Action-value estimation is a critical component of many reinforcement learning (RL) methods whereby sample complexity relies heavily on how fast a good estimator for action value can be learned. By viewing this problem through the lens of representation learning, good representations of both state and action can facilitate action-value estimation. While advances in deep learning have seamlessly driven progress in learning state representations, given the specificity of the notion of agency to RL, little attention has been paid to learning action representations. We conjecture that leveraging the combinatorial structure of multi-dimensional action spaces is a key ingredient for learning good representations of action. To test this, we set forth the action hypergraph networks framework -- a class of functions for learning action representations in multi-dimensional discrete action spaces with a structural inductive bias. Using this framework we realise an agent class based on a combination with deep Q-networks, which we dub hypergraph Q-networks. We show the effectiveness of our approach on a myriad of domains: illustrative prediction problems under minimal confounding effects, Atari 2600 games, and discretised physical control benchmarks.
Reinforcement learning algorithms require many samples when solving complex hierarchical tasks with sparse and delayed rewards. For such complex tasks, the recently proposed RUDDER uses reward redistribution to leverage steps in the Q-function that are associated with accomplishing sub-tasks. However, often only few episodes with high rewards are available as demonstrations since current exploration strategies cannot discover them in reasonable time. In this work, we introduce Align-RUDDER, which utilizes a profile model for reward redistribution that is obtained from multiple sequence alignment of demonstrations. Consequently, Align-RUDDER employs reward redistribution effectively and, thereby, drastically improves learning on few demonstrations. Align-RUDDER outperforms competitors on complex artificial tasks with delayed rewards and few demonstrations. On the Minecraft ObtainDiamond task, Align-RUDDER is able to mine a diamond, though not frequently. Code is available at https://github.com/ml-jku/align-rudder. YouTube: https://youtu.be/HO-_8ZUl-UY
In this paper, we present a novel developmental reinforcement learning-based controller for a quadcopter with thrust vectoring capabilities. This multirotor UAV design has tilt-enabled rotors. It utilizes the rotor force magnitude and direction to achieve the desired state during flight. The control policy of this robot is learned using the policy transfer from the learned controller of the quadcopter (comparatively simple UAV design without thrust vectoring). This approach allows learning a control policy for systems with multiple inputs and multiple outputs. The performance of the learned policy is evaluated by physics-based simulations for the tasks of hovering and way-point navigation. The flight simulations utilize a flight controller based on reinforcement learning without any additional PID components. The results show faster learning with the presented approach as opposed to learning the control policy from scratch for this new UAV design created by modifications in a conventional quadcopter, i.e., the addition of more degrees of freedom (4-actuators in conventional quadcopter to 8-actuators in tilt-rotor quadcopter). We demonstrate the robustness of our learned policy by showing the recovery of the tilt-rotor platform in the simulation from various non-static initial conditions in order to reach a desired state. The developmental policy for the tilt-rotor UAV also showed superior fault tolerance when compared with the policy learned from the scratch. The results show the ability of the presented approach to bootstrap the learned behavior from a simpler system (lower-dimensional action-space) to a more complex robot (comparatively higher-dimensional action-space) and reach better performance faster.
While deep reinforcement learning excels at solving tasks where large amounts of data can be collected through virtually unlimited interaction with the environment, learning from limited interaction remains a key challenge. We posit that an agent can learn more efficiently if we augment reward maximization with self-supervised objectives based on structure in its visual input and sequential interaction with the environment. Our method, Self-Predictive Representations(SPR), trains an agent to predict its own latent state representations multiple steps into the future. We compute target representations for future states using an encoder which is an exponential moving average of the agent's parameters and we make predictions using a learned transition model. On its own, this future prediction objective outperforms prior methods for sample-efficient deep RL from pixels. We further improve performance by adding data augmentation to the future prediction loss, which forces the agent's representations to be consistent across multiple views of an observation. Our full self-supervised objective, which combines future prediction and data augmentation, achieves a median human-normalized score of 0.415 on Atari in a setting limited to 100k steps of environment interaction, which represents a 55% relative improvement over the previous state-of-the-art. Notably, even in this limited data regime, SPR exceeds expert human scores on 7 out of 26 games. The code associated with this work is available at https://github.com/mila-iqia/spr
Deep model-based Reinforcement Learning (RL) has the potential to substantially improve the sample-efficiency of deep RL. While various challenges have long held it back, a number of papers have recently come out reporting success with deep model-based methods. This is a great development, but the lack of a consistent metric to evaluate such methods makes it difficult to compare various approaches. For example, the common single-task sample-efficiency metric conflates improvements due to model-based learning with various other aspects, such as representation learning, making it difficult to assess true progress on model-based RL. To address this, we introduce an experimental setup to evaluate model-based behavior of RL methods, inspired by work from neuroscience on detecting model-based behavior in humans and animals. Our metric based on this setup, the Local Change Adaptation (LoCA) regret, measures how quickly an RL method adapts to a local change in the environment. Our metric can identify model-based behavior, even if the method uses a poor representation and provides insight in how close a method's behavior is from optimal model-based behavior. We use our setup to evaluate the model-based behavior of MuZero on a variation of the classic Mountain Car task.
Many dynamic processes, including common scenarios in robotic control and reinforcement learning (RL), involve a set of interacting subprocesses. Though the subprocesses are not independent, their interactions are often sparse, and the dynamics at any given time step can often be decomposed into locally independent causal mechanisms. Such local causal structures can be leveraged to improve the sample efficiency of sequence prediction and off-policy reinforcement learning. We formalize this by introducing local causal models (LCMs), which are induced from a global causal model by conditioning on a subset of the state space. We propose an approach to inferring these structures given an object-oriented state representation, as well as a novel algorithm for Counterfactual Data Augmentation (CoDA). CoDA uses local structures and an experience replay to generate counterfactual experiences that are causally valid in the global model. We find that CoDA significantly improves the performance of RL agents in locally factored tasks, including the batch-constrained and goal-conditioned settings.
Increasing the scale of reinforcement learning experiments has allowed researchers to achieve unprecedented results in both training sophisticated agents for video games, and in sim-to-real transfer for robotics. Typically such experiments rely on large distributed systems and require expensive hardware setups, limiting wider access to this exciting area of research. In this work we aim to solve this problem by optimizing the efficiency and resource utilization of reinforcement learning algorithms instead of relying on distributed computation. We present the "Sample Factory", a high-throughput training system optimized for a single-machine setting. Our architecture combines a highly efficient, asynchronous, GPU-based sampler with off-policy correction techniques, allowing us to achieve throughput higher than $10^5$ environment frames/second on non-trivial control problems in 3D without sacrificing sample efficiency. We extend Sample Factory to support self-play and population-based training and apply these techniques to train highly capable agents for a multiplayer first-person shooter game. The source code is available at https://github.com/alex-petrenko/sample-factory
The challenge of developing powerful and general Reinforcement Learning (RL) agents has received increasing attention in recent years. Much of this effort has focused on the single-agent setting, in which an agent maximizes a predefined extrinsic reward function. However, a long-term question inevitably arises: how will such independent agents cooperate when they are continually learning and acting in a shared multi-agent environment? Observing that humans often provide incentives to influence others' behavior, we propose to equip each RL agent in a multi-agent environment with the ability to give rewards directly to other agents, using a learned incentive function. Each agent learns its own incentive function by explicitly accounting for its impact on the learning of recipients and, through them, the impact on its own extrinsic objective. We demonstrate in experiments that such agents significantly outperform standard RL and opponent-shaping agents in challenging general-sum Markov games, often by finding a near-optimal division of labor. Our work points toward more opportunities and challenges along the path to ensure the common good in a multi-agent future.
With advances in reinforcement learning (RL), agents are now being developed in high-stakes application domains such as healthcare and transportation. Explaining the behavior of these agents is challenging, as the environments in which they act have large state spaces, and their decision-making can be affected by delayed rewards, making it difficult to analyze their behavior. To address this problem, several approaches have been developed. Some approaches attempt to convey the $\textit{global}$ behavior of the agent, describing the actions it takes in different states. Other approaches devised $\textit{local}$ explanations which provide information regarding the agent's decision-making in a particular state. In this paper, we combine global and local explanation methods, and evaluate their joint and separate contributions, providing (to the best of our knowledge) the first user study of combined local and global explanations for RL agents. Specifically, we augment strategy summaries that extract important trajectories of states from simulations of the agent with saliency maps which show what information the agent attends to. Our results show that the choice of what states to include in the summary (global information) strongly affects people's understanding of agents: participants shown summaries that included important states significantly outperformed participants who were presented with agent behavior in a randomly set of chosen world-states. We find mixed results with respect to augmenting demonstrations with saliency maps (local information), as the addition of saliency maps did not significantly improve performance in most cases. However, we do find some evidence that saliency maps can help users better understand what information the agent relies on in its decision making, suggesting avenues for future work that can further improve explanations of RL agents.
A significant challenge in developing AI that can generalize well is designing agents that learn about their world without being told what to learn, and apply that learning to challenges with sparse rewards. Moreover, most traditional reinforcement learning approaches explicitly separate learning and decision making in a way that does not correspond to biological learning. We implement an architecture founded in principles of experimental neuroscience, by combining computationally efficient abstractions of biological algorithms. Our approach is inspired by research on spike-timing dependent plasticity, the transition between short and long term memory, and the role of various neurotransmitters in rewarding curiosity. The Neurons-in-a-Box architecture can learn in a wholly generalizable manner, and demonstrates an efficient way to build and apply representations without explicitly optimizing over a set of criteria or actions. We find it performs well in many environments including OpenAI Gym's Mountain Car, which has no reward besides touching a hard-to-reach flag on a hill, Inverted Pendulum, where it learns simple strategies to improve the time it holds a pendulum up, a video stream, where it spontaneously learns to distinguish an open and closed hand, as well as other environments like Google Chrome's Dinosaur Game.
Transfer learning is an important new subfield of multiagent reinforcement learning that aims to help an agent learn about a problem by using knowledge that it has gained solving another problem, or by using knowledge that is communicated to it by an agent who already knows the problem. This is useful when one wishes to change the architecture or learning algorithm of an agent (so that the new knowledge need not be built "from scratch"), when new agents are frequently introduced to the environment with no knowledge, or when an agent must adapt to similar but different problems. Great progress has been made in the agent-to-agent case using the Teacher/Student framework proposed by (Torrey and Taylor 2013). However, that approach requires that learning from a teacher be treated differently from learning in every other reinforcement learning context. In this paper, I propose a method which allows the teacher/student framework to be applied in a way that fits directly and naturally into the more general reinforcement learning framework by integrating the teacher feedback into the reward signal received by the learning agent. I show that this approach can significantly improve the rate of learning for an agent playing a one-player stochastic game; I give examples of potential pitfalls of the approach; and I propose further areas of research building on this framework.
We study infinite horizon discounted Mean Field Control (MFC) problems with common noise through the lens of Mean Field Markov Decision Processes (MFMDP). We allow the agents to use actions that are randomized not only at the individual level but also at the level of the population. This common randomization allows us to establish connections between both closed-loop and open-loop policies for MFC and Markov policies for the MFMDP. In particular, we show that there exists an optimal closed-loop policy for the original MFC. Building on this framework and the notion of state-action value function, we then propose reinforcement learning (RL) methods for such problems, by adapting existing tabular and deep RL methods to the mean-field setting. The main difficulty is the treatment of the population state, which is an input of the policy and the value function. We provide convergence guarantees for tabular algorithms based on discretizations of the simplex. Neural network based algorithms are more suitable for continuous spaces and allow us to avoid discretizing the mean field state space. Numerical examples are provided.
Owing to their ability to both effectively integrate information over long time horizons and scale to massive amounts of data, self-attention architectures have recently shown breakthrough success in natural language processing (NLP), achieving state-of-the-art results in domains such as language modeling and machine translation. Harnessing the transformer's ability to process long time horizons of information could provide a similar performance boost in partially observable reinforcement learning (RL) domains, but the large-scale transformers used in NLP have yet to be successfully applied to the RL setting. In this work we demonstrate that the standard transformer architecture is difficult to optimize, which was previously observed in the supervised learning setting but becomes especially pronounced with RL objectives. We propose architectural modifications that substantially improve the stability and learning speed of the original Transformer and XL variant. The proposed architecture, the Gated Transformer-XL (GTrXL), surpasses LSTMs on challenging memory environments and achieves state-of-the-art results on the multi-task DMLab-30 benchmark suite, exceeding the performance of an external memory architecture. We show that the GTrXL, trained using the same losses, has stability and performance that consistently matches or exceeds a competitive LSTM baseline, including on more reactive tasks where memory is less critical. GTrXL offers an easy-to-train, simple-to-implement but substantially more expressive architectural alternative to the standard multi-layer LSTM ubiquitously used for RL agents in partially observable environments.
This paper introduces a principled approach for the design of a scalable general reinforcement learning agent. This approach is based on a direct approximation of AIXI, a Bayesian optimality notion for general reinforcement learning agents. Previously, it has been unclear whether the theory of AIXI could motivate the design of practical algorithms. We answer this hitherto open question in the affirmative, by providing the first computationally feasible approximation to the AIXI agent. To develop our approximation, we introduce a Monte Carlo Tree Search algorithm along with an agent-specific extension of the Context Tree Weighting algorithm. Empirically, we present a set of encouraging results on a number of stochastic, unknown, and partially observable domains.
OpenSpiel is a collection of environments and algorithms for research in general reinforcement learning and search/planning in games. OpenSpiel supports n-player (single- and multi- agent) zero-sum, cooperative and general-sum, one-shot and sequential, strictly turn-taking and simultaneous-move, perfect and imperfect information games, as well as traditional multiagent environments such as (partially- and fully- observable) grid worlds and social dilemmas. OpenSpiel also includes tools to analyze learning dynamics and other common evaluation metrics. This document serves both as an overview of the code base and an introduction to the terminology, core concepts, and algorithms across the fields of reinforcement learning, computational game theory, and search.
Consistent and reproducible evaluation of Deep Reinforcement Learning (DRL) is not straightforward. In the Arcade Learning Environment (ALE), small changes in environment parameters such as stochasticity or the maximum allowed play time can lead to very different performance. In this work, we discuss the difficulties of comparing different agents trained on ALE. In order to take a step further towards reproducible and comparable DRL, we introduce SABER, a Standardized Atari BEnchmark for general Reinforcement learning algorithms. Our methodology extends previous recommendations and contains a complete set of environment parameters as well as train and test procedures. We then use SABER to evaluate the current state of the art, Rainbow. Furthermore, we introduce a human world records baseline, and argue that previous claims of expert or superhuman performance of DRL might not be accurate. Finally, we propose Rainbow-IQN by extending Rainbow with Implicit Quantile Networks (IQN) leading to new state-of-the-art performance. Source code is available for reproducibility.
Automatic data abstraction is an important capability for both benchmarking machine intelligence and supporting summarization applications. In the former one asks whether a machine can `understand' enough about the meaning of input data to produce a meaningful but more compact abstraction. In the latter this capability is exploited for saving space or human time by summarizing the essence of input data. In this paper we study a general reinforcement learning based framework for learning to abstract sequential data in a goal-driven way. The ability to define different abstraction goals uniquely allows different aspects of the input data to be preserved according to the ultimate purpose of the abstraction. Our reinforcement learning objective does not require human-defined examples of ideal abstraction. Importantly our model processes the input sequence holistically without being constrained by the original input order. Our framework is also domain agnostic -- we demonstrate applications to sketch, video and text data and achieve promising results in all domains.
The successful application of general reinforcement learning algorithms to real-world robotics applications is often limited by their high data requirements. We introduce Regularized Hierarchical Policy Optimization (RHPO) to improve data-efficiency for domains with multiple dominant tasks and ultimately reduce required platform time. To this end, we employ compositional inductive biases on multiple levels and corresponding mechanisms for sharing off-policy transition data across low-level controllers and tasks as well as scheduling of tasks. The presented algorithm enables stable and fast learning for complex, real-world domains in the parallel multitask and sequential transfer case. We show that the investigated types of hierarchy enable positive transfer while partially mitigating negative interference and evaluate the benefits of additional incentives for efficient, compositional task solutions in single task domains. Finally, we demonstrate substantial data-efficiency and final performance gains over competitive baselines in a week-long, physical robot stacking experiment.
In an effort to better understand the different ways in which the discount factor affects the optimization process in reinforcement learning, we designed a set of experiments to study each effect in isolation. Our analysis reveals that the common perception that poor performance of low discount factors is caused by (too) small action-gaps requires revision. We propose an alternative hypothesis that identifies the size-difference of the action-gap across the state-space as the primary cause. We then introduce a new method that enables more homogeneous action-gaps by mapping value estimates to a logarithmic space. We prove convergence for this method under standard assumptions and demonstrate empirically that it indeed enables lower discount factors for approximate reinforcement-learning methods. This in turn allows tackling a class of reinforcement-learning problems that are challenging to solve with traditional methods.
We consider undiscounted reinforcement learning in Markov decision processes (MDPs) where both the reward functions and the state-transition probabilities may vary (gradually or abruptly) over time. For this problem setting, we propose an algorithm and provide performance guarantees for the regret evaluated against the optimal non-stationary policy. The upper bound on the regret is given in terms of the total variation in the MDP. This is the first variational regret bound for the general reinforcement learning setting.