that make RTRL scalable. We show that by either decomposing the network into independent modules or learning the network in stages, we can make RTRL scale linearly with the number of parameters. Unlike prior scalable gradient estimation algorithms, such as UORO and Truncated-BPTT, our algorithms do not add noise or bias to the gradient estimate. Instead, they trade off the functional capacity of the network for computationally efficient learning. We demonstrate the effectiveness of our approach over Truncated-BPTT on a prediction benchmark inspired by animal learning and by doing policy evaluation of pre-trained policies for Atari 2600 games.

Andrew Patterson, Victor Liao, Martha White

Most value function learning algorithms in reinforcement learning are based on the mean squared (projected) Bellman error. However, squared errors are known to be sensitive to outliers, both skewing the solution of the objective and resulting in high-magnitude and high-variance gradients. To control these high-magnitude updates, typical strategies in RL involve clipping gradients, clipping rewards, rescaling rewards, or clipping errors. While these strategies appear to be related to robust losses -- like the Huber loss -- they are built on semi-gradient update rules which do not minimize a known loss. In this work, we build on recent insights reformulating squared Bellman errors as a saddlepoint optimization problem and propose a saddlepoint reformulation for a Huber Bellman error and Absolute Bellman error. We start from a formalization of robust losses, then derive sound gradient-based approaches to minimize these losses in both the online off-policy prediction and control settings. We characterize the solutions of the robust losses, providing insight into the problem settings where the robust losses define notably better solutions than the mean squared Bellman error. Finally, we show that the resulting gradient-based algorithms are more stable, for both prediction and control, with less sensitivity to meta-parameters.

Brett Daley, Martha White, Christopher Amato et al.

Off-policy learning from multistep returns is crucial for sample-efficient reinforcement learning, but counteracting off-policy bias without exacerbating variance is challenging. Classically, off-policy bias is corrected in a per-decision manner: past temporal-difference errors are re-weighted by the instantaneous Importance Sampling (IS) ratio after each action via eligibility traces. Many off-policy algorithms rely on this mechanism, along with differing protocols for cutting the IS ratios to combat the variance of the IS estimator. Unfortunately, once a trace has been fully cut, the effect cannot be reversed. This has led to the development of credit-assignment strategies that account for multiple past experiences at a time. These trajectory-aware methods have not been extensively analyzed, and their theoretical justification remains uncertain. In this paper, we propose a multistep operator that can express both per-decision and trajectory-aware methods. We prove convergence conditions for our operator in the tabular setting, establishing the first guarantees for several existing methods as well as many new ones. Finally, we introduce Recency-Bounded Importance Sampling (RBIS), which leverages trajectory awareness to perform robustly across $λ$-values in an off-policy control task.

Lingwei Zhu, Zheng Chen, Matthew Schlegel et al.

Many policy optimization approaches in reinforcement learning incorporate a Kullback-Leilbler (KL) divergence to the previous policy, to prevent the policy from changing too quickly. This idea was initially proposed in a seminal paper on Conservative Policy Iteration, with approximations given by algorithms like TRPO and Munchausen Value Iteration (MVI). We continue this line of work by investigating a generalized KL divergence -- called the Tsallis KL divergence -- which use the $q$-logarithm in the definition. The approach is a strict generalization, as $q = 1$ corresponds to the standard KL divergence; $q > 1$ provides a range of new options. We characterize the types of policies learned under the Tsallis KL, and motivate when $q >1$ could be beneficial. To obtain a practical algorithm that incorporates Tsallis KL regularization, we extend MVI, which is one of the simplest approaches to incorporate KL regularization. We show that this generalized MVI($q$) obtains significant improvements over the standard MVI($q = 1$) across 35 Atari games.

Kirby Banman, Liam Peet-Pare, Nidhi Hegde et al.

Most convergence guarantees for stochastic gradient descent with momentum (SGDm) rely on iid sampling. Yet, SGDm is often used outside this regime, in settings with temporally correlated input samples such as continual learning and reinforcement learning. Existing work has shown that SGDm with a decaying step-size can converge under Markovian temporal correlation. In this work, we show that SGDm under covariate shift with a fixed step-size can be unstable and diverge. In particular, we show SGDm under covariate shift is a parametric oscillator, and so can suffer from a phenomenon known as resonance. We approximate the learning system as a time varying system of ordinary differential equations, and leverage existing theory to characterize the system's divergence/convergence as resonant/nonresonant modes. The theoretical result is limited to the linear setting with periodic covariate shift, so we empirically supplement this result to show that resonance phenomena persist even under non-periodic covariate shift, nonlinear dynamics with neural networks, and optimizers other than SGDm.

Matthew Vandergrift, Martha White, Yury Polyanskiy et al.

Many learning problems require predicting how populations evolve under an unknown transformation. A natural representation for such populations is a probability measure, with point clouds as a key example. In this work, we study the measure-to-measure (M2M) regression problem, in which one seeks to learn a map between probability measures from a finite collection of observed input-output pairs. In contrast to classical regression, where individual samples are transformed independently, M2M regression treats entire distributions as the data points. This perspective is vital in certain scientific applications, for example, cellular and molecular biology, where cells are known to evolve not as independent data points but as a collection. However, few existing approaches address the problem of M2M regression with sufficient expressivity and scalability. We present a formalization of nonlinear M2M regression and introduce two easy-to-use, expressive, and scalable approaches to learn such operators: transformers as static M2M maps and transformers as dynamic M2M velocity fields. Our approach leverages the natural measure-dependent and mean-field structure of transformers to learn nonlinear M2M maps on the space of probability distributions. We illustrate the effectiveness of our proposed method to generalize to unseen measures on synthetic experiments, interacting particle systems, and a large-scale patient-derived organoid dataset for predicting treatment response in colorectal cancer.

Maximilian Bloor, Martha White, Ehecatl Antonio del Rio Chanona et al.

Electrified chemical processes are incentivized by exposure to time-varying electricity markets to operate flexibly, but participating in demand response schemes can require satisfying terminal constraints over long horizons. Specifically, terminal constraints may be required when computing optimal schedules in order to preserve dynamic stability. Model-based optimization methods are computationally costly, and data-driven scheduling via reinforcement learning (RL) faces severe credit-assignment challenges. We integrate Goal-Space Planning (GSP) with Deep Deterministic Policy Gradient (DDPG), using learned temporally abstract models over discrete subgoals to propagate value across extended horizons. Using a simulated air separation benchmark, we demonstrate the proposed approach improves sample efficiency over standard DDPG while satisfying terminal storage constraints, mitigating myopic control behavior.

Andrew Patterson, Samuel Neumann, Raksha Kumaraswamy et al.

This paper introduces a new empirical methodology, the Cross-environment Hyperparameter Setting Benchmark, that compares RL algorithms across environments using a single hyperparameter setting, encouraging algorithmic development which is insensitive to hyperparameters. We demonstrate that this benchmark is robust to statistical noise and obtains qualitatively similar results across repeated applications, even when using few samples. This robustness makes the benchmark computationally cheap to apply, allowing statistically sound insights at low cost. We demonstrate two example instantiations of the CHS, on a set of six small control environments (SC-CHS) and on the entire DM Control suite of 28 environments (DMC-CHS). Finally, to illustrate the applicability of the CHS to modern RL algorithms on challenging environments, we conduct a novel empirical study of an open question in the continuous control literature. We show, with high confidence, that there is no meaningful difference in performance between Ornstein-Uhlenbeck noise and uncorrelated Gaussian noise for exploration with the DDPG algorithm on the DMC-CHS.

Esraa Elelimy, Adam White, Michael Bowling et al.

Recurrent Neural Networks (RNNs) are used to learn representations in partially observable environments. For agents that learn online and continually interact with the environment, it is desirable to train RNNs with real-time recurrent learning (RTRL); unfortunately, RTRL is prohibitively expensive for standard RNNs. A promising direction is to use linear recurrent architectures (LRUs), where dense recurrent weights are replaced with a complex-valued diagonal, making RTRL efficient. In this work, we build on these insights to provide a lightweight but effective approach for training RNNs in online RL. We introduce Recurrent Trace Units (RTUs), a small modification on LRUs that we nonetheless find to have significant performance benefits over LRUs when trained with RTRL. We find RTUs significantly outperform other recurrent architectures across several partially observable environments while using significantly less computation.

Parham Mohammad Panahi, Andrew Patterson, Martha White et al.

Experience replay, the reuse of past data to improve sample efficiency, is ubiquitous in reinforcement learning. Though a variety of smart sampling schemes have been introduced to improve performance, uniform sampling by far remains the most common approach. One exception is Prioritized Experience Replay (PER), where sampling is done proportionally to TD errors, inspired by the success of prioritized sweeping in dynamic programming. The original work on PER showed improvements in Atari, but follow-up results were mixed. In this paper, we investigate several variations on PER, to attempt to understand where and when PER may be useful. Our findings in prediction tasks reveal that while PER can improve value propagation in tabular settings, behavior is significantly different when combined with neural networks. Certain mitigations $-$ like delaying target network updates to control generalization and using estimates of expected TD errors in PER to avoid chasing stochasticity $-$ can avoid large spikes in error with PER and neural networks but generally do not outperform uniform replay. In control tasks, none of the prioritized variants consistently outperform uniform replay. We present new insight into the interaction between prioritization, bootstrapping, and neural networks and propose several improvements for PER in tabular settings and noisy domains.

Esraa Elelimy, Brett Daley, Andrew Patterson et al.

Achieving fast and stable off-policy learning in deep reinforcement learning (RL) is challenging. Most existing methods rely on semi-gradient temporal-difference (TD) methods for their simplicity and efficiency, but are consequently susceptible to divergence. While more principled approaches like Gradient TD (GTD) methods have strong convergence guarantees, they have rarely been used in deep RL. Recent work introduced the generalized Projected Bellman Error ($\overline{\text{PBE}}$), enabling GTD methods to work efficiently with nonlinear function approximation. However, this work is limited to one-step methods, which are slow at credit assignment and require a large number of samples. In this paper, we extend the generalized $\overline{\text{PBE}}$ objective to support multistep credit assignment based on the $λ$-return and derive three gradient-based methods that optimize this new objective. We provide both a forward-view formulation compatible with experience replay and a backward-view formulation compatible with streaming algorithms. Finally, we evaluate the proposed algorithms and show that they outperform both PPO and StreamQ in MuJoCo and MinAtar environments, respectively. Code available at https://github.com/esraaelelimy/gtd\_algos

Khurram Javed, Martha White

A continual learning agent should be able to build on top of existing knowledge to learn on new data quickly while minimizing forgetting. Current intelligent systems based on neural network function approximators arguably do the opposite---they are highly prone to forgetting and rarely trained to facilitate future learning. One reason for this poor behavior is that they learn from a representation that is not explicitly trained for these two goals. In this paper, we propose OML, an objective that directly minimizes catastrophic interference by learning representations that accelerate future learning and are robust to forgetting under online updates in continual learning. We show that it is possible to learn naturally sparse representations that are more effective for online updating. Moreover, our algorithm is complementary to existing continual learning strategies, such as MER and GEM. Finally, we demonstrate that a basic online updating strategy on representations learned by OML is competitive with rehearsal based methods for continual learning. We release an implementation of our method at https://github.com/khurramjaved96/mrcl .

Muhammad Kamran Janjua, Haseeb Shah, Martha White et al.

In this paper we investigate the use of reinforcement-learning based prediction approaches for a real drinking-water treatment plant. Developing such a prediction system is a critical step on the path to optimizing and automating water treatment. Before that, there are many questions to answer about the predictability of the data, suitable neural network architectures, how to overcome partial observability and more. We first describe this dataset, and highlight challenges with seasonality, nonstationarity, partial observability, and heterogeneity across sensors and operation modes of the plant. We then describe General Value Function (GVF) predictions -- discounted cumulative sums of observations -- and highlight why they might be preferable to classical n-step predictions common in time series prediction. We discuss how to use offline data to appropriately pre-train our temporal difference learning (TD) agents that learn these GVF predictions, including how to select hyperparameters for online fine-tuning in deployment. We find that the TD-prediction agent obtains an overall lower normalized mean-squared error than the n-step prediction agent. Finally, we show the importance of learning in deployment, by comparing a TD agent trained purely offline with no online updating to a TD agent that learns online. This final result is one of the first to motivate the importance of adapting predictions in real-time, for non-stationary high-volume systems in the real world.

Ehsan Imani, Kai Luedemann, Sam Scholnick-Hughes et al.

It is becoming increasingly common in regression to train neural networks that model the entire distribution even if only the mean is required for prediction. This additional modeling often comes with performance gain and the reasons behind the improvement are not fully known. This paper investigates a recent approach to regression, the Histogram Loss, which involves learning the conditional distribution of the target variable by minimizing the cross-entropy between a target distribution and a flexible histogram prediction. We design theoretical and empirical analyses to determine why and when this performance gain appears, and how different components of the loss contribute to it. Our results suggest that the benefits of learning distributions in this setup come from improvements in optimization rather than modelling extra information. We then demonstrate the viability of the Histogram Loss in common deep learning applications without a need for costly hyperparameter tuning.

Gautham Vasan, Mohamed Elsayed, Alireza Azimi et al.

Modern deep policy gradient methods achieve effective performance on simulated robotic tasks, but they all require large replay buffers or expensive batch updates, or both, making them incompatible for real systems with resource-limited computers. We show that these methods fail catastrophically when limited to small replay buffers or during incremental learning, where updates only use the most recent sample without batch updates or a replay buffer. We propose a novel incremental deep policy gradient method -- Action Value Gradient (AVG) and a set of normalization and scaling techniques to address the challenges of instability in incremental learning. On robotic simulation benchmarks, we show that AVG is the only incremental method that learns effectively, often achieving final performance comparable to batch policy gradient methods. This advancement enabled us to show for the first time effective deep reinforcement learning with real robots using only incremental updates, employing a robotic manipulator and a mobile robot.

Brett Daley, Martha White, Marlos C. Machado

Multistep returns, such as $n$-step returns and $λ$-returns, are commonly used to improve the sample efficiency of reinforcement learning (RL) methods. The variance of the multistep returns becomes the limiting factor in their length; looking too far into the future increases variance and reverses the benefits of multistep learning. In our work, we demonstrate the ability of compound returns -- weighted averages of $n$-step returns -- to reduce variance. We prove for the first time that any compound return with the same contraction modulus as a given $n$-step return has strictly lower variance. We additionally prove that this variance-reduction property improves the finite-sample complexity of temporal-difference learning under linear function approximation. Because general compound returns can be expensive to implement, we introduce two-bootstrap returns which reduce variance while remaining efficient, even when using minibatched experience replay. We conduct experiments showing that compound returns often increase the sample efficiency of $n$-step deep RL agents like DQN and PPO.

Vincent Liu, Prabhat Nagarajan, Andrew Patterson et al.

Offline reinforcement learning algorithms often require careful hyperparameter tuning. Consequently, before deployment, we need to select amongst a set of candidate policies. As yet, however, there is little understanding about the fundamental limits of this offline policy selection (OPS) problem. In this work we aim to provide clarity on when sample efficient OPS is possible, primarily by connecting OPS to off-policy policy evaluation (OPE) and Bellman error (BE) estimation. We first show a hardness result, that in the worst case, OPS is just as hard as OPE, by proving a reduction of OPE to OPS. As a result, no OPS method can be more sample efficient than OPE in the worst case. We then propose a BE method for OPS, called Identifiable BE Selection (IBES), that has a straightforward method for selecting its own hyperparameters. We highlight that using IBES for OPS generally has more requirements than OPE methods, but if satisfied, can be more sample efficient. We conclude with an empirical study comparing OPE and IBES, and by showing the difficulty of OPS on an offline Atari benchmark dataset.

Golnaz Mesbahi, Parham Mohammad Panahi, Olya Mastikhina et al.

In continual RL we want agents capable of never-ending learning, and yet our evaluation methodologies do not reflect this. The standard practice in RL is to assume unfettered access to the deployment environment for the full lifetime of the agent. For example, agent designers select the best performing hyperparameters in Atari by testing each for 200 million frames and then reporting results on 200 million frames. In this position paper, we argue and demonstrate the pitfalls of this inappropriate empirical methodology: lifetime tuning. We provide empirical evidence to support our position by testing DQN and SAC across several of continuing and non-stationary environments with two main findings: (1) lifetime tuning does not allow us to identify algorithms that work well for continual learning -- all algorithms equally succeed; (2) recently developed continual RL algorithms outperform standard non-continual algorithms when tuning is limited to a fraction of the agent's lifetime. The goal of this paper is to provide an explanation for why recent progress in continual RL has been mixed and motivate the development of empirical practices that better match the goals of continual RL.

Han Wang, Adam White, Martha White

Fine-tuning policies learned offline remains a major challenge in application domains. Monotonic performance improvement during \emph{fine-tuning} is often challenging, as agents typically experience performance degradation at the early fine-tuning stage. The community has identified multiple difficulties in fine-tuning a learned network online, however, the majority of progress has focused on improving learning efficiency during fine-tuning. In practice, this comes at a serious cost during fine-tuning: initially, agent performance degrades as the agent explores and effectively overrides the policy learned offline. We show across a range of settings, many offline-to-online algorithms exhibit either (1) performance degradation or (2) slow learning (sometimes effectively no improvement) during fine-tuning. We introduce a new fine-tuning algorithm, based on an algorithm called Jump Start, that gradually allows more exploration based on online estimates of performance. Empirically, this approach achieves fast fine-tuning and significantly reduces performance degradations compared with existing algorithms designed to do the same.

Lingwei Zhu, Zheng Chen, Han Wang et al.

This paper introduces symmetric divergences to behavior regularization policy optimization (BRPO) to establish a novel offline RL framework. Existing methods focus on asymmetric divergences such as KL to obtain analytic regularized policies and a practical minimization objective. We show that symmetric divergences do not permit an analytic policy as regularization and can incur numerical issues as loss. We tackle these challenges by the Taylor series of $f$-divergence. Specifically, we prove that an analytic policy can be obtained with a finite series. For loss, we observe that symmetric divergences can be decomposed into an asymmetry and a conditional symmetry term, Taylor-expanding the latter alleviates numerical issues. Summing together, we propose Symmetric $f$ Actor-Critic (S$f$-AC), the first practical BRPO algorithm with symmetric divergences. Experimental results on distribution approximation and MuJoCo verify that S$f$-AC performs competitively.

Brett Daley, Prabhat Nagarajan, Martha White et al.

The hallmark feature of temporal-difference (TD) learning is bootstrapping: using value predictions to generate new value predictions. The vast majority of TD methods for control learn a policy by bootstrapping from a single action-value function (e.g., Q-learning and Sarsa). Significantly less attention has been given to methods that bootstrap from two asymmetric value functions: i.e., methods that learn state values as an intermediate step in learning action values. Existing algorithms in this vein can be categorized as either QV-learning or AV-learning. Though these algorithms have been investigated to some degree in prior work, it remains unclear if and when it is advantageous to learn two value functions instead of just one -- and whether such approaches are theoretically sound in general. In this paper, we analyze these algorithmic families in terms of convergence and sample efficiency. We find that while both families are more efficient than Expected Sarsa in the prediction setting, only AV-learning methods offer any major benefit over Q-learning in the control setting. Finally, we introduce a new AV-learning algorithm called Regularized Dueling Q-learning (RDQ), which significantly outperforms Dueling DQN in the MinAtar benchmark.

Prabhat Nagarajan, Martha White, Marlos C. Machado

Overestimation is pervasive in reinforcement learning (RL), including in Q-learning, which forms the algorithmic basis for many value-based deep RL algorithms. Double Q-learning is an algorithm introduced to address Q-learning's overestimation by training two Q-functions and using both to de-correlate action-selection and action-evaluation in bootstrap targets. Shortly after Q-learning was adapted to deep RL in the form of deep Q-networks (DQN), Double Q-learning was adapted to deep RL in the form of Double DQN. However, Double DQN only loosely adapts Double Q-learning, forgoing the training of two different Q-functions that bootstrap off one another. In this paper, we study algorithms that adapt this core idea of Double Q-learning for value-based deep RL. We term such algorithms Deep Double Q-learning (DDQL). Our aim is to understand whether DDQL exhibits less overestimation than Double DQN and whether performant instantiations of DDQL exist. We answer both questions affirmatively, demonstrating that DDQL reduces overestimation and outperforms Double DQN in aggregate across 57 Atari 2600 games, without requiring additional hyperparameters. We also study several aspects of DDQL, including its network architecture, replay ratio, and minibatch sampling strategy.

Scott M. Jordan, Adam White, Bruno Castro da Silva et al.

Novel reinforcement learning algorithms, or improvements on existing ones, are commonly justified by evaluating their performance on benchmark environments and are compared to an ever-changing set of standard algorithms. However, despite numerous calls for improvements, experimental practices continue to produce misleading or unsupported claims. One reason for the ongoing substandard practices is that conducting rigorous benchmarking experiments requires substantial computational time. This work investigates the sources of increased computation costs in rigorous experiment designs. We show that conducting rigorous performance benchmarks will likely have computational costs that are often prohibitive. As a result, we argue for using an additional experimentation paradigm to overcome the limitations of benchmarking.

Brett Daley, Marlos C. Machado, Martha White

The recency heuristic in reinforcement learning is the assumption that stimuli that occurred closer in time to an acquired reward should be more heavily reinforced. The recency heuristic is one of the key assumptions made by TD($λ$), which reinforces recent experiences according to an exponentially decaying weighting. In fact, all other widely used return estimators for TD learning, such as $n$-step returns, satisfy a weaker (i.e., non-monotonic) recency heuristic. Why is the recency heuristic effective for temporal credit assignment? What happens when credit is assigned in a way that violates this heuristic? In this paper, we analyze the specific mathematical implications of adopting the recency heuristic in TD learning. We prove that any return estimator satisfying this heuristic: 1) is guaranteed to converge to the correct value function, 2) has a relatively fast contraction rate, and 3) has a long window of effective credit assignment, yet bounded worst-case variance. We also give a counterexample where on-policy, tabular TD methods violating the recency heuristic diverge. Our results offer some of the first theoretical evidence that credit assignment based on the recency heuristic facilitates learning.

Kevin Roice, Parham Mohammad Panahi, Scott M. Jordan et al.

This paper investigates a new approach to model-based reinforcement learning using background planning: mixing (approximate) dynamic programming updates and model-free updates, similar to the Dyna architecture. Background planning with learned models is often worse than model-free alternatives, such as Double DQN, even though the former uses significantly more memory and computation. The fundamental problem is that learned models can be inaccurate and often generate invalid states, especially when iterated many steps. In this paper, we avoid this limitation by constraining background planning to a set of (abstract) subgoals and learning only local, subgoal-conditioned models. This goal-space planning (GSP) approach is more computationally efficient, naturally incorporates temporal abstraction for faster long-horizon planning and avoids learning the transition dynamics entirely. We show that our GSP algorithm can propagate value from an abstract space in a manner that helps a variety of base learners learn significantly faster in different domains.

Hugo Silva, Martha White

Oftentimes, machine learning applications using neural networks involve solving discrete optimization problems, such as in pruning, parameter-isolation-based continual learning and training of binary networks. Still, these discrete problems are combinatorial in nature and are also not amenable to gradient-based optimization. Additionally, classical approaches used in discrete settings do not scale well to large neural networks, forcing scientists and empiricists to rely on alternative methods. Among these, two main distinct sources of top-down information can be used to lead the model to good solutions: (1) extrapolating gradient information from points outside of the solution set (2) comparing evaluations between members of a subset of the valid solutions. We take continuation path (CP) methods to represent using purely the former and Monte Carlo (MC) methods to represent the latter, while also noting that some hybrid methods combine the two. The main goal of this work is to compare both approaches. For that purpose, we first overview the two classes while also discussing some of their drawbacks analytically. Then, on the experimental section, we compare their performance, starting with smaller microworld experiments, which allow more fine-grained control of problem variables, and gradually moving towards larger problems, including neural network regression and neural network pruning for image classification, where we additionally compare against magnitude-based pruning.

James E. Kostas, Scott M. Jordan, Yash Chandak et al.

Coagent networks for reinforcement learning (RL) [Thomas and Barto, 2011] provide a powerful and flexible framework for deriving principled learning rules for arbitrary stochastic neural networks. The coagent framework offers an alternative to backpropagation-based deep learning (BDL) that overcomes some of backpropagation's main limitations. For example, coagent networks can compute different parts of the network \emph{asynchronously} (at different rates or at different times), can incorporate non-differentiable components that cannot be used with backpropagation, and can explore at levels higher than their action spaces (that is, they can be designed as hierarchical networks for exploration and/or temporal abstraction). However, the coagent framework is not just an alternative to BDL; the two approaches can be blended: BDL can be combined with coagent learning rules to create architectures with the advantages of both approaches. This work generalizes the coagent theory and learning rules provided by previous works; this generalization provides more flexibility for network architecture design within the coagent framework. This work also studies one of the chief disadvantages of coagent networks: high variance updates for networks that have many coagents and do not use backpropagation. We show that a coagent algorithm with a policy network that does not use backpropagation can scale to a challenging RL domain with a high-dimensional state and action space (the MuJoCo Ant environment), learning reasonable (although not state-of-the-art) policies. These contributions motivate and provide a more general theoretical foundation for future work that studies coagent networks.

Han Wang, Erfan Miahi, Martha White et al.

In this paper we investigate the properties of representations learned by deep reinforcement learning systems. Much of the early work on representations for reinforcement learning focused on designing fixed-basis architectures to achieve properties thought to be desirable, such as orthogonality and sparsity. In contrast, the idea behind deep reinforcement learning methods is that the agent designer should not encode representational properties, but rather that the data stream should determine the properties of the representation -- good representations emerge under appropriate training schemes. In this paper we bring these two perspectives together, empirically investigating the properties of representations that support transfer in reinforcement learning. We introduce and measure six representational properties over more than 25 thousand agent-task settings. We consider Deep Q-learning agents with different auxiliary losses in a pixel-based navigation environment, with source and transfer tasks corresponding to different goal locations. We develop a method to better understand why some representations work better for transfer, through a systematic approach varying task similarity and measuring and correlating representation properties with transfer performance. We demonstrate the generality of the methodology by investigating representations learned by a Rainbow agent that successfully transfer across games modes in Atari 2600.

Matthew McLeod, Chunlok Lo, Matthew Schlegel et al.

Learning auxiliary tasks, such as multiple predictions about the world, can provide many benefits to reinforcement learning systems. A variety of off-policy learning algorithms have been developed to learn such predictions, but as yet there is little work on how to adapt the behavior to gather useful data for those off-policy predictions. In this work, we investigate a reinforcement learning system designed to learn a collection of auxiliary tasks, with a behavior policy learning to take actions to improve those auxiliary predictions. We highlight the inherent non-stationarity in this continual auxiliary task learning problem, for both prediction learners and the behavior learner. We develop an algorithm based on successor features that facilitates tracking under non-stationary rewards, and prove the separation into learning successor features and rewards provides convergence rate improvements. We conduct an in-depth study into the resulting multi-prediction learning system.

Samuele Tosatto, Andrew Patterson, Martha White et al.

The policy gradient theorem (Sutton et al., 2000) prescribes the usage of a cumulative discounted state distribution under the target policy to approximate the gradient. Most algorithms based on this theorem, in practice, break this assumption, introducing a distribution shift that can cause the convergence to poor solutions. In this paper, we propose a new approach of reconstructing the policy gradient from the start state without requiring a particular sampling strategy. The policy gradient calculation in this form can be simplified in terms of a gradient critic, which can be recursively estimated due to a new Bellman equation of gradients. By using temporal-difference updates of the gradient critic from an off-policy data stream, we develop the first estimator that sidesteps the distribution shift issue in a model-free way. We prove that, under certain realizability conditions, our estimator is unbiased regardless of the sampling strategy. We empirically show that our technique achieves a superior bias-variance trade-off and performance in presence of off-policy samples.

Shivam Garg, Samuele Tosatto, Yangchen Pan et al.

Policy gradient (PG) estimators are ineffective in dealing with softmax policies that are sub-optimally saturated, which refers to the situation when the policy concentrates its probability mass on sub-optimal actions. Sub-optimal policy saturation may arise from bad policy initialization or sudden changes in the environment that occur after the policy has already converged. Current softmax PG estimators require a large number of updates to overcome policy saturation, which causes low sample efficiency and poor adaptability to new situations. To mitigate this problem, we propose a novel PG estimator for softmax policies that utilizes the bias in the critic estimate and the noise present in the reward signal to escape the saturated regions of the policy parameter space. Our theoretical analysis and experiments, conducted on bandits and various reinforcement learning environments, show that this new estimator is significantly more robust to policy saturation.

Ehsan Imani, Wei Hu, Martha White

It is now a standard for neural network representations to be trained on large, publicly available datasets, and used for new problems. The reasons for why neural network representations have been so successful for transfer, however, are still not fully understood. In this paper we show that, after training, neural network representations align their top singular vectors to the targets. We investigate this representation alignment phenomenon in a variety of neural network architectures and find that (a) alignment emerges across a variety of different architectures and optimizers, with more alignment arising from depth (b) alignment increases for layers closer to the output and (c) existing high-performance deep CNNs exhibit high levels of alignment. We then highlight why alignment between the top singular vectors and the targets can speed up learning and show in a classic synthetic transfer problem that representation alignment correlates with positive and negative transfer to similar and dissimilar tasks.

Eric Graves, Ehsan Imani, Raksha Kumaraswamy et al.

A variety of theoretically-sound policy gradient algorithms exist for the on-policy setting due to the policy gradient theorem, which provides a simplified form for the gradient. The off-policy setting, however, has been less clear due to the existence of multiple objectives and the lack of an explicit off-policy policy gradient theorem. In this work, we unify these objectives into one off-policy objective, and provide a policy gradient theorem for this unified objective. The derivation involves emphatic weightings and interest functions. We show multiple strategies to approximate the gradients, in an algorithm called Actor Critic with Emphatic weightings (ACE). We prove in a counterexample that previous (semi-gradient) off-policy actor-critic methods--particularly Off-Policy Actor-Critic (OffPAC) and Deterministic Policy Gradient (DPG)--converge to the wrong solution whereas ACE finds the optimal solution. We also highlight why these semi-gradient approaches can still perform well in practice, suggesting strategies for variance reduction in ACE. We empirically study several variants of ACE on two classic control environments and an image-based environment designed to illustrate the tradeoffs made by each gradient approximation. We find that by approximating the emphatic weightings directly, ACE performs as well as or better than OffPAC in all settings tested.

Vincent Liu, James R. Wright, Martha White

Offline reinforcement learning -- learning a policy from a batch of data -- is known to be hard for general MDPs. These results motivate the need to look at specific classes of MDPs where offline reinforcement learning might be feasible. In this work, we explore a restricted class of MDPs to obtain guarantees for offline reinforcement learning. The key property, which we call Action Impact Regularity (AIR), is that actions primarily impact a part of the state (an endogenous component) and have limited impact on the remaining part of the state (an exogenous component). AIR is a strong assumption, but it nonetheless holds in a number of real-world domains including financial markets. We discuss algorithms that exploit the AIR property, and provide a theoretical analysis for an algorithm based on Fitted-Q Iteration. Finally, we demonstrate that the algorithm outperforms existing offline reinforcement learning algorithms across different data collection policies in simulated and real world environments where the regularity holds.

Alan Chan, Hugo Silva, Sungsu Lim et al.

Approximate Policy Iteration (API) algorithms alternate between (approximate) policy evaluation and (approximate) greedification. Many different approaches have been explored for approximate policy evaluation, but less is understood about approximate greedification and what choices guarantee policy improvement. In this work, we investigate approximate greedification when reducing the KL divergence between the parameterized policy and the Boltzmann distribution over action values. In particular, we investigate the difference between the forward and reverse KL divergences, with varying degrees of entropy regularization. We show that the reverse KL has stronger policy improvement guarantees, but that reducing the forward KL can result in a worse policy. We also demonstrate, however, that a large enough reduction of the forward KL can induce improvement under additional assumptions. Empirically, we show on simple continuous-action environments that the forward KL can induce more exploration, but at the cost of a more suboptimal policy. No significant differences were observed in the discrete-action setting or on a suite of benchmark problems. Throughout, we highlight that many policy gradient methods can be seen as an instance of API, with either the forward or reverse KL for the policy update, and discuss next steps for understanding and improving our policy optimization algorithms.

Qingfeng Lan, Luke Kumar, Martha White et al.

To effectively perform the task of next-word prediction, long short-term memory networks (LSTMs) must keep track of many types of information. Some information is directly related to the next word's identity, but some is more secondary (e.g. discourse-level features or features of downstream words). Correlates of secondary information appear in LSTM representations even though they are not part of an \emph{explicitly} supervised prediction task. In contrast, in reinforcement learning (RL), techniques that explicitly supervise representations to predict secondary information have been shown to be beneficial. Inspired by that success, we propose Predictive Representation Learning (PRL), which explicitly constrains LSTMs to encode specific predictions, like those that might need to be learned implicitly. We show that PRL 1) significantly improves two strong language modeling methods, 2) converges more quickly, and 3) performs better when data is limited. Our work shows that explicitly encoding a simple predictive task facilitates the search for a more effective language model.

Andrew Patterson, Adam White, Martha White

Many reinforcement learning algorithms rely on value estimation, however, the most widely used algorithms -- namely temporal difference algorithms -- can diverge under both off-policy sampling and nonlinear function approximation. Many algorithms have been developed for off-policy value estimation based on the linear mean squared projected Bellman error (MSPBE) and are sound under linear function approximation. Extending these methods to the nonlinear case has been largely unsuccessful. Recently, several methods have been introduced that approximate a different objective -- the mean-squared Bellman error (MSBE) -- which naturally facilitate nonlinear approximation. In this work, we build on these insights and introduce a new generalized MSPBE that extends the linear MSPBE to the nonlinear setting. We show how this generalized objective unifies previous work and obtain new bounds for the value error of the solutions of the generalized objective. We derive an easy-to-use, but sound, algorithm to minimize the generalized objective, and show that it is more stable across runs, is less sensitive to hyperparameters, and performs favorably across four control domains with neural network function approximation.

Khurram Javed, Martha White, Rich Sutton

Structural credit assignment for recurrent learning is challenging. An algorithm called RTRL can compute gradients for recurrent networks online but is computationally intractable for large networks. Alternatives, such as BPTT, are not online. In this work, we propose a credit-assignment algorithm -- \algoname{} -- that approximates the gradients for recurrent learning in real-time using $O(n)$ operations and memory per-step. Our method builds on the idea that for modular recurrent networks, composed of columns with scalar states, it is sufficient for a parameter to only track its influence on the state of its column. We empirically show that as long as connections between columns are sparse, our method approximates the true gradient well. In the special case when there are no connections between columns, the $O(n)$ gradient estimate is exact. We demonstrate the utility of the approach for both recurrent state learning and meta-learning by comparing the estimated gradient to the true gradient on a synthetic test-bed.

Sebastian Höfer, Kostas Bekris, Ankur Handa et al.

This report presents the debates, posters, and discussions of the Sim2Real workshop held in conjunction with the 2020 edition of the "Robotics: Science and System" conference. Twelve leaders of the field took competing debate positions on the definition, viability, and importance of transferring skills from simulation to the real world in the context of robotics problems. The debaters also joined a large panel discussion, answering audience questions and outlining the future of Sim2Real in robotics. Furthermore, we invited extended abstracts to this workshop which are summarized in this report. Based on the workshop, this report concludes with directions for practitioners exploiting this technology and for researchers further exploring open problems in this area.

Yash Chandak, Scott M. Jordan, Georgios Theocharous et al.

Many real-world sequential decision-making problems involve critical systems with financial risks and human-life risks. While several works in the past have proposed methods that are safe for deployment, they assume that the underlying problem is stationary. However, many real-world problems of interest exhibit non-stationarity, and when stakes are high, the cost associated with a false stationarity assumption may be unacceptable. We take the first steps towards ensuring safety, with high confidence, for smoothly-varying non-stationary decision problems. Our proposed method extends a type of safe algorithm, called a Seldonian algorithm, through a synthesis of model-free reinforcement learning with time-series analysis. Safety is ensured using sequential hypothesis testing of a policy's forecasted performance, and confidence intervals are obtained using wild bootstrap.

Maryam Hashemzadeh, Greta Kaufeld, Martha White et al.

The representations generated by many models of language (word embeddings, recurrent neural networks and transformers) correlate to brain activity recorded while people read. However, these decoding results are usually based on the brain's reaction to syntactically and semantically sound language stimuli. In this study, we asked: how does an LSTM (long short term memory) language model, trained (by and large) on semantically and syntactically intact language, represent a language sample with degraded semantic or syntactic information? Does the LSTM representation still resemble the brain's reaction? We found that, even for some kinds of nonsensical language, there is a statistically significant relationship between the brain's activity and the representations of an LSTM. This indicates that, at least in some instances, LSTMs and the human brain handle nonsensical data similarly.

Yangchen Pan, Jincheng Mei, Amir-massoud Farahmand et al.

Prioritized Experience Replay (ER) has been empirically shown to improve sample efficiency across many domains and attracted great attention; however, there is little theoretical understanding of why such prioritized sampling helps and its limitations. In this work, we take a deep look at the prioritized ER. In a supervised learning setting, we show the equivalence between the error-based prioritized sampling method for mean squared error and uniform sampling for cubic power loss. We then provide theoretical insight into why it improves convergence rate upon uniform sampling during early learning. Based on the insight, we further point out two limitations of the prioritized ER method: 1) outdated priorities and 2) insufficient coverage of the sample space. To mitigate the limitations, we propose our model-based stochastic gradient Langevin dynamics sampling method. We show that our method does provide states distributed close to an ideal prioritized sampling distribution estimated by the brute-force method, which does not suffer from the two limitations. We conduct experiments on both discrete and continuous control problems to show our approach's efficacy and examine the practical implication of our method in an autonomous driving application.

Vincent Liu, Adam White, Hengshuai Yao et al.

Catastrophic interference is common in many network-based learning systems, and many proposals exist for mitigating it. But, before we overcome interference we must understand it better. In this work, we provide a definition of interference for control in reinforcement learning. We systematically evaluate our new measures, by assessing correlation with several measures of learning performance, including stability, sample efficiency, and online and offline control performance across a variety of learning architectures. Our new interference measure allows us to ask novel scientific questions about commonly used deep learning architectures. In particular we show that target network frequency is a dominating factor for interference, and that updates on the last layer result in significantly higher interference than updates internal to the network. This new measure can be expensive to compute; we conclude with motivation for an efficient proxy measure and empirically demonstrate it is correlated with our definition of interference.

Zaheer Abbas, Samuel Sokota, Erin J. Talvitie et al.

In model-based reinforcement learning, planning with an imperfect model of the environment has the potential to harm learning progress. But even when a model is imperfect, it may still contain information that is useful for planning. In this paper, we investigate the idea of using an imperfect model selectively. The agent should plan in parts of the state space where the model would be helpful but refrain from using the model where it would be harmful. An effective selective planning mechanism requires estimating predictive uncertainty, which arises out of aleatoric uncertainty, parameter uncertainty, and model inadequacy, among other sources. Prior work has focused on parameter uncertainty for selective planning. In this work, we emphasize the importance of model inadequacy. We show that heteroscedastic regression can signal predictive uncertainty arising from model inadequacy that is complementary to that which is detected by methods designed for parameter uncertainty, indicating that considering both parameter uncertainty and model inadequacy may be a more promising direction for effective selective planning than either in isolation.

Sina Ghiassian, Andrew Patterson, Shivam Garg et al.

It is still common to use Q-learning and temporal difference (TD) learning-even though they have divergence issues and sound Gradient TD alternatives exist-because divergence seems rare and they typically perform well. However, recent work with large neural network learning systems reveals that instability is more common than previously thought. Practitioners face a difficult dilemma: choose an easy to use and performant TD method, or a more complex algorithm that is more sound but harder to tune and all but unexplored with non-linear function approximation or control. In this paper, we introduce a new method called TD with Regularized Corrections (TDRC), that attempts to balance ease of use, soundness, and performance. It behaves as well as TD, when TD performs well, but is sound in cases where TD diverges. We empirically investigate TDRC across a range of problems, for both prediction and control, and for both linear and non-linear function approximation, and show, potentially for the first time, that gradient TD methods could be a better alternative to TD and Q-learning.

Khurram Javed, Martha White, Yoshua Bengio

Predictive models -- learned from observational data not covering the complete data distribution -- can rely on spurious correlations in the data for making predictions. These correlations make the models brittle and hinder generalization. One solution for achieving strong generalization is to incorporate causal structures in the models; such structures constrain learning by ignoring correlations that contradict them. However, learning these structures is a hard problem in itself. Moreover, it's not clear how to incorporate the machinery of causality with online continual learning. In this work, we take an indirect approach to discovering causal models. Instead of searching for the true causal model directly, we propose an online algorithm that continually detects and removes spurious features. Our algorithm works on the idea that the correlation of a spurious feature with a target is not constant over-time. As a result, the weight associated with that feature is constantly changing. We show that by continually removing such features, our method converges to solutions that have strong generalization. Moreover, our method combined with random search can also discover non-spurious features from raw sensory data. Finally, our work highlights that the information present in the temporal structure of the problem -- destroyed by shuffling the data -- is essential for detecting spurious features online.

Taher Jafferjee, Ehsan Imani, Erin Talvitie et al.

Dyna-style reinforcement learning (RL) agents improve sample efficiency over model-free RL agents by updating the value function with simulated experience generated by an environment model. However, it is often difficult to learn accurate models of environment dynamics, and even small errors may result in failure of Dyna agents. In this paper, we investigate one type of model error: hallucinated states. These are states generated by the model, but that are not real states of the environment. We present the Hallucinated Value Hypothesis (HVH): updating values of real states towards values of hallucinated states results in misleading state-action values which adversely affect the control policy. We discuss and evaluate four Dyna variants; three which update real states toward simulated -- and therefore potentially hallucinated -- states and one which does not. The experimental results provide evidence for the HVH thus suggesting a fruitful direction toward developing Dyna algorithms robust to model error.