SBEED: Convergent Reinforcement Learning with Nonlinear Function Approximation
Google Inc.（谷歌公司）；Microsoft Research, Redmond（微软研究院（雷德蒙德））；University of Illinois at Urbana Champaign（伊利诺伊大学厄巴纳-香槟分校）；Tencent AI Lab, Bellevue（腾讯人工智能实验室（贝尔维尤））
Bo Dai, Albert Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, Le Song
备注：
28 pages, 13 figures. To appear at the 35th International Conference on Machine Learning (ICML 2018)
When function approximation is used, solving the Bellman optimality equation with stability guarantees has remained a major open problem in reinforcement learning for decades. The fundamental difficulty is that the Bellman operator may become an expansion in general, resulting in oscillating and even divergent behavior of popular algorithms like Q-learning. In this paper, we revisit the Bellman equation, and reformulate it into a novel primal-dual optimization problem using Nesterov's smoothing technique and the Legendre-Fenchel transformation. We then develop a new algorithm, called Smoothed Bellman Error Embedding, to solve this optimization problem where any differentiable function class may be used. We provide what we believe to be the first convergence guarantee for general nonlinear function approximation, and analyze the algorithm's sample complexity. Empirically, our algorithm compares favorably to state-of-the-art baselines in several benchmark control problems.
The \textit{Smoothed Bellman Error Embedding} algorithm~\citep{dai2018sbeed}, known as SBEED, was proposed as a provably convergent reinforcement learning algorithm with general nonlinear function approximation. It has been successfully implemented with neural networks and achieved strong empirical results. In this work, we study the theoretical behavior of SBEED in batch-mode reinforcement learning. We prove a near-optimal performance guarantee that depends on the representation power of the used function classes and a tight notion of the distribution shift. Our results improve upon prior gua…
The University of Chicago（芝加哥大学）；Princeton University（普林斯顿大学）；Northwestern University（西北大学）
We study the safe reinforcement learning problem with nonlinear function approximation, where policy optimization is formulated as a constrained optimization problem with both the objective and the constraint being nonconvex functions. For such a problem, we construct a sequence of surrogate convex constrained optimization problems by replacing the nonconvex functions locally with convex quadratic functions obtained from policy gradient estimators. We prove that the solutions to these surrogate problems converge to a stationary point of the original nonconvex problem. Furthermore, to extend ou…
Reinforcement learning is commonly used with function approximation. However, very few positive results are known about the convergence of function approximation based RL control algorithms. In this paper we show that TD(0) and Sarsa(0) with linear function approximation is convergent for a simple class of problems, where the system is linear and the costs are quadratic (the LQ control problem). Furthermore, we show that for systems with Gaussian noise and non-completely observable states (the LQG problem), the mentioned RL algorithms are still convergent, if they are combined with Kalman filt…
With the development of experimental quantum technology, quantum control has attracted increasing attention due to the realization of controllable artificial quantum systems. However, because quantum-mechanical systems are often too difficult to analytically deal with, heuristic strategies and numerical algorithms which search for proper control protocols are adopted, and, deep learning, especially deep reinforcement learning (RL), is a promising generic candidate solution for the control problems. Although there have been a few successful applications of deep RL to quantum control problems, m…
In Reinforcement Learning the Q-learning algorithm provably converges to the optimal solution. However, as others have demonstrated, Q-learning can also overestimate the values and thereby spend too long exploring unhelpful states. Double Q-learning is a provably convergent alternative that mitigates some of the overestimation issues, though sometimes at the expense of slower convergence. We introduce an alternative algorithm that replaces the max operation with an average, resulting also in a provably convergent off-policy algorithm which can mitigate overestimation yet retain similar converg…
Polytechnique Montréal（蒙特利尔综合理工学院）；University of Basel（巴塞尔大学）；Johannes Kepler University Linz（林茨约翰·开普勒大学）
Evolutionary strategies have recently been shown to achieve competing levels of performance for complex optimization problems in reinforcement learning. In such problems, one often needs to optimize an objective function subject to a set of constraints, including for instance constraints on the entropy of a policy or to restrict the possible set of actions or states accessible to an agent. Convergence guarantees for evolutionary strategies to optimize stochastic constrained problems are however lacking in the literature. In this work, we address this problem by designing a novel optimization a…
[Zhang, ICML 2018] provided the first decentralized actor-critic algorithm for multi-agent reinforcement learning (MARL) that offers convergence guarantees. In that work, policies are stochastic and are defined on finite action spaces. We extend those results to offer a provably-convergent decentralized actor-critic algorithm for learning deterministic policies on continuous action spaces. Deterministic policies are important in real-world settings. To handle the lack of exploration inherent in deterministic policies, we consider both off-policy and on-policy settings. We provide the expressio…
Facebook AI Research（脸书人工智能研究院）；MILA, Université de Montréal（蒙特利尔大学米拉）；MILA, McGill University（麦吉尔大学米拉）；Canadian Institute for Advanced Research (CIFAR)（加拿大高级研究院）
Off-policy learning is key to scaling up reinforcement learning as it allows to learn about a target policy from the experience generated by a different behavior policy. Unfortunately, it has been challenging to combine off-policy learning with function approximation and multi-step bootstrapping in a way that leads to both stable and efficient algorithms. In this work, we show that the \textsc{Tree Backup} and \textsc{Retrace} algorithms are unstable with linear function approximation, both in theory and in practice with specific examples. Based on our analysis, we then derive stable and effic…
University of Tokyo（东京大学）；RIKEN Center for Emergent Matter Science（理化学研究所 emergent matter科学中心）
Despite the empirical success of the deep Q network (DQN) reinforcement learning algorithm and its variants, DQN is still not well understood and it does not guarantee convergence. In this work, we show that DQN can indeed diverge and cease to operate in realistic settings. Although there exist gradient-based convergent methods, we show that they actually have inherent problems in learning dynamics which cause them to fail even in simple tasks. To overcome these problems, we propose a convergent DQN algorithm (C-DQN) that is guaranteed to converge and can work with large discount factors (0.99…
Learning the value function of a given policy (target policy) from the data samples obtained from a different policy (behavior policy) is an important problem in Reinforcement Learning (RL). This problem is studied under the setting of off-policy prediction. Temporal Difference (TD) learning algorithms are a popular class of algorithms for solving the prediction problem. TD algorithms with linear function approximation are shown to be convergent when the samples are generated from the target policy (known as on-policy prediction). However, it has been well established in the literature that of…
Actor-Critic based approaches were among the first to address reinforcement learning in a general setting. Recently, these algorithms have gained renewed interest due to their generality, good convergence properties, and possible biological relevance. In this paper, we introduce an online temporal difference based actor-critic algorithm which is proved to converge to a neighborhood of a local maximum of the average reward. Linear function approximation is used by the critic in order estimate the value function, and the temporal difference signal, which is passed from the critic to the actor. T…
A softmax operator applied to a set of values acts somewhat like the maximization function and somewhat like an average. In sequential decision making, softmax is often used in settings where it is necessary to maximize utility but also to hedge against problems that arise from putting all of one's weight behind a single maximum utility decision. The Boltzmann softmax operator is the most commonly used softmax operator in this setting, but we show that this operator is prone to misbehavior. In this work, we study a differentiable softmax operator that, among other properties, is a non-expansio…
Microsoft Research（微软研究院）；Max Planck Institute for Intelligent Systems（马克斯·普朗克智能系统研究所）
We present a general convergent class of reinforcement learning algorithms that is founded on two distinct principles: (1) mapping value estimates to a different space using arbitrary functions from a broad class, and (2) linearly decomposing the reward signal into multiple channels. The first principle enables incorporating specific properties into the value estimator that can enhance learning. The second principle, on the other hand, allows for the value function to be represented as a composition of multiple utility functions. This can be leveraged for various purposes, e.g. dealing with hi…
Hajim School of Engineering and Applied Sciences, University of Rochester（罗切斯特大学哈吉姆工程与应用科学学院）
In this paper, we set forth a new vision of reinforcement learning developed by us over the past few years, one that yields mathematically rigorous solutions to longstanding important questions that have remained unresolved: (i) how to design reliable, convergent, and robust reinforcement learning algorithms (ii) how to guarantee that reinforcement learning satisfies pre-specified "safety" guarantees, and remains in a stable region of the parameter space (iii) how to design "off-policy" temporal difference learning algorithms in a reliable and stable manner, and finally (iv) how to integrate t…
Khoury College of Computer Sciences（库里计算机科学学院）；Northeastern University（东北大学）
Off-policy learning from multistep returns is crucial for sample-efficient reinforcement learning, particularly in the experience replay setting now commonly used with deep neural networks. Classically, off-policy estimation bias is corrected in a per-decision manner: past temporal-difference errors are re-weighted by the instantaneous Importance Sampling (IS) ratio (via eligibility traces) after each action. Many important off-policy algorithms such as Tree Backup and Retrace rely on this mechanism along with differing protocols for truncating ("cutting") the ratios ("traces") to counteract t…
University of Alberta（阿尔伯塔大学）；Amii（阿尔伯塔机器智能研究所）；DeepMind
We extend the options framework for temporal abstraction in reinforcement learning from discounted Markov decision processes (MDPs) to average-reward MDPs. Our contributions include general convergent off-policy inter-option learning algorithms, intra-option algorithms for learning values and models, as well as sample-based planning variants of our learning algorithms. Our algorithms and convergence proofs extend those recently developed by Wan, Naik, and Sutton. We also extend the notion of option-interrupting behavior from the discounted to the average-reward formulation. We show the efficac…
Huawei R&D UK（华为英国研发中心）；Imperial College London（帝国理工学院）；University College London（伦敦大学学院）；University of Cambridge（剑桥大学）
Reinforcement learning algorithms, though successful, tend to over-fit to training environments hampering their application to the real-world. This paper proposes $\text{W}\text{R}^{2}\text{L}$ -- a robust reinforcement learning algorithm with significant robust performance on low and high-dimensional control tasks. Our method formalises robust reinforcement learning as a novel min-max game with a Wasserstein constraint for a correct and convergent solver. Apart from the formulation, we also propose an efficient and scalable solver following a novel zero-order optimisation method that we belie…
Recurrent neural networks are often used for learning time-series data. Based on a few assumptions we model this learning task as a minimization problem of a nonlinear least-squares cost function. The special structure of the cost function allows us to build a connection to reinforcement learning. We exploit this connection and derive a convergent, policy iteration-based algorithm. Furthermore, we argue that RNN training can be fit naturally into the reinforcement learning framework.
We propose a reinforcement learning algorithm for stationary mean-field games, where the goal is to learn a pair of mean-field state and stationary policy that constitutes the Nash equilibrium. When viewing the mean-field state and the policy as two players, we propose a fictitious play algorithm which alternatively updates the mean-field state and the policy via gradient-descent and proximal policy optimization, respectively. Our algorithm is in stark contrast with previous literature which solves each single-agent reinforcement learning problem induced by the iterates mean-field states to th…
In this paper, we aim to develop a simple and scalable reinforcement learning algorithm that uses standard supervised learning methods as subroutines. Our goal is an algorithm that utilizes only simple and convergent maximum likelihood loss functions, while also being able to leverage off-policy data. Our proposed approach, which we refer to as advantage-weighted regression (AWR), consists of two standard supervised learning steps: one to regress onto target values for a value function, and another to regress onto weighted target actions for the policy. The method is simple and general, can ac…
Gradient temporal difference (Gradient TD) algorithms are a popular class of stochastic approximation (SA) algorithms used for policy evaluation in reinforcement learning. Here, we consider Gradient TD algorithms with an additional heavy ball momentum term and provide choice of step size and momentum parameter that ensures almost sure convergence of these algorithms asymptotically. In doing so, we decompose the heavy ball Gradient TD iterates into three separate iterates with different step sizes. We first analyze these iterates under one-timescale SA setting using results from current literat…
Foundations of Multivariate Distributional Reinforcement Learning
Mila — Québec AI Institute（米拉-魁北克人工智能研究所）；McGill University（麦吉尔大学）；Google DeepMind（谷歌DeepMind）；Gatsby Unit, University College London（伦敦大学学院盖茨比计算神经科学单元）
In this paper we revisit the method of off-policy corrections for reinforcement learning (COP-TD) pioneered by Hallak et al. (2017). Under this method, online updates to the value function are reweighted to avoid divergence issues typical of off-policy learning. While Hallak et al.'s solution is appealing, it cannot easily be transferred to nonlinear function approximation. First, it requires a projection step onto the probability simplex; second, even though the operator describing the expected behavior of the off-policy learning algorithm is convergent, it is not known to be a contraction ma…
To widen their accessibility and increase their utility, intelligent agents must be able to learn complex behaviors as specified by (non-expert) human users. Moreover, they will need to learn these behaviors within a reasonable amount of time while efficiently leveraging the sparse feedback a human trainer is capable of providing. Recent work has shown that human feedback can be characterized as a critique of an agent's current behavior rather than as an alternative reward signal to be maximized, culminating in the COnvergent Actor-Critic by Humans (COACH) algorithm for making direct policy up…
MIT（麻省理工学院）；University of California, Berkeley（加州大学伯克利分校）
We introduce the first direct policy search algorithm which provably converges to the globally optimal $\textit{dynamic}$ filter for the classical problem of predicting the outputs of a linear dynamical system, given noisy, partial observations. Despite the ubiquity of partial observability in practice, theoretical guarantees for direct policy search algorithms, one of the backbones of modern reinforcement learning, have proven difficult to achieve. This is primarily due to the degeneracies which arise when optimizing over filters that maintain internal state. In this paper, we provide a new p…
Learning complicated value functions in high dimensional state space by function approximation is a challenging task, partially due to that the max-operator used in temporal difference updates can theoretically cause instability for most linear or non-linear approximation schemes. Mellowmax is a recently proposed differentiable and non-expansion softmax operator that allows a convergent behavior in learning and planning. Unfortunately, the performance bound for the fixed point it converges to remains unclear, and in practice, its parameter is sensitive to various domains and has to be tuned ca…
University of Ottawa（渥太华大学）；Aswan University（阿斯旺大学）；University of Texas at Arlington（德克萨斯大学阿灵顿分校）
The control problem of the flexible wing aircraft is challenging due to the prevailing and high nonlinear deformations in the flexible wing system. This urged for new control mechanisms that are robust to the real-time variations in the wing's aerodynamics. An online control mechanism based on a value iteration reinforcement learning process is developed for flexible wing aerial structures. It employs a model-free control policy framework and a guaranteed convergent adaptive learning architecture to solve the system's Bellman optimality equation. A Riccati equation is derived and shown to be e…
The Intelligent decision of the unmanned combat aerial vehicle (UCAV) has long been a challenging problem. The conventional search method can hardly satisfy the real-time demand during high dynamics air combat scenarios. The reinforcement learning (RL) method can significantly shorten the decision time via using neural networks. However, the sparse reward problem limits its convergence speed and the artificial prior experience reward can easily deviate its optimal convergent direction of the original task, which raises great difficulties for the RL air combat application. In this paper, we pro…
Beijing Institute of Technology（北京理工大学）；Delft University of Technology（代尔夫特理工大学）
The deep reinforcement learning-based energy management strategies (EMS) have become a promising solution for hybrid electric vehicles (HEVs). When driving cycles are changed, the neural network will be retrained, which is a time-consuming and laborious task. A more efficient way of choosing EMS is to combine deep reinforcement learning (DRL) with transfer learning, which can transfer knowledge of one domain to the other new domain, making the network of the new domain reach convergence values quickly. Different exploration methods of DRL, including adding action space noise and parameter spac…
To accumulate knowledge and improve its policy of behaviour, a reinforcement learning agent can learn `off-policy' about policies that differ from the policy used to generate its experience. This is important to learn counterfactuals, or because the experience was generated out of its own control. However, off-policy learning is non-trivial, and standard reinforcement-learning algorithms can be unstable and divergent. In this paper we discuss a novel family of off-policy prediction algorithms which are convergent by construction. The idea is to first learn on-policy about the data-generating b…
