Despite their success, policy gradient methods suffer from high variance of the gradient estimate, which can result in unsatisfactory sample complexity. Recently, numerous variance-reduced extensions of policy gradient methods with provably better sample complexity and competitive numerical performance have been proposed. After a compact survey on some of the main variance-reduced REINFORCE-type methods, we propose ProbAbilistic Gradient Estimation for Policy Gradient (PAGE-PG), a novel loopless variance-reduced policy gradient method based on a probabilistic switch between two types of updates. Our method is inspired by the PAGE estimator for supervised learning and leverages importance sampling to obtain an unbiased gradient estimator. We show that PAGE-PG enjoys a $\mathcal{O}\left( \epsilon^{-3} \right)$ average sample complexity to reach an $\epsilon$-stationary solution, which matches the sample complexity of its most competitive counterparts under the same setting. A numerical evaluation confirms the competitive performance of our method on classical control tasks.
Enabling reinforcement learning (RL) agents to leverage a knowledge base while learning from experience promises to advance RL in knowledge intensive domains. However, it has proven difficult to leverage knowledge that is not manually tailored to the environment. We propose to use the subclass relationships present in open-source knowledge graphs to abstract away from specific objects. We develop a residual policy gradient method that is able to integrate knowledge across different abstraction levels in the class hierarchy. Our method results in improved sample efficiency and generalisation to unseen objects in commonsense games, but we also investigate failure modes, such as excessive noise in the extracted class knowledge or environments with little class structure.
Natural policy gradient methods are popular reinforcement learning methods that improve the stability of policy gradient methods by utilizing second-order approximations to precondition the gradient with the inverse of the Fisher-information matrix. However, to the best of the authors' knowledge, there has not been a study that has investigated the effects of different second-order approximations in a comprehensive and systematic manner. To address this, five different second-order approximations were studied and compared across multiple key metrics including performance, stability, sample efficiency, and computation time. Furthermore, hyperparameters which aren't typically acknowledged in the literature are studied including the effect of different batch sizes and optimizing the critic network with the natural gradient. Experimental results show that on average, improved second-order approximations achieve the best performance and that using properly tuned hyperparameters can lead to large improvements in performance and sample efficiency ranging up to +181%. We also make the code in this study available at https://github.com/gebob19/natural-policy-gradient-reinforcement-learning.
We propose a new policy gradient method, named homotopic policy mirror descent (HPMD), for solving discounted, infinite horizon MDPs with finite state and action spaces. HPMD performs a mirror descent type policy update with an additional diminishing regularization term, and possesses several computational properties that seem to be new in the literature. We first establish the global linear convergence of HPMD instantiated with Kullback-Leibler divergence, for both the optimality gap, and a weighted distance to the set of optimal policies. Then local superlinear convergence is obtained for both quantities without any assumption. With local acceleration and diminishing regularization, we establish the first result among policy gradient methods on certifying and characterizing the limiting policy, by showing, with a non-asymptotic characterization, that the last-iterate policy converges to the unique optimal policy with the maximal entropy. We then extend all the aforementioned results to HPMD instantiated with a broad class of decomposable Bregman divergences, demonstrating the generality of the these computational properties. As a by product, we discover the finite-time exact convergence for some commonly used Bregman divergences, implying the continuing convergence of HPMD to the limiting policy even if the current policy is already optimal. Finally, we develop a stochastic version of HPMD and establish similar convergence properties. By exploiting the local acceleration, we show that for small optimality gap, a better than $\tilde{\mathcal{O}}(\left|\mathcal{S}\right| \left|\mathcal{A}\right| / \epsilon^2)$ sample complexity holds with high probability, when assuming a generative model for policy evaluation.
We consider infinite-horizon discounted Markov decision problems with finite state and action spaces and study the convergence rates of the projected policy gradient method and a general class of policy mirror descent methods, all with direct parametrization in the policy space. First, we develop a theory of weak gradient-mapping dominance and use it to prove sharper sublinear convergence rate of the projected policy gradient method. Then we show that with geometrically increasing step sizes, a general class of policy mirror descent methods, including the natural policy gradient method and a projected Q-descent method, all enjoy a linear rate of convergence without relying on entropy or other strongly convex regularization. Finally, we also analyze the convergence rate of an inexact policy mirror descent method and estimate its sample complexity under a simple generative model.
Engineering problems that are modeled using sophisticated mathematical methods or are characterized by expensive-to-conduct tests or experiments, are encumbered with limited budget or finite computational resources. Moreover, practical scenarios in the industry, impose restrictions, based on logistics and preference, on the manner in which the experiments can be conducted. For example, material supply may enable only a handful of experiments in a single-shot or in the case of computational models one may face significant wait-time based on shared computational resources. In such scenarios, one usually resorts to performing experiments in a manner that allows for maximizing one's state-of-knowledge while satisfying the above mentioned practical constraints. Sequential design of experiments (SDOE) is a popular suite of methods, that has yielded promising results in recent years across different engineering and practical problems. A common strategy, that leverages Bayesian formalism is the Bayesian SDOE, which usually works best in the one-step-ahead or myopic scenario of selecting a single experiment at each step of a sequence of experiments. In this work, we aim to extend the SDOE strategy, to query the experiment or computer code at a batch of inputs. To this end, we leverage deep reinforcement learning (RL) based policy gradient methods, to propose batches of queries that are selected taking into account entire budget in hand. The algorithm retains the sequential nature, inherent in the SDOE, while incorporating elements of reward based on task from the domain of deep RL. A unique capability of the proposed methodology is its ability to be applied to multiple tasks, for example optimization of a function, once its trained. We demonstrate the performance of the proposed algorithm on a synthetic problem, and a challenging high-dimensional engineering problem.
We propose a novel policy gradient method for multi-agent reinforcement learning, which leverages two different variance-reduction techniques and does not require large batches over iterations. Specifically, we propose a momentum-based decentralized policy gradient tracking (MDPGT) where a new momentum-based variance reduction technique is used to approximate the local policy gradient surrogate with importance sampling, and an intermediate parameter is adopted to track two consecutive policy gradient surrogates. Moreover, MDPGT provably achieves the best available sample complexity of $\mathcal{O}(N^{-1}\epsilon^{-3})$ for converging to an $\epsilon$-stationary point of the global average of $N$ local performance functions (possibly nonconcave). This outperforms the state-of-the-art sample complexity in decentralized model-free reinforcement learning, and when initialized with a single trajectory, the sample complexity matches those obtained by the existing decentralized policy gradient methods. We further validate the theoretical claim for the Gaussian policy function. When the required error tolerance $\epsilon$ is small enough, MDPGT leads to a linear speed up, which has been previously established in decentralized stochastic optimization, but not for reinforcement learning. Lastly, we provide empirical results on a multi-agent reinforcement learning benchmark environment to support our theoretical findings.
We introduce a novel training procedure for policy gradient methods wherein episodic memory is used to optimize the hyperparameters of reinforcement learning algorithms on-the-fly. Unlike other hyperparameter searches, we formulate hyperparameter scheduling as a standard Markov Decision Process and use episodic memory to store the outcome of used hyperparameters and their training contexts. At any policy update step, the policy learner refers to the stored experiences, and adaptively reconfigures its learning algorithm with the new hyperparameters determined by the memory. This mechanism, dubbed as Episodic Policy Gradient Training (EPGT), enables an episodic learning process, and jointly learns the policy and the learning algorithm's hyperparameters within a single run. Experimental results on both continuous and discrete environments demonstrate the advantage of using the proposed method in boosting the performance of various policy gradient algorithms.
Owing to the growth of interest in Reinforcement Learning in the last few years, gradient based policy control methods have been gaining popularity for Control problems as well. And rightly so, since gradient policy methods have the advantage of optimizing a metric of interest in an end-to-end manner, along with being relatively easy to implement without complete knowledge of the underlying system. In this paper, we study the global convergence of gradient-based policy optimization methods for quadratic control of discrete-time and model-free Markovian jump linear systems (MJLS). We surmount myriad challenges that arise because of more than one states coupled with lack of knowledge of the system dynamics and show global convergence of the policy using gradient descent and natural policy gradient methods. We also provide simulation studies to corroborate our claims.
In reinforcement learning, continuous time is often discretized by a time scale $\delta$, to which the resulting performance is known to be highly sensitive. In this work, we seek to find a $\delta$-invariant algorithm for policy gradient (PG) methods, which performs well regardless of the value of $\delta$. We first identify the underlying reasons that cause PG methods to fail as $\delta \to 0$, proving that the variance of the PG estimator can diverge to infinity in stochastic environments under a certain assumption of stochasticity. While durative actions or action repetition can be employed to have $\delta$-invariance, previous action repetition methods cannot immediately react to unexpected situations in stochastic environments. We thus propose a novel $\delta$-invariant method named Safe Action Repetition (SAR) applicable to any existing PG algorithm. SAR can handle the stochasticity of environments by adaptively reacting to changes in states during action repetition. We empirically show that our method is not only $\delta$-invariant but also robust to stochasticity, outperforming previous $\delta$-invariant approaches on eight MuJoCo environments with both deterministic and stochastic settings. Our code is available at https://vision.snu.ac.kr/projects/sar.
In this paper, we establish the global optimality and convergence rate of an off-policy actor critic algorithm in the tabular setting without using density ratio to correct the discrepancy between the state distribution of the behavior policy and that of the target policy. Our work goes beyond existing works on the optimality of policy gradient methods in that existing works use the exact policy gradient for updating the policy parameters while we use an approximate and stochastic update step. Our update step is not a gradient update because we do not use a density ratio to correct the state distribution, which aligns well with what practitioners do. Our update is approximate because we use a learned critic instead of the true value function. Our update is stochastic because at each step the update is done for only the current state action pair. Moreover, we remove several restrictive assumptions from existing works in our analysis. Central to our work is the finite sample analysis of a generic stochastic approximation algorithm with time-inhomogeneous update operators on time-inhomogeneous Markov chains, based on its uniform contraction properties.
Reinforcement learning methods for continuous control tasks have evolved in recent years generating a family of policy gradient methods that rely primarily on a Gaussian distribution for modeling a stochastic policy. However, the Gaussian distribution has an infinite support, whereas real world applications usually have a bounded action space. This dissonance causes an estimation bias that can be eliminated if the Beta distribution is used for the policy instead, as it presents a finite support. In this work, we investigate how this Beta policy performs when it is trained by the Proximal Policy Optimization (PPO) algorithm on two continuous control tasks from OpenAI gym. For both tasks, the Beta policy is superior to the Gaussian policy in terms of agent's final expected reward, also showing more stability and faster convergence of the training process. For the CarRacing environment with high-dimensional image input, the agent's success rate was improved by 63% over the Gaussian policy.
Policy gradient methods have been frequently applied to problems in control and reinforcement learning with great success, yet existing convergence analysis still relies on non-intuitive, impractical and often opaque conditions. In particular, existing rates are achieved in limited settings, under strict regularity conditions. In this work, we establish explicit convergence rates of policy gradient methods, extending the convergence regime to weakly smooth policy classes with $L_2$ integrable gradient. We provide intuitive examples to illustrate the insight behind these new conditions. Notably, our analysis also shows that convergence rates are achievable for both the standard policy gradient and the natural policy gradient algorithms under these assumptions. Lastly we provide performance guarantees for the converged policies.
Entropy regularization is an efficient technique for encouraging exploration and preventing a premature convergence of (vanilla) policy gradient methods in reinforcement learning (RL). However, the theoretical understanding of entropy-regularized RL algorithms has been limited. In this paper, we revisit the classical entropy regularized policy gradient methods with the soft-max policy parametrization, whose convergence has so far only been established assuming access to exact gradient oracles. To go beyond this scenario, we propose the first set of (nearly) unbiased stochastic policy gradient estimators with trajectory-level entropy regularization, with one being an unbiased visitation measure-based estimator and the other one being a nearly unbiased yet more practical trajectory-based estimator. We prove that although the estimators themselves are unbounded in general due to the additional logarithmic policy rewards introduced by the entropy term, the variances are uniformly bounded. We then propose a two-phase stochastic policy gradient (PG) algorithm that uses a large batch size in the first phase to overcome the challenge of the stochastic approximation due to the non-coercive landscape, and uses a small batch size in the second phase by leveraging the curvature information around the optimal policy. We establish a global optimality convergence result and a sample complexity of $\widetilde{\mathcal{O}}(\frac{1}{\epsilon^2})$ for the proposed algorithm. Our result is the first global convergence and sample complexity results for the stochastic entropy-regularized vanilla PG method.
Policy gradient methods have become popular in multi-agent reinforcement learning, but they suffer from high variance due to the presence of environmental stochasticity and exploring agents (i.e., non-stationarity), which is potentially worsened by the difficulty in credit assignment. As a result, there is a need for a method that is not only capable of efficiently solving the above two problems but also robust enough to solve a variety of tasks. To this end, we propose a new multi-agent policy gradient method, called Robust Local Advantage (ROLA) Actor-Critic. ROLA allows each agent to learn an individual action-value function as a local critic as well as ameliorating environment non-stationarity via a novel centralized training approach based on a centralized critic. By using this local critic, each agent calculates a baseline to reduce variance on its policy gradient estimation, which results in an expected advantage action-value over other agents' choices that implicitly improves credit assignment. We evaluate ROLA across diverse benchmarks and show its robustness and effectiveness over a number of state-of-the-art multi-agent policy gradient algorithms.
Stabilizing an unknown control system is one of the most fundamental problems in control systems engineering. In this paper, we provide a simple, model-free algorithm for stabilizing fully observed dynamical systems. While model-free methods have become increasingly popular in practice due to their simplicity and flexibility, stabilization via direct policy search has received surprisingly little attention. Our algorithm proceeds by solving a series of discounted LQR problems, where the discount factor is gradually increased. We prove that this method efficiently recovers a stabilizing controller for linear systems, and for smooth, nonlinear systems within a neighborhood of their equilibria. Our approach overcomes a significant limitation of prior work, namely the need for a pre-given stabilizing control policy. We empirically evaluate the effectiveness of our approach on common control benchmarks.
An agent's functionality is largely determined by its design, i.e., skeletal structure and joint attributes (e.g., length, size, strength). However, finding the optimal agent design for a given function is extremely challenging since the problem is inherently combinatorial and the design space is prohibitively large. Additionally, it can be costly to evaluate each candidate design which requires solving for its optimal controller. To tackle these problems, our key idea is to incorporate the design procedure of an agent into its decision-making process. Specifically, we learn a conditional policy that, in an episode, first applies a sequence of transform actions to modify an agent's skeletal structure and joint attributes, and then applies control actions under the new design. To handle a variable number of joints across designs, we use a graph-based policy where each graph node represents a joint and uses message passing with its neighbors to output joint-specific actions. Using policy gradient methods, our approach enables joint optimization of agent design and control as well as experience sharing across different designs, which improves sample efficiency substantially. Experiments show that our approach, Transform2Act, outperforms prior methods significantly in terms of convergence speed and final performance. Notably, Transform2Act can automatically discover plausible designs similar to giraffes, squids, and spiders. Code and videos are available at https://sites.google.com/view/transform2act.
When designing algorithms for finite-time-horizon episodic reinforcement learning problems, a common approach is to introduce a fictitious discount factor and use stationary policies for approximations. Empirically, it has been shown that the fictitious discount factor helps reduce variance, and stationary policies serve to save the per-iteration computational cost. Theoretically, however, there is no existing work on convergence analysis for algorithms with this fictitious discount recipe. This paper takes the first step towards analyzing these algorithms. It focuses on two vanilla policy gradient (VPG) variants: the first being a widely used variant with discounted advantage estimations (DAE), the second with an additional fictitious discount factor in the score functions of the policy gradient estimators. Non-asymptotic convergence guarantees are established for both algorithms, and the additional discount factor is shown to reduce the bias introduced in DAE and thus improve the algorithm convergence asymptotically. A key ingredient of our analysis is to connect three settings of Markov decision processes (MDPs): the finite-time-horizon, the average reward and the discounted settings. To our best knowledge, this is the first theoretical guarantee on fictitious discount algorithms for the episodic reinforcement learning of finite-time-horizon MDPs, which also leads to the (first) global convergence of policy gradient methods for finite-time-horizon episodic reinforcement learning.
Opinion summarization has been traditionally approached with unsupervised, weakly-supervised and few-shot learning techniques. In this work, we collect a large dataset of summaries paired with user reviews for over 31,000 products, enabling supervised training. However, the number of reviews per product is large (320 on average), making summarization - and especially training a summarizer - impractical. Moreover, the content of many reviews is not reflected in the human-written summaries, and, thus, the summarizer trained on random review subsets hallucinates. In order to deal with both of these challenges, we formulate the task as jointly learning to select informative subsets of reviews and summarizing the opinions expressed in these subsets. The choice of the review subset is treated as a latent variable, predicted by a small and simple selector. The subset is then fed into a more powerful summarizer. For joint training, we use amortized variational inference and policy gradient methods. Our experiments demonstrate the importance of selecting informative reviews resulting in improved quality of summaries and reduced hallucinations.
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.