Reinforcement Learning (RL) has achieved remarkable success in safety-critical areas, but it can be weakened by adversarial attacks. Recent studies have introduced "smoothed policies" in order to enhance its robustness. Yet, it is still challenging to establish a provable guarantee to certify the bound of its total reward. Prior methods relied primarily on computing bounds using Lipschitz continuity or calculating the probability of cumulative reward above specific thresholds. However, these techniques are only suited for continuous perturbations on the RL agent's observations and are restricted to perturbations bounded by the $l_2$-norm. To address these limitations, this paper proposes a general black-box certification method capable of directly certifying the cumulative reward of the smoothed policy under various $l_p$-norm bounded perturbations. Furthermore, we extend our methodology to certify perturbations on action spaces. Our approach leverages f-divergence to measure the distinction between the original distribution and the perturbed distribution, subsequently determining the certification bound by solving a convex optimisation problem. We provide a comprehensive theoretical analysis and run sufficient experiments in multiple environments. Our results show that our method not only improves the certified lower bound of mean cumulative reward but also demonstrates better efficiency than state-of-the-art techniques.
The large action space is one fundamental obstacle to deploying Reinforcement Learning methods in the real world. The numerous redundant actions will cause the agents to make repeated or invalid attempts, even leading to task failure. Although current algorithms conduct some initial explorations for this issue, they either suffer from rule-based systems or depend on expert demonstrations, which significantly limits their applicability in many real-world settings. In this work, we examine the theoretical analysis of what action can be eliminated in policy optimization and propose a novel redundant action filtering mechanism. Unlike other works, our method constructs the similarity factor by estimating the distance between the state distributions, which requires no prior knowledge. In addition, we combine the modified inverse model to avoid extensive computation in high-dimensional state space. We reveal the underlying structure of action spaces and propose a simple yet efficient redundant action filtering mechanism named No Prior Mask (NPM) based on the above techniques. We show the superior performance of our method by conducting extensive experiments on high-dimensional, pixel-input, and stochastic problems with various action redundancy. Our code is public online at https://github.com/zhongdy15/npm.
The promotion of large-scale applications of reinforcement learning (RL) requires efficient training computation. While existing parallel RL frameworks encompass a variety of RL algorithms and parallelization techniques, the excessively burdensome communication frameworks hinder the attainment of the hardware's limit for final throughput and training effects on a single desktop. In this paper, we propose Spreeze, a lightweight parallel framework for RL that efficiently utilizes a single desktop hardware resource to approach the throughput limit. We asynchronously parallelize the experience sampling, network update, performance evaluation, and visualization operations, and employ multiple efficient data transmission techniques to transfer various types of data between processes. The framework can automatically adjust the parallelization hyperparameters based on the computing ability of the hardware device in order to perform efficient large-batch updates. Based on the characteristics of the "Actor-Critic" RL algorithm, our framework uses dual GPUs to independently update the network of actors and critics in order to further improve throughput. Simulation results show that our framework can achieve up to 15,000Hz experience sampling and 370,000Hz network update frame rate using only a personal desktop computer, which is an order of magnitude higher than other mainstream parallel RL frameworks, resulting in a 73% reduction of training time. Our work on fully utilizing the hardware resources of a single desktop computer is fundamental to enabling efficient large-scale distributed RL training.
Reinforcement learning and Imitation Learning approaches utilize policy learning strategies that are difficult to generalize well with just a few examples of a task. In this work, we propose a language-conditioned semantic search-based method to produce an online search-based policy from the available demonstration dataset of state-action trajectories. Here we directly acquire actions from the most similar manipulation trajectories found in the dataset. Our approach surpasses the performance of the baselines on the CALVIN benchmark and exhibits strong zero-shot adaptation capabilities. This holds great potential for expanding the use of our online search-based policy approach to tasks typically addressed by Imitation Learning or Reinforcement Learning-based policies.
Safe reinforcement learning deals with mitigating or avoiding unsafe situations by reinforcement learning (RL) agents. Safe RL approaches are based on specific risk representations for particular problems or domains. In order to analyze agent behaviors, compare safe RL approaches, and effectively transfer techniques between application domains, it is necessary to understand the types of risk specific to safe RL problems. We performed a systematic literature mapping with the objective to characterize risk in safe RL. Based on the obtained results, we present definitions, characteristics, and types of risk that hold on multiple application domains. Our literature mapping covers literature from the last 5 years (2017-2022), from a variety of knowledge areas (AI, finance, engineering, medicine) where RL approaches emphasize risk representation and management. Our mapping covers 72 papers filtered systematically from over thousands of papers on the topic. Our proposed notion of risk covers a variety of representations, disciplinary differences, common training exercises, and types of techniques. We encourage researchers to include explicit and detailed accounts of risk in future safe RL research reports, using this mapping as a starting point. With this information, researchers and practitioners could draw stronger conclusions on the effectiveness of techniques on different problems.
Structural pruning has become an integral part of neural network optimization, used to achieve architectural configurations which can be deployed and run more efficiently on embedded devices. Previous results showed that pruning is possible with minimum performance loss by utilizing a reinforcement learning agent which makes decisions about the sparsity level of each neural layer by maximizing as a reward the accuracy of the network. We introduce a novel information-theoretic reward function which minimizes the spatial entropy of convolutional activations. This minimization ultimately acts as a proxy for maintaining accuracy, although these two criteria are not related in any way. Our method shows that there is another possibility to preserve accuracy without the need to directly optimize it in the agent's reward function. In our experiments, we were able to reduce the total number of FLOPS of multiple popular neural network architectures by 5-10x, incurring minimal or no performance drop and being on par with the solution found by maximizing the accuracy.
Parallel Reinforcement Learning (RL) frameworks are essential for mapping RL workloads to multiple computational resources, allowing for faster generation of samples, estimation of values, and policy improvement. These computational paradigms require a seamless integration of training, serving, and simulation workloads. Existing frameworks, such as Ray, are not managing this orchestration efficiently, especially in RL tasks that demand intensive input/output and synchronization between actors on a single node. In this study, we have proposed a solution implementing the reactor model, which enforces a set of actors to have a fixed communication pattern. This allows the scheduler to eliminate work needed for synchronization, such as acquiring and releasing locks for each actor or sending and processing coordination-related messages. Our framework, Lingua Franca (LF), a coordination language based on the reactor model, also supports true parallelism in Python and provides a unified interface that allows users to automatically generate dataflow graphs for RL tasks. In comparison to Ray on a single-node multi-core compute platform, LF achieves 1.21x and 11.62x higher simulation throughput in OpenAI Gym and Atari environments, reduces the average training time of synchronized parallel Q-learning by 31.2%, and accelerates multi-agent RL inference by 5.12x.
We present DARLEI, a framework that combines evolutionary algorithms with parallelized reinforcement learning for efficiently training and evolving populations of UNIMAL agents. Our approach utilizes Proximal Policy Optimization (PPO) for individual agent learning and pairs it with a tournament selection-based generational learning mechanism to foster morphological evolution. By building on Nvidia's Isaac Gym, DARLEI leverages GPU accelerated simulation to achieve over 20x speedup using just a single workstation, compared to previous work which required large distributed CPU clusters. We systematically characterize DARLEI's performance under various conditions, revealing factors impacting diversity of evolved morphologies. For example, by enabling inter-agent collisions within the simulator, we find that we can simulate some multi-agent interactions between the same morphology, and see how it influences individual agent capabilities and long-term evolutionary adaptation. While current results demonstrate limited diversity across generations, we hope to extend DARLEI in future work to include interactions between diverse morphologies in richer environments, and create a platform that allows for coevolving populations and investigating emergent behaviours in them. Our source code is also made publicly at https://saeejithnair.github.io/darlei.
Communication networks able to withstand hostile environments are critically important for disaster relief operations. In this paper, we consider a challenging scenario where drones have been compromised in the supply chain, during their manufacture, and harbour malicious software capable of wide-ranging and infectious disruption. We investigate multi-agent deep reinforcement learning as a tool for learning defensive strategies that maximise communications bandwidth despite continual adversarial interference. Using a public challenge for learning network resilience strategies, we propose a state-of-the-art expert technique and study its superiority over deep reinforcement learning agents. Correspondingly, we identify three specific methods for improving the performance of our learning-based agents: (1) ensuring each observation contains the necessary information, (2) using expert agents to provide a curriculum for learning, and (3) paying close attention to reward. We apply our methods and present a new mixed strategy enabling expert and learning-based agents to work together and improve on all prior results.
The past decade has seen vast progress in deep reinforcement learning (RL) on the back of algorithms manually designed by human researchers. Recently, it has been shown that it is possible to meta-learn update rules, with the hope of discovering algorithms that can perform well on a wide range of RL tasks. Despite impressive initial results from algorithms such as Learned Policy Gradient (LPG), there remains a generalization gap when these algorithms are applied to unseen environments. In this work, we examine how characteristics of the meta-training distribution impact the generalization performance of these algorithms. Motivated by this analysis and building on ideas from Unsupervised Environment Design (UED), we propose a novel approach for automatically generating curricula to maximize the regret of a meta-learned optimizer, in addition to a novel approximation of regret, which we name algorithmic regret (AR). The result is our method, General RL Optimizers Obtained Via Environment Design (GROOVE). In a series of experiments, we show that GROOVE achieves superior generalization to LPG, and evaluate AR against baseline metrics from UED, identifying it as a critical component of environment design in this setting. We believe this approach is a step towards the discovery of truly general RL algorithms, capable of solving a wide range of real-world environments.
Despite the impressive feats demonstrated by Reinforcement Learning (RL), these algorithms have seen little adoption in high-risk, real-world applications due to current difficulties in explaining RL agent actions and building user trust. We present Counterfactual Demonstrations for Explanation (CODEX), a method that incorporates semantic clustering, which can effectively summarize RL agent behavior in the state-action space. Experimentation on the MiniGrid and StarCraft II gaming environments reveals the semantic clusters retain temporal as well as entity information, which is reflected in the constructed summary of agent behavior. Furthermore, clustering the discrete+continuous game-state latent representations identifies the most crucial episodic events, demonstrating a relationship between the latent and semantic spaces. This work contributes to the growing body of work that strives to unlock the power of RL for widespread use by leveraging and extending techniques from Natural Language Processing.
Reinforcement learning (RL) is a versatile framework for optimizing long-term goals. Although many real-world problems can be formalized with RL, learning and deploying a performant RL policy requires a system designed to address several important challenges, including the exploration-exploitation dilemma, partial observability, dynamic action spaces, and safety concerns. While the importance of these challenges has been well recognized, existing open-source RL libraries do not explicitly address them. This paper introduces Pearl, a Production-Ready RL software package designed to embrace these challenges in a modular way. In addition to presenting benchmarking results, we also highlight examples of Pearl's ongoing industry adoption to demonstrate its advantages for production use cases. Pearl is open sourced on GitHub at github.com/facebookresearch/pearl and its official website is pearlagent.github.io.
Transferring knowledge in cross-domain reinforcement learning is a challenging setting in which learning is accelerated by reusing knowledge from a task with different observation and/or action space. However, it is often necessary to carefully select the source of knowledge for the receiving end to benefit from the transfer process. In this article, we study how to measure the similarity between cross-domain reinforcement learning tasks to select a source of knowledge that will improve the performance of the learning agent. We developed a semi-supervised alignment loss to match different spaces with a set of encoder-decoders, and use them to measure similarity and transfer policies across tasks. In comparison to prior works, our method does not require data to be aligned, paired or collected by expert policies. Experimental results, on a set of varied Mujoco control tasks, show the robustness of our method in effectively selecting and transferring knowledge, without the supervision of a tailored set of source tasks.
Offline Multi-agent Reinforcement Learning (MARL) is valuable in scenarios where online interaction is impractical or risky. While independent learning in MARL offers flexibility and scalability, accurately assigning credit to individual agents in offline settings poses challenges because interactions with an environment are prohibited. In this paper, we propose a new framework, namely Multi-Agent Causal Credit Assignment (MACCA), to address credit assignment in the offline MARL setting. Our approach, MACCA, characterizing the generative process as a Dynamic Bayesian Network, captures relationships between environmental variables, states, actions, and rewards. Estimating this model on offline data, MACCA can learn each agent's contribution by analyzing the causal relationship of their individual rewards, ensuring accurate and interpretable credit assignment. Additionally, the modularity of our approach allows it to seamlessly integrate with various offline MARL methods. Theoretically, we proved that under the setting of the offline dataset, the underlying causal structure and the function for generating the individual rewards of agents are identifiable, which laid the foundation for the correctness of our modeling. In our experiments, we demonstrate that MACCA not only outperforms state-of-the-art methods but also enhances performance when integrated with other backbones.
Active feature acquisition (AFA) agents, crucial in domains like healthcare where acquiring features is often costly or harmful, determine the optimal set of features for a subsequent classification task. As deploying an AFA agent introduces a shift in missingness distribution, it's vital to assess its expected performance at deployment using retrospective data. In a companion paper, we introduce a semi-offline reinforcement learning (RL) framework for active feature acquisition performance evaluation (AFAPE) where features are assumed to be time-dependent. Here, we study and extend the AFAPE problem to cover static feature settings, where features are time-invariant, and hence provide more flexibility to the AFA agents in deciding the order of the acquisitions. In this static feature setting, we derive and adapt new inverse probability weighting (IPW), direct method (DM), and double reinforcement learning (DRL) estimators within the semi-offline RL framework. These estimators can be applied when the missingness in the retrospective dataset follows a missing-at-random (MAR) pattern. They also can be applied to missing-not-at-random (MNAR) patterns in conjunction with appropriate existing missing data techniques. We illustrate the improved data efficiency offered by the semi-offline RL estimators in synthetic and real-world data experiments under synthetic MAR and MNAR missingness.
In this paper, we introduce a novel algorithm - the Skill-Driven Skill Recombination Algorithm (SDSRA) - an innovative framework that significantly enhances the efficiency of achieving maximum entropy in reinforcement learning tasks. We find that SDSRA achieves faster convergence compared to the traditional Soft Actor-Critic (SAC) algorithm and produces improved policies. By integrating skill-based strategies within the robust Actor-Critic framework, SDSRA demonstrates remarkable adaptability and performance across a wide array of complex and diverse benchmarks.
In this work, we investigate the means of using curiosity on replay buffers to improve offline multi-task continual reinforcement learning when tasks, which are defined by the non-stationarity in the environment, are non labeled and not evenly exposed to the learner in time. In particular, we investigate the use of curiosity both as a tool for task boundary detection and as a priority metric when it comes to retaining old transition tuples, which we respectively use to propose two different buffers. Firstly, we propose a Hybrid Reservoir Buffer with Task Separation (HRBTS), where curiosity is used to detect task boundaries that are not known due to the task agnostic nature of the problem. Secondly, by using curiosity as a priority metric when it comes to retaining old transition tuples, a Hybrid Curious Buffer (HCB) is proposed. We ultimately show that these buffers, in conjunction with regular reinforcement learning algorithms, can be used to alleviate the catastrophic forgetting issue suffered by the state of the art on replay buffers when the agent's exposure to tasks is not equal along time. We evaluate catastrophic forgetting and the efficiency of our proposed buffers against the latest works such as the Hybrid Reservoir Buffer (HRB) and the Multi-Time Scale Replay Buffer (MTR) in three different continual reinforcement learning settings. Experiments were done on classical control tasks and Metaworld environment. Experiments show that our proposed replay buffers display better immunity to catastrophic forgetting compared to existing works in most of the settings.
We propose Pgx, a suite of board game reinforcement learning (RL) environments written in JAX and optimized for GPU/TPU accelerators. By leveraging JAX's auto-vectorization and parallelization over accelerators, Pgx can efficiently scale to thousands of simultaneous simulations over accelerators. In our experiments on a DGX-A100 workstation, we discovered that Pgx can simulate RL environments 10-100x faster than existing implementations available in Python. Pgx includes RL environments commonly used as benchmarks in RL research, such as backgammon, chess, shogi, and Go. Additionally, Pgx offers miniature game sets and baseline models to facilitate rapid research cycles. We demonstrate the efficient training of the Gumbel AlphaZero algorithm with Pgx environments. Overall, Pgx provides high-performance environment simulators for researchers to accelerate their RL experiments. Pgx is available at http://github.com/sotetsuk/pgx.
A major challenge in reinforcement learning is to determine which state-action pairs are responsible for future rewards that are delayed. Reward redistribution serves as a solution to re-assign credits for each time step from observed sequences. While the majority of current approaches construct the reward redistribution in an uninterpretable manner, we propose to explicitly model the contributions of state and action from a causal perspective, resulting in an interpretable reward redistribution and preserving policy invariance. In this paper, we start by studying the role of causal generative models in reward redistribution by characterizing the generation of Markovian rewards and trajectory-wise long-term return and further propose a framework, called Generative Return Decomposition (GRD), for policy optimization in delayed reward scenarios. Specifically, GRD first identifies the unobservable Markovian rewards and causal relations in the generative process. Then, GRD makes use of the identified causal generative model to form a compact representation to train policy over the most favorable subspace of the state space of the agent. Theoretically, we show that the unobservable Markovian reward function is identifiable, as well as the underlying causal structure and causal models. Experimental results show that our method outperforms state-of-the-art methods and the provided visualization further demonstrates the interpretability of our method. The project page is located at https://reedzyd.github.io/GenerativeReturnDecomposition/.
Advances in artificial intelligence (AI) have led to its application in many areas of everyday life. In the context of control engineering, reinforcement learning (RL) represents a particularly promising approach as it is centred around the idea of allowing an agent to freely interact with its environment to find an optimal strategy. One of the challenges professionals face when training and deploying RL agents is that the latter often have to run on dedicated embedded devices. This could be to integrate them into an existing toolchain or to satisfy certain performance criteria like real-time constraints. Conventional RL libraries, however, cannot be easily utilised in conjunction with that kind of hardware. In this paper, we present a framework named LExCI, the Learning and Experiencing Cycle Interface, which bridges this gap and provides end-users with a free and open-source tool for training agents on embedded systems using the open-source library RLlib. Its operability is demonstrated with two state-of-the-art RL-algorithms and a rapid control prototyping system.