Marlos C. Machado

Zaheer Abbas, Rosie Zhao, Joseph Modayil et al. · deepmind

The ability to learn continually is essential in a complex and changing world. In this paper, we characterize the behavior of canonical value-based deep reinforcement learning (RL) approaches under varying degrees of non-stationarity. In particular, we demonstrate that deep RL agents lose their ability to learn good policies when they cycle through a sequence of Atari 2600 games. This phenomenon is alluded to in prior work under various guises -- e.g., loss of plasticity, implicit under-parameterization, primacy bias, and capacity loss. We investigate this phenomenon closely at scale and analyze how the weights, gradients, and activations change over time in several experiments with varying dimensions (e.g., similarity between games, number of games, number of frames per game), with some experiments spanning 50 days and 2 billion environment interactions. Our analysis shows that the activation footprint of the network becomes sparser, contributing to the diminishing gradients. We investigate a remarkably simple mitigation strategy -- Concatenated ReLUs (CReLUs) activation function -- and demonstrate its effectiveness in facilitating continual learning in a changing environment.

Akram Erraqabi, Marlos C. Machado, Mingde Zhao et al. · meta-ai, mila

In reinforcement learning, the graph Laplacian has proved to be a valuable tool in the task-agnostic setting, with applications ranging from skill discovery to reward shaping. Recently, learning the Laplacian representation has been framed as the optimization of a temporally-contrastive objective to overcome its computational limitations in large (or continuous) state spaces. However, this approach requires uniform access to all states in the state space, overlooking the exploration problem that emerges during the representation learning process. In this work, we propose an alternative method that is able to recover, in a non-uniform-prior setting, the expressiveness and the desired properties of the Laplacian representation. We do so by combining the representation learning with a skill-based covering policy, which provides a better training distribution to extend and refine the representation. We also show that a simple augmentation of the representation objective with the learned temporal abstractions improves dynamics-awareness and helps exploration. We find that our method succeeds as an alternative to the Laplacian in the non-uniform setting and scales to challenging continuous control environments. Finally, even if our method is not optimized for skill discovery, the learned skills can successfully solve difficult continuous navigation tasks with sparse rewards, where standard skill discovery approaches are no so effective.

Ruo Yu Tao, Adam White, Marlos C. Machado · deepmind

In many, if not every realistic sequential decision-making task, the decision-making agent is not able to model the full complexity of the world. The environment is often much larger and more complex than the agent, a setting also known as partial observability. In such settings, the agent must leverage more than just the current sensory inputs; it must construct an agent state that summarizes previous interactions with the world. Currently, a popular approach for tackling this problem is to learn the agent-state function via a recurrent network from the agent's sensory stream as input. Many impressive reinforcement learning applications have instead relied on environment-specific functions to aid the agent's inputs for history summarization. These augmentations are done in multiple ways, from simple approaches like concatenating observations to more complex ones such as uncertainty estimates. Although ubiquitous in the field, these additional inputs, which we term auxiliary inputs, are rarely emphasized, and it is not clear what their role or impact is. In this work we explore this idea further, and relate these auxiliary inputs to prior classic approaches to state construction. We present a series of examples illustrating the different ways of using auxiliary inputs for reinforcement learning. We show that these auxiliary inputs can be used to discriminate between observations that would otherwise be aliased, leading to more expressive features that smoothly interpolate between different states. Finally, we show that this approach is complementary to state-of-the-art methods such as recurrent neural networks and truncated back-propagation through time, and acts as a heuristic that facilitates longer temporal credit assignment, leading to better performance.

Martin Klissarov, Marlos C. Machado

Selecting exploratory actions that generate a rich stream of experience for better learning is a fundamental challenge in reinforcement learning (RL). An approach to tackle this problem consists in selecting actions according to specific policies for an extended period of time, also known as options. A recent line of work to derive such exploratory options builds upon the eigenfunctions of the graph Laplacian. Importantly, until now these methods have been mostly limited to tabular domains where (1) the graph Laplacian matrix was either given or could be fully estimated, (2) performing eigendecomposition on this matrix was computationally tractable, and (3) value functions could be learned exactly. Additionally, these methods required a separate option discovery phase. These assumptions are fundamentally not scalable. In this paper we address these limitations and show how recent results for directly approximating the eigenfunctions of the Laplacian can be leveraged to truly scale up options-based exploration. To do so, we introduce a fully online deep RL algorithm for discovering Laplacian-based options and evaluate our approach on a variety of pixel-based tasks. We compare to several state-of-the-art exploration methods and show that our approach is effective, general, and especially promising in non-stationary settings.

Alex Lewandowski, Haruto Tanaka, Dale Schuurmans et al.

Loss of plasticity is a phenomenon in which neural networks lose their ability to learn from new experience. Despite being empirically observed in several problem settings, little is understood about the mechanisms that lead to loss of plasticity. In this paper, we offer a consistent explanation for loss of plasticity: Neural networks lose directions of curvature during training and that loss of plasticity can be attributed to this reduction in curvature. To support such a claim, we provide a systematic investigation of loss of plasticity across continual learning tasks using MNIST, CIFAR-10 and ImageNet. Our findings illustrate that loss of curvature directions coincides with loss of plasticity, while also showing that previous explanations are insufficient to explain loss of plasticity in all settings. Lastly, we show that regularizers which mitigate loss of plasticity also preserve curvature, motivating a simple distributional regularizer that proves to be effective across the problem settings we considered.

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.

Diego Gomez, Michael Bowling, Marlos C. Machado

The ability to learn good representations of states is essential for solving large reinforcement learning problems, where exploration, generalization, and transfer are particularly challenging. The Laplacian representation is a promising approach to address these problems by inducing informative state encoding and intrinsic rewards for temporally-extended action discovery and reward shaping. To obtain the Laplacian representation one needs to compute the eigensystem of the graph Laplacian, which is often approximated through optimization objectives compatible with deep learning approaches. These approximations, however, depend on hyperparameters that are impossible to tune efficiently, converge to arbitrary rotations of the desired eigenvectors, and are unable to accurately recover the corresponding eigenvalues. In this paper we introduce a theoretically sound objective and corresponding optimization algorithm for approximating the Laplacian representation. Our approach naturally recovers both the true eigenvectors and eigenvalues while eliminating the hyperparameter dependence of previous approximations. We provide theoretical guarantees for our method and we show that those results translate empirically into robust learning across multiple environments.

Subhojeet Pramanik, Esraa Elelimy, Marlos C. Machado et al.

In this paper we investigate transformer architectures designed for partially observable online reinforcement learning. The self-attention mechanism in the transformer architecture is capable of capturing long-range dependencies and it is the main reason behind its effectiveness in processing sequential data. Nevertheless, despite their success, transformers have two significant drawbacks that still limit their applicability in online reinforcement learning: (1) in order to remember all past information, the self-attention mechanism requires access to the whole history to be provided as context. (2) The inference cost in transformers is expensive. In this paper, we introduce recurrent alternatives to the transformer self-attention mechanism that offer context-independent inference cost, leverage long-range dependencies effectively, and performs well in online reinforcement learning task. We quantify the impact of the different components of our architecture in a diagnostic environment and assess performance gains in 2D and 3D pixel-based partially-observable environments (e.g. T-Maze, Mystery Path, Craftax, and Memory Maze). Compared with a state-of-the-art architecture, GTrXL, inference in our approach is at least 40% cheaper while reducing memory use more than 50%. Our approach either performs similarly or better than GTrXL, improving more than 37% upon GTrXL performance in harder tasks.

Siddarth Chandrasekar, Marlos C. Machado

Across scientific disciplines, Laplacian eigenvectors serve as a fundamental basis for simplifying complex systems, from signal processing to quantum mechanics. In reinforcement learning (RL), they similarly form a basis over the state space, enabling reward functions to be approximated by projection onto a small set of eigenvectors. This projection makes zero-shot control possible, but it also imposes a fundamental limitation: the induced policies are only as expressive as the linear span of the chosen eigenvectors. We introduce the Laplacian Keyboard (LK), a hierarchical framework that goes beyond this linear span. LK constructs a task-agnostic library of behaviors from these eigenvectors, forming a behavior basis guaranteed to contain the optimal policy for any reward within the linear span. A meta-policy learns to stitch these behaviors dynamically, enabling efficient learning of policies outside the original linear constraints. We establish theoretical bounds on zero-shot approximation error and demonstrate empirically that LK improves over the zero-shot solution while achieving better sample efficiency compared to standard RL methods.

Hon Tik Tse, Marlos C. Machado

In computational reinforcement learning, the default representation (DR) and its principal eigenvector have been shown to be effective for a wide variety of applications, including reward shaping, count-based exploration, option discovery, and transfer. However, in prior investigations, the eigenvectors of the DR were computed by first approximating the DR matrix, and then performing an eigendecomposition. This procedure is computationally expensive and does not scale to high-dimensional spaces. In this paper, we derive an objective for directly approximating the principal eigenvector of the DR with a neural network. We empirically demonstrate the effectiveness of the objective in a number of environments, and apply the learned eigenvectors for reward shaping.

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

Alex Lewandowski, Adtiya A. Ramesh, Edan Meyer et al.

Continual learning is often motivated by the idea, known as the big world hypothesis, that "the world is bigger" than the agent. Recent problem formulations capture this idea by explicitly constraining an agent relative to the environment. These constraints lead to solutions in which the agent continually adapts to best use its limited capacity, rather than converging to a fixed solution. However, explicit constraints can be ad hoc, difficult to incorporate, and may limit the effectiveness of scaling up the agent's capacity. In this paper, we characterize a problem setting in which an agent, regardless of its capacity, is constrained by being embedded in the environment. In particular, we introduce a computationally-embedded perspective that represents an embedded agent as an automaton simulated within a universal (formal) computer. Such an automaton is always constrained; we prove that it is equivalent to an agent that interacts with a partially observable Markov decision process over a countably infinite state-space. We propose an objective for this setting, which we call interactivity, that measures an agent's ability to continually adapt its behaviour by learning new predictions. We then develop a model-based reinforcement learning algorithm for interactivity-seeking, and use it to construct a synthetic problem to evaluate continual learning capability. Our results show that deep nonlinear networks struggle to sustain interactivity, whereas deep linear networks sustain higher interactivity as capacity increases.

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.

Martin Klissarov, Mikael Henaff, Roberta Raileanu et al.

Describing skills in natural language has the potential to provide an accessible way to inject human knowledge about decision-making into an AI system. We present MaestroMotif, a method for AI-assisted skill design, which yields high-performing and adaptable agents. MaestroMotif leverages the capabilities of Large Language Models (LLMs) to effectively create and reuse skills. It first uses an LLM's feedback to automatically design rewards corresponding to each skill, starting from their natural language description. Then, it employs an LLM's code generation abilities, together with reinforcement learning, for training the skills and combining them to implement complex behaviors specified in language. We evaluate MaestroMotif using a suite of complex tasks in the NetHack Learning Environment (NLE), demonstrating that it surpasses existing approaches in both performance and usability.

Alex Lewandowski, Dale Schuurmans, Marlos C. Machado

Deep neural networks can struggle to learn continually in the face of non-stationarity. This phenomenon is known as loss of plasticity. In this paper, we identify underlying principles that lead to plastic algorithms. In particular, we provide theoretical results showing that linear function approximation, as well as a special case of deep linear networks, do not suffer from loss of plasticity. We then propose deep Fourier features, which are the concatenation of a sine and cosine in every layer, and we show that this combination provides a dynamic balance between the trainability obtained through linearity and the effectiveness obtained through the nonlinearity of neural networks. Deep networks composed entirely of deep Fourier features are highly trainable and sustain their trainability over the course of learning. Our empirical results show that continual learning performance can be drastically improved by replacing ReLU activations with deep Fourier features. These results hold for different continual learning scenarios (e.g., label noise, class incremental learning, pixel permutations) on all major supervised learning datasets used for continual learning research, such as CIFAR10, CIFAR100, and tiny-ImageNet.

Martin Klissarov, Akhil Bagaria, Ziyan Luo et al.

Developing agents capable of exploring, planning and learning in complex open-ended environments is a grand challenge in artificial intelligence (AI). Hierarchical reinforcement learning (HRL) offers a promising solution to this challenge by discovering and exploiting the temporal structure within a stream of experience. The strong appeal of the HRL framework has led to a rich and diverse body of literature attempting to discover a useful structure. However, it is still not clear how one might define what constitutes good structure in the first place, or the kind of problems in which identifying it may be helpful. This work aims to identify the benefits of HRL from the perspective of the fundamental challenges in decision-making, as well as highlight its impact on the performance trade-offs of AI agents. Through these benefits, we then cover the families of methods that discover temporal structure in HRL, ranging from learning directly from online experience to offline datasets, to leveraging large language models (LLMs). Finally, we highlight the challenges of temporal structure discovery and the domains that are particularly well-suited for such endeavours.

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.

Dikshant Shehmar, Matthew Schlegel, Matthew E. Taylor et al.

Planning with a learned model remains a key challenge in model-based reinforcement learning (RL). In decision-time planning, state representations are critical as they must support local cost computation while preserving long-horizon structure. In this paper, we show that the Laplacian representation provides an effective latent space for planning by capturing state-space distances at multiple time scales. This representation preserves meaningful distances and naturally decomposes long-horizon problems into subgoals, also mitigating the compounding errors that arise over long prediction horizons. Building on these properties, we introduce ALPS, a hierarchical planning algorithm, and demonstrate that it outperforms commonly used baselines on a selection of offline goal-conditioned RL tasks from OGBench, a benchmark previously dominated by model-free methods.

Alex Lewandowski, Marlos C. Machado, Dale Schuurmans

Language models now provide an interface to express and often solve general problems in natural language, yet their ultimate computational capabilities remain a major topic of scientific debate. Unlike a formal computer, a language model is trained to autoregressively predict successive elements in human-generated text. We prove that chaining a language model's autoregressive output is sufficient to perform universal computation. That is, a language model can simulate the execution of any algorithm on any input. The challenge of eliciting desired computational behaviour can thus be reframed in terms of programmability: the ease of finding a suitable prompt. Strikingly, we demonstrate that even randomly initialized language models are capable of universal computation before training. This implies that training does not give rise to computational expressiveness -- rather, it improves programmability, enabling a natural language interface for accessing these intrinsic capabilities.

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.

Harshil Kotamreddy, Marlos C. Machado

Options, which impose an inductive bias toward temporal and hierarchical structure, offer a powerful framework for reinforcement learning (RL). While effective in sequential decision-making, they are often handcrafted rather than learned. Among approaches for discovering options, eigenoptions have shown strong performance in exploration, but their role in credit assignment remains underexplored. In this paper, we investigate whether eigenoptions can accelerate credit assignment in model-free RL, evaluating them in tabular and pixel-based gridworlds. We find that pre-specified eigenoptions aid not only exploration but also credit assignment, whereas online discovery can bias the agent's experience too strongly and hinder learning. In the context of deep RL, we also propose a method for learning option-values under non-linear function approximation, highlighting the impact of termination conditions on performance. Our findings reveal both the promise and complexity of using eigenoptions, and options more broadly, to simultaneously support credit assignment and exploration in reinforcement learning.

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.

Mohamed A. Mohamed, Kateryna Nekhomiazh, Vedant Vyas et al.

Continual reinforcement learning (RL) concerns agents that are expected to learn continually, rather than converge to a policy that is then fixed for evaluation. Such an approach is well suited to environments the agent perceives as changing, which renders any static policy ineffective over time. The few simulators explicitly designed for empirical research in continual RL are often limited in scope or complexity, and it is now common for researchers to modify episodic RL environments by artificially incorporating abrupt task changes during interaction. In this paper, we introduce AgarCL, a research platform for continual RL that allows for a progression of increasingly sophisticated behaviour. AgarCL is based on the game Agar.io, a non-episodic, high-dimensional problem featuring stochastic, ever-evolving dynamics, continuous actions, and partial observability. Additionally, we provide benchmark results reporting the performance of DQN, PPO, and SAC in both the primary, challenging continual RL problem, and across a suite of smaller tasks within AgarCL, each of which isolates aspects of the full environment and allow us to characterize the challenges posed by different aspects of the game.

Hon Tik Tse, Siddarth Chandrasekar, Marlos C. Machado

In recent years, the successor representation (SR) has attracted increasing attention in reinforcement learning (RL), and it has been used to address some of its key challenges, such as exploration, credit assignment, and generalization. The SR can be seen as representing the underlying credit assignment structure of the environment by implicitly encoding its induced transition dynamics. However, the SR is reward-agnostic. In this paper, we discuss a similar representation that also takes into account the reward dynamics of the problem. We study the default representation (DR), a recently proposed representation with limited theoretical (and empirical) analysis. Here, we lay some of the theoretical foundation underlying the DR in the tabular case by (1) deriving dynamic programming and (2) temporal-difference methods to learn the DR, (3) characterizing the basis for the vector space of the DR, and (4) formally extending the DR to the function approximation case through default features. Empirically, we analyze the benefits of the DR in many of the settings in which the SR has been applied, including (1) reward shaping, (2) option discovery, (3) exploration, and (4) transfer learning. Our results show that, compared to the SR, the DR gives rise to qualitatively different, reward-aware behaviour and quantitatively better performance in several settings.

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.

Alex Lewandowski, Michał Bortkiewicz, Saurabh Kumar et al.

Loss of plasticity is a phenomenon where neural networks can become more difficult to train over the course of learning. Continual learning algorithms seek to mitigate this effect by sustaining good performance while maintaining network trainability. We develop a new technique for improving continual learning inspired by the observation that the singular values of the neural network parameters at initialization are an important factor for trainability during early phases of learning. From this perspective, we derive a new spectral regularizer for continual learning that better sustains these beneficial initialization properties throughout training. In particular, the regularizer keeps the maximum singular value of each layer close to one. Spectral regularization directly ensures that gradient diversity is maintained throughout training, which promotes continual trainability, while minimally interfering with performance in a single task. We present an experimental analysis that shows how the proposed spectral regularizer can sustain trainability and performance across a range of model architectures in continual supervised and reinforcement learning settings. Spectral regularization is less sensitive to hyperparameters while demonstrating better training in individual tasks, sustaining trainability as new tasks arrive, and achieving better generalization performance.

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.

Richard S. Sutton, Marlos C. Machado, G. Zacharias Holland et al.

To achieve the ambitious goals of artificial intelligence, reinforcement learning must include planning with a model of the world that is abstract in state and time. Deep learning has made progress with state abstraction, but temporal abstraction has rarely been used, despite extensively developed theory based on the options framework. One reason for this is that the space of possible options is immense, and the methods previously proposed for option discovery do not take into account how the option models will be used in planning. Options are typically discovered by posing subsidiary tasks, such as reaching a bottleneck state or maximizing the cumulative sum of a sensory signal other than reward. Each subtask is solved to produce an option, and then a model of the option is learned and made available to the planning process. In most previous work, the subtasks ignore the reward on the original problem, whereas we propose subtasks that use the original reward plus a bonus based on a feature of the state at the time the option terminates. We show that option models obtained from such reward-respecting subtasks are much more likely to be useful in planning than eigenoptions, shortest path options based on bottleneck states, or reward-respecting options generated by the option-critic. Reward respecting subtasks strongly constrain the space of options and thereby also provide a partial solution to the problem of option discovery. Finally, we show how values, policies, options, and models can all be learned online and off-policy using standard algorithms and general value functions.

Marlos C. Machado, Andre Barreto, Doina Precup et al.

Reasoning at multiple levels of temporal abstraction is one of the key attributes of intelligence. In reinforcement learning, this is often modeled through temporally extended courses of actions called options. Options allow agents to make predictions and to operate at different levels of abstraction within an environment. Nevertheless, approaches based on the options framework often start with the assumption that a reasonable set of options is known beforehand. When this is not the case, there are no definitive answers for which options one should consider. In this paper, we argue that the successor representation (SR), which encodes states based on the pattern of state visitation that follows them, can be seen as a natural substrate for the discovery and use of temporal abstractions. To support our claim, we take a big picture view of recent results, showing how the SR can be used to discover options that facilitate either temporally-extended exploration or planning. We cast these results as instantiations of a general framework for option discovery in which the agent's representation is used to identify useful options, which are then used to further improve its representation. This results in a virtuous, never-ending, cycle in which both the representation and the options are constantly refined based on each other. Beyond option discovery itself, we also discuss how the SR allows us to augment a set of options into a combinatorially large counterpart without additional learning. This is achieved through the combination of previously learned options. Our empirical evaluation focuses on options discovered for exploration and on the use of the SR to combine them. The results of our experiments shed light on important design decisions involved in the definition of options and demonstrate the synergy of different methods based on the SR, such as eigenoptions and the option keyboard.

Adrien Ali Taïga, William Fedus, Marlos C. Machado et al.

Research on exploration in reinforcement learning, as applied to Atari 2600 game-playing, has emphasized tackling difficult exploration problems such as Montezuma's Revenge (Bellemare et al., 2016). Recently, bonus-based exploration methods, which explore by augmenting the environment reward, have reached above-human average performance on such domains. In this paper we reassess popular bonus-based exploration methods within a common evaluation framework. We combine Rainbow (Hessel et al., 2018) with different exploration bonuses and evaluate its performance on Montezuma's Revenge, Bellemare et al.'s set of hard of exploration games with sparse rewards, and the whole Atari 2600 suite. We find that while exploration bonuses lead to higher score on Montezuma's Revenge they do not provide meaningful gains over the simpler $ε$-greedy scheme. In fact, we find that methods that perform best on that game often underperform $ε$-greedy on easy exploration Atari 2600 games. We find that our conclusions remain valid even when hyperparameters are tuned for these easy-exploration games. Finally, we find that none of the methods surveyed benefit from additional training samples (1 billion frames, versus Rainbow's 200 million) on Bellemare et al.'s hard exploration games. Our results suggest that recent gains in Montezuma's Revenge may be better attributed to architecture change, rather than better exploration schemes; and that the real pace of progress in exploration research for Atari 2600 games may have been obfuscated by good results on a single domain.

Sharan Vaswani, Olivier Bachem, Simone Totaro et al.

Common policy gradient methods rely on the maximization of a sequence of surrogate functions. In recent years, many such surrogate functions have been proposed, most without strong theoretical guarantees, leading to algorithms such as TRPO, PPO or MPO. Rather than design yet another surrogate function, we instead propose a general framework (FMA-PG) based on functional mirror ascent that gives rise to an entire family of surrogate functions. We construct surrogate functions that enable policy improvement guarantees, a property not shared by most existing surrogate functions. Crucially, these guarantees hold regardless of the choice of policy parameterization. Moreover, a particular instantiation of FMA-PG recovers important implementation heuristics (e.g., using forward vs reverse KL divergence) resulting in a variant of TRPO with additional desirable properties. Via experiments on simple bandit problems, we evaluate the algorithms instantiated by FMA-PG. The proposed framework also suggests an improved variant of PPO, whose robustness and efficiency we empirically demonstrate on the MuJoCo suite.

Rishabh Agarwal, Marlos C. Machado, Pablo Samuel Castro et al.

Reinforcement learning methods trained on few environments rarely learn policies that generalize to unseen environments. To improve generalization, we incorporate the inherent sequential structure in reinforcement learning into the representation learning process. This approach is orthogonal to recent approaches, which rarely exploit this structure explicitly. Specifically, we introduce a theoretically motivated policy similarity metric (PSM) for measuring behavioral similarity between states. PSM assigns high similarity to states for which the optimal policies in those states as well as in future states are similar. We also present a contrastive representation learning procedure to embed any state similarity metric, which we instantiate with PSM to obtain policy similarity embeddings (PSEs). We demonstrate that PSEs improve generalization on diverse benchmarks, including LQR with spurious correlations, a jumping task from pixels, and Distracting DM Control Suite.

Wesley Chung, Valentin Thomas, Marlos C. Machado et al.

Bandit and reinforcement learning (RL) problems can often be framed as optimization problems where the goal is to maximize average performance while having access only to stochastic estimates of the true gradient. Traditionally, stochastic optimization theory predicts that learning dynamics are governed by the curvature of the loss function and the noise of the gradient estimates. In this paper we demonstrate that this is not the case for bandit and RL problems. To allow our analysis to be interpreted in light of multi-step MDPs, we focus on techniques derived from stochastic optimization principles (e.g., natural policy gradient and EXP3) and we show that some standard assumptions from optimization theory are violated in these problems. We present theoretical results showing that, at least for bandit problems, curvature and noise are not sufficient to explain the learning dynamics and that seemingly innocuous choices like the baseline can determine whether an algorithm converges. These theoretical findings match our empirical evaluation, which we extend to multi-state MDPs.

Dibya Ghosh, Marlos C. Machado, Nicolas Le Roux

We cast policy gradient methods as the repeated application of two operators: a policy improvement operator $\mathcal{I}$, which maps any policy $π$ to a better one $\mathcal{I}π$, and a projection operator $\mathcal{P}$, which finds the best approximation of $\mathcal{I}π$ in the set of realizable policies. We use this framework to introduce operator-based versions of traditional policy gradient methods such as REINFORCE and PPO, which leads to a better understanding of their original counterparts. We also use the understanding we develop of the role of $\mathcal{I}$ and $\mathcal{P}$ to propose a new global lower bound of the expected return. This new perspective allows us to further bridge the gap between policy-based and value-based methods, showing how REINFORCE and the Bellman optimality operator, for example, can be seen as two sides of the same coin.

Adrien Ali Taïga, William Fedus, Marlos C. Machado et al.

This paper provides an empirical evaluation of recently developed exploration algorithms within the Arcade Learning Environment (ALE). We study the use of different reward bonuses that incentives exploration in reinforcement learning. We do so by fixing the learning algorithm used and focusing only on the impact of the different exploration bonuses in the agent's performance. We use Rainbow, the state-of-the-art algorithm for value-based agents, and focus on some of the bonuses proposed in the last few years. We consider the impact these algorithms have on performance within the popular game Montezuma's Revenge which has gathered a lot of interest from the exploration community, across the the set of seven games identified by Bellemare et al. (2016) as challenging for exploration, and easier games where exploration is not an issue. We find that, in our setting, recently developed bonuses do not provide significantly improved performance on Montezuma's Revenge or hard exploration games. We also find that existing bonus-based methods may negatively impact performance on games in which exploration is not an issue and may even perform worse than $ε$-greedy exploration.

Jesse Farebrother, Marlos C. Machado, Michael Bowling

Deep reinforcement learning algorithms have shown an impressive ability to learn complex control policies in high-dimensional tasks. However, despite the ever-increasing performance on popular benchmarks, policies learned by deep reinforcement learning algorithms can struggle to generalize when evaluated in remarkably similar environments. In this paper we propose a protocol to evaluate generalization in reinforcement learning through different modes of Atari 2600 games. With that protocol we assess the generalization capabilities of DQN, one of the most traditional deep reinforcement learning algorithms, and we provide evidence suggesting that DQN overspecializes to the training environment. We then comprehensively evaluate the impact of dropout and $\ell_2$ regularization, as well as the impact of reusing learned representations to improve the generalization capabilities of DQN. Despite regularization being largely underutilized in deep reinforcement learning, we show that it can, in fact, help DQN learn more general features. These features can be reused and fine-tuned on similar tasks, considerably improving DQN's sample efficiency.

Marlos C. Machado, Marc G. Bellemare, Michael Bowling

In this paper we introduce a simple approach for exploration in reinforcement learning (RL) that allows us to develop theoretically justified algorithms in the tabular case but that is also extendable to settings where function approximation is required. Our approach is based on the successor representation (SR), which was originally introduced as a representation defining state generalization by the similarity of successor states. Here we show that the norm of the SR, while it is being learned, can be used as a reward bonus to incentivize exploration. In order to better understand this transient behavior of the norm of the SR we introduce the substochastic successor representation (SSR) and we show that it implicitly counts the number of times each state (or feature) has been observed. We use this result to introduce an algorithm that performs as well as some theoretically sample-efficient approaches. Finally, we extend these ideas to a deep RL algorithm and show that it achieves state-of-the-art performance in Atari 2600 games when in a low sample-complexity regime.

Craig Sherstan, Marlos C. Machado, Patrick M. Pilarski

Here we propose using the successor representation (SR) to accelerate learning in a constructive knowledge system based on general value functions (GVFs). In real-world settings like robotics for unstructured and dynamic environments, it is infeasible to model all meaningful aspects of a system and its environment by hand due to both complexity and size. Instead, robots must be capable of learning and adapting to changes in their environment and task, incrementally constructing models from their own experience. GVFs, taken from the field of reinforcement learning (RL), are a way of modeling the world as predictive questions. One approach to such models proposes a massive network of interconnected and interdependent GVFs, which are incrementally added over time. It is reasonable to expect that new, incrementally added predictions can be learned more swiftly if the learning process leverages knowledge gained from past experience. The SR provides such a means of separating the dynamics of the world from the prediction targets and thus capturing regularities that can be reused across multiple GVFs. As a primary contribution of this work, we show that using SR-based predictions can improve sample efficiency and learning speed in a continual learning setting where new predictions are incrementally added and learned over time. We analyze our approach in a grid-world and then demonstrate its potential on data from a physical robot arm.

Miao Liu, Marlos C. Machado, Gerald Tesauro et al.

Eigenoptions (EOs) have been recently introduced as a promising idea for generating a diverse set of options through the graph Laplacian, having been shown to allow efficient exploration. Despite its initial promising results, a couple of issues in current algorithms limit its application, namely: (1) EO methods require two separate steps (eigenoption discovery and reward maximization) to learn a control policy, which can incur a significant amount of storage and computation; (2) EOs are only defined for problems with discrete state-spaces and; (3) it is not easy to take the environment's reward function into consideration when discovering EOs. To addresses these issues, we introduce an algorithm termed eigenoption-critic (EOC) based on the Option-critic (OC) framework [Bacon17], a general hierarchical reinforcement learning (RL) algorithm that allows learning the intra-option policies simultaneously with the policy over options. We also propose a generalization of EOC to problems with continuous state-spaces through the Nyström approximation. EOC can also be seen as extending OC to nonstationary settings, where the discovered options are not tailored for a single task.

Marlos C. Machado, Clemens Rosenbaum, Xiaoxiao Guo et al.

Options in reinforcement learning allow agents to hierarchically decompose a task into subtasks, having the potential to speed up learning and planning. However, autonomously learning effective sets of options is still a major challenge in the field. In this paper we focus on the recently introduced idea of using representation learning methods to guide the option discovery process. Specifically, we look at eigenoptions, options obtained from representations that encode diffusive information flow in the environment. We extend the existing algorithms for eigenoption discovery to settings with stochastic transitions and in which handcrafted features are not available. We propose an algorithm that discovers eigenoptions while learning non-linear state representations from raw pixels. It exploits recent successes in the deep reinforcement learning literature and the equivalence between proto-value functions and the successor representation. We use traditional tabular domains to provide intuition about our approach and Atari 2600 games to demonstrate its potential.

Marlos C. Machado, Marc G. Bellemare, Erik Talvitie et al.

The Arcade Learning Environment (ALE) is an evaluation platform that poses the challenge of building AI agents with general competency across dozens of Atari 2600 games. It supports a variety of different problem settings and it has been receiving increasing attention from the scientific community, leading to some high-profile success stories such as the much publicized Deep Q-Networks (DQN). In this article we take a big picture look at how the ALE is being used by the research community. We show how diverse the evaluation methodologies in the ALE have become with time, and highlight some key concerns when evaluating agents in the ALE. We use this discussion to present some methodological best practices and provide new benchmark results using these best practices. To further the progress in the field, we introduce a new version of the ALE that supports multiple game modes and provides a form of stochasticity we call sticky actions. We conclude this big picture look by revisiting challenges posed when the ALE was introduced, summarizing the state-of-the-art in various problems and highlighting problems that remain open.

Marlos C. Machado, Marc G. Bellemare, Michael Bowling

Representation learning and option discovery are two of the biggest challenges in reinforcement learning (RL). Proto-value functions (PVFs) are a well-known approach for representation learning in MDPs. In this paper we address the option discovery problem by showing how PVFs implicitly define options. We do it by introducing eigenpurposes, intrinsic reward functions derived from the learned representations. The options discovered from eigenpurposes traverse the principal directions of the state space. They are useful for multiple tasks because they are discovered without taking the environment's rewards into consideration. Moreover, different options act at different time scales, making them helpful for exploration. We demonstrate features of eigenpurposes in traditional tabular domains as well as in Atari 2600 games.

Craig Sherstan, Adam White, Marlos C. Machado et al.

Agents of general intelligence deployed in real-world scenarios must adapt to ever-changing environmental conditions. While such adaptive agents may leverage engineered knowledge, they will require the capacity to construct and evaluate knowledge themselves from their own experience in a bottom-up, constructivist fashion. This position paper builds on the idea of encoding knowledge as temporally extended predictions through the use of general value functions. Prior work has focused on learning predictions about externally derived signals about a task or environment (e.g. battery level, joint position). Here we advocate that the agent should also predict internally generated signals regarding its own learning process - for example, an agent's confidence in its learned predictions. Finally, we suggest how such information would be beneficial in creating an introspective agent that is able to learn to make good decisions in a complex, changing world.

Marlos C. Machado, Michael Bowling

Artificial intelligence is commonly defined as the ability to achieve goals in the world. In the reinforcement learning framework, goals are encoded as reward functions that guide agent behaviour, and the sum of observed rewards provide a notion of progress. However, some domains have no such reward signal, or have a reward signal so sparse as to appear absent. Without reward feedback, agent behaviour is typically random, often dithering aimlessly and lacking intentionality. In this paper we present an algorithm capable of learning purposeful behaviour in the absence of rewards. The algorithm proceeds by constructing temporally extended actions (options), through the identification of purposes that are "just out of reach" of the agent's current behaviour. These purposes establish intrinsic goals for the agent to learn, ultimately resulting in a suite of behaviours that encourage the agent to visit different parts of the state space. Moreover, the approach is particularly suited for settings where rewards are very sparse, and such behaviours can help in the exploration of the environment until reward is observed.

Harm van Seijen, A. Rupam Mahmood, Patrick M. Pilarski et al.

The temporal-difference methods TD($λ$) and Sarsa($λ$) form a core part of modern reinforcement learning. Their appeal comes from their good performance, low computational cost, and their simple interpretation, given by their forward view. Recently, new versions of these methods were introduced, called true online TD($λ$) and true online Sarsa($λ$), respectively (van Seijen & Sutton, 2014). These new versions maintain an exact equivalence with the forward view at all times, whereas the traditional versions only approximate it for small step-sizes. We hypothesize that these true online methods not only have better theoretical properties, but also dominate the regular methods empirically. In this article, we put this hypothesis to the test by performing an extensive empirical comparison. Specifically, we compare the performance of true online TD($λ$)/Sarsa($λ$) with regular TD($λ$)/Sarsa($λ$) on random MRPs, a real-world myoelectric prosthetic arm, and a domain from the Arcade Learning Environment. We use linear function approximation with tabular, binary, and non-binary features. Our results suggest that the true online methods indeed dominate the regular methods. Across all domains/representations the learning speed of the true online methods are often better, but never worse than that of the regular methods. An additional advantage is that no choice between traces has to be made for the true online methods. Besides the empirical results, we provide an in-depth analysis of the theory behind true online temporal-difference learning. In addition, we show that new true online temporal-difference methods can be derived by making changes to the online forward view and then rewriting the update equations.

Yitao Liang, Marlos C. Machado, Erik Talvitie et al.

The recently introduced Deep Q-Networks (DQN) algorithm has gained attention as one of the first successful combinations of deep neural networks and reinforcement learning. Its promise was demonstrated in the Arcade Learning Environment (ALE), a challenging framework composed of dozens of Atari 2600 games used to evaluate general competency in AI. It achieved dramatically better results than earlier approaches, showing that its ability to learn good representations is quite robust and general. This paper attempts to understand the principles that underlie DQN's impressive performance and to better contextualize its success. We systematically evaluate the importance of key representational biases encoded by DQN's network by proposing simple linear representations that make use of these concepts. Incorporating these characteristics, we obtain a computationally practical feature set that achieves competitive performance to DQN in the ALE. Besides offering insight into the strengths and weaknesses of DQN, we provide a generic representation for the ALE, significantly reducing the burden of learning a representation for each game. Moreover, we also provide a simple, reproducible benchmark for the sake of comparison to future work in the ALE.

Marlos C. Machado, Sriram Srinivasan, Michael Bowling

In Reinforcement Learning (RL), it is common to use optimistic initialization of value functions to encourage exploration. However, such an approach generally depends on the domain, viz., the scale of the rewards must be known, and the feature representation must have a constant norm. We present a simple approach that performs optimistic initialization with less dependence on the domain.

Marlos C. Machado

AI is gradually receiving more attention as a fundamental feature to increase the immersion in digital games. Among the several AI approaches, player modeling is becoming an important one. The main idea is to understand and model the player characteristics and behaviors in order to develop a better AI. In this work, we discuss several aspects of this new field. We proposed a taxonomy to organize the area, discussing several facets of this topic, ranging from implementation decisions up to what a model attempts to describe. We then classify, in our taxonomy, some of the most important works in this field. We also presented a generic approach to deal with player modeling using ML, and we instantiated this approach to model players' preferences in the game Civilization IV. The instantiation of this approach has several steps. We first discuss a generic representation, regardless of what is being modeled, and evaluate it performing experiments with the strategy game Civilization IV. Continuing the instantiation of the proposed approach we evaluated the applicability of using game score information to distinguish different preferences. We presented a characterization of virtual agents in the game, comparing their behavior with their stated preferences. Once we have characterized these agents, we were able to observe that different preferences generate different behaviors, measured by several game indicators. We then tackled the preference modeling problem as a binary classification task, with a supervised learning approach. We compared four different methods, based on different paradigms (SVM, AdaBoost, NaiveBayes and JRip), evaluating them on a set of matches played by different virtual agents. We conclude our work using the learned models to infer human players' preferences. Using some of the evaluated classifiers we obtained accuracies over 60% for most of the inferred preferences.