Recently, safe reinforcement learning (RL) with the actor-critic structure for continuous control tasks has received increasing attention. It is still challenging to learn a near-optimal control policy with safety and convergence guarantees. Also, few works have addressed the safe RL algorithm design under time-varying safety constraints. This paper proposes a safe RL algorithm for optimal control of nonlinear systems with time-varying state and control constraints. In the proposed approach, we construct a novel barrier force-based control policy structure to guarantee control safety. A multi-step policy evaluation mechanism is proposed to predict the policy's safety risk under time-varying safety constraints and guide the policy to update safely. Theoretical results on stability and robustness are proven. Also, the convergence of the actor-critic implementation is analyzed. The performance of the proposed algorithm outperforms several state-of-the-art RL algorithms in the simulated Safety Gym environment. Furthermore, the approach is applied to the integrated path following and collision avoidance problem for two real-world intelligent vehicles. A differential-drive vehicle and an Ackermann-drive one are used to verify offline deployment and online learning performance, respectively. Our approach shows an impressive sim-to-real transfer capability and a satisfactory online control performance in the experiment.
An important problem in reinforcement learning is designing agents that learn to solve tasks safely in an environment. A common solution is for a human expert to define either a penalty in the reward function or a cost to be minimised when reaching unsafe states. However, this is non-trivial, since too small a penalty may lead to agents that reach unsafe states, while too large a penalty increases the time to convergence. Additionally, the difficulty in designing reward or cost functions can increase with the complexity of the problem. Hence, for a given environment with a given set of unsafe states, we are interested in finding the upper bound of rewards at unsafe states whose optimal policies minimise the probability of reaching those unsafe states, irrespective of task rewards. We refer to this exact upper bound as the "Minmax penalty", and show that it can be obtained by taking into account both the controllability and diameter of an environment. We provide a simple practical model-free algorithm for an agent to learn this Minmax penalty while learning the task policy, and demonstrate that using it leads to agents that learn safe policies in high-dimensional continuous control environments.
We examine online safe multi-agent reinforcement learning using constrained Markov games in which agents compete by maximizing their expected total rewards under a constraint on expected total utilities. Our focus is confined to an episodic two-player zero-sum constrained Markov game with independent transition functions that are unknown to agents, adversarial reward functions, and stochastic utility functions. For such a Markov game, we employ an approach based on the occupancy measure to formulate it as an online constrained saddle-point problem with an explicit constraint. We extend the Lagrange multiplier method in constrained optimization to handle the constraint by creating a generalized Lagrangian with minimax decision primal variables and a dual variable. Next, we develop an upper confidence reinforcement learning algorithm to solve this Lagrangian problem while balancing exploration and exploitation. Our algorithm updates the minimax decision primal variables via online mirror descent and the dual variable via projected gradient step and we prove that it enjoys sublinear rate $ O((|X|+|Y|) L \sqrt{T(|A|+|B|)}))$ for both regret and constraint violation after playing $T$ episodes of the game. Here, $L$ is the horizon of each episode, $(|X|,|A|)$ and $(|Y|,|B|)$ are the state/action space sizes of the min-player and the max-player, respectively. To the best of our knowledge, we provide the first provably efficient online safe reinforcement learning algorithm in constrained Markov games.
Reinforcement learning (RL) is an area of significant research interest, and safe RL in particular is attracting attention due to its ability to handle safety-driven constraints that are crucial for real-world applications. This work proposes a novel approach to RL training, called control invariant set (CIS) enhanced RL, which leverages the advantages of utilizing the explicit form of CIS to improve stability guarantees and sampling efficiency. Furthermore, the robustness of the proposed approach is investigated in the presence of uncertainty. The approach consists of two learning stages: offline and online. In the offline stage, CIS is incorporated into the reward design, initial state sampling, and state reset procedures. This incorporation of CIS facilitates improved sampling efficiency during the offline training process. In the online stage, RL is retrained whenever the predicted next step state is outside of the CIS, which serves as a stability criterion, by introducing a Safety Supervisor to examine the safety of the action and make necessary corrections. The stability analysis is conducted for both cases, with and without uncertainty. To evaluate the proposed approach, we apply it to a simulated chemical reactor. The results show a significant improvement in sampling efficiency during offline training and closed-loop stability guarantee in the online implementation, with and without uncertainty.
Due to the trial-and-error nature, it is typically challenging to apply RL algorithms to safety-critical real-world applications, such as autonomous driving, human-robot interaction, robot manipulation, etc, where such errors are not tolerable. Recently, safe RL (i.e. constrained RL) has emerged rapidly in the literature, in which the agents explore the environment while satisfying constraints. Due to the diversity of algorithms and tasks, it remains difficult to compare existing safe RL algorithms. To fill that gap, we introduce GUARD, a Generalized Unified SAfe Reinforcement Learning Development Benchmark. GUARD has several advantages compared to existing benchmarks. First, GUARD is a generalized benchmark with a wide variety of RL agents, tasks, and safety constraint specifications. Second, GUARD comprehensively covers state-of-the-art safe RL algorithms with self-contained implementations. Third, GUARD is highly customizable in tasks and algorithms. We present a comparison of state-of-the-art safe RL algorithms in various task settings using GUARD and establish baselines that future work can build on.
This paper proposes a safe reinforcement learning (RL) algorithm that approximately solves the state-constrained optimal control problem for continuous-time uncertain nonlinear systems. We formulate the safe RL problem as the minimization of a Lagrangian that includes the cost functional and a user-defined barrier Lyapunov function (BLF) encoding the state constraints. We show that the analytical solution obtained by the application of Karush-Kuhn-Tucker (KKT) conditions contains a state-dependent expression for the Lagrange multiplier, which is a function of uncertain terms in the system dynamics. We argue that a naive estimation of the Lagrange multiplier may lead to safety constraint violations. To obviate this challenge, we propose an Actor-Critic-Identifier-Lagrangian (ACIL) algorithm that learns optimal control policies from online data without compromising safety. We provide safety and boundedness guarantees with the proposed algorithm and compare its performance with existing offline/online RL methods via a simulation study.
AI systems empowered by reinforcement learning (RL) algorithms harbor the immense potential to catalyze societal advancement, yet their deployment is often impeded by significant safety concerns. Particularly in safety-critical applications, researchers have raised concerns about unintended harms or unsafe behaviors of unaligned RL agents. The philosophy of safe reinforcement learning (SafeRL) is to align RL agents with harmless intentions and safe behavioral patterns. In SafeRL, agents learn to develop optimal policies by receiving feedback from the environment, while also fulfilling the requirement of minimizing the risk of unintended harm or unsafe behavior. However, due to the intricate nature of SafeRL algorithm implementation, combining methodologies across various domains presents a formidable challenge. This had led to an absence of a cohesive and efficacious learning framework within the contemporary SafeRL research milieu. In this work, we introduce a foundational framework designed to expedite SafeRL research endeavors. Our comprehensive framework encompasses an array of algorithms spanning different RL domains and places heavy emphasis on safety elements. Our efforts are to make the SafeRL-related research process more streamlined and efficient, therefore facilitating further research in AI safety. Our project is released at: https://github.com/PKU-Alignment/omnisafe.
The massive integration of renewable-based distributed energy resources (DERs) inherently increases the energy system's complexity, especially when it comes to defining its operational schedule. Deep reinforcement learning (DRL) algorithms arise as a promising solution due to their data-driven and model-free features. However, current DRL algorithms fail to enforce rigorous operational constraints (e.g., power balance, ramping up or down constraints) limiting their implementation in real systems. To overcome this, in this paper, a DRL algorithm (namely MIP-DQN) is proposed, capable of \textit{strictly} enforcing all operational constraints in the action space, ensuring the feasibility of the defined schedule in real-time operation. This is done by leveraging recent optimization advances for deep neural networks (DNNs) that allow their representation as a MIP formulation, enabling further consideration of any action space constraints. Comprehensive numerical simulations show that the proposed algorithm outperforms existing state-of-the-art DRL algorithms, obtaining a lower error when compared with the optimal global solution (upper boundary) obtained after solving a mathematical programming formulation with perfect forecast information; while strictly enforcing all operational constraints (even in unseen test days).
In many real-world applications, safety constraints for reinforcement learning (RL) algorithms are either unknown or not explicitly defined. We propose a framework that concurrently learns safety constraints and optimal RL policies in such environments, supported by theoretical guarantees. Our approach merges a logically-constrained RL algorithm with an evolutionary algorithm to synthesize signal temporal logic (STL) specifications. The framework is underpinned by theorems that establish the convergence of our joint learning process and provide error bounds between the discovered policy and the true optimal policy. We showcased our framework in grid-world environments, successfully identifying both acceptable safety constraints and RL policies while demonstrating the effectiveness of our theorems in practice.
Safety is the priority concern when applying reinforcement learning (RL) algorithms to real-world control problems. While policy iteration provides a fundamental algorithm for standard RL, an analogous theoretical algorithm for safe RL remains absent. In this paper, we propose feasible policy iteration (FPI), the first foundational dynamic programming algorithm for safe RL. FPI alternates between policy evaluation, region identification and policy improvement. This follows actor-critic-scenery (ACS) framework where scenery refers to a feasibility function that represents a feasible region. A region-wise update rule is developed for the policy improvement step, which maximizes state-value function inside the feasible region and minimizes feasibility function outside it. With this update rule, FPI guarantees monotonic expansion of feasible region, monotonic improvement of state-value function, and geometric convergence to the optimal safe policy. Experimental results demonstrate that FPI achieves strictly zero constraint violation on low-dimensional tasks and outperforms existing methods in constraint adherence and reward performance on high-dimensional tasks.
Safe reinforcement learning (RL) with hard constraint guarantees is a promising optimal control direction for multi-energy management systems. It only requires the environment-specific constraint functions itself a priori and not a complete model. The project-specific upfront and ongoing engineering efforts are therefore still reduced, better representations of the underlying system dynamics can still be learnt, and modelling bias is kept to a minimum. However, even the constraint functions alone are not always trivial to accurately provide in advance, leading to potentially unsafe behaviour. In this paper, we present two novel advancements: (I) combining the OptLayer and SafeFallback method, named OptLayerPolicy, to increase the initial utility while keeping a high sample efficiency and the possibility to formulate equality constraints. (II) introducing self-improving hard constraints, to increase the accuracy of the constraint functions as more and new data becomes available so that better policies can be learnt. Both advancements keep the constraint formulation decoupled from the RL formulation, so new (presumably better) RL algorithms can act as drop-in replacements. We have shown that, in a simulated multi-energy system case study, the initial utility is increased to 92.4% (OptLayerPolicy) compared to 86.1% (OptLayer) and that the policy after training is increased to 104.9% (GreyOptLayerPolicy) compared to 103.4% (OptLayer) - all relative to a vanilla RL benchmark. Although introducing surrogate functions into the optimisation problem requires special attention, we conclude that the newly presented GreyOptLayerPolicy method is the most advantageous.
Reinforcement learning (RL) has demonstrated impressive performance in various areas such as video games and robotics. However, ensuring safety and stability, which are two critical properties from a control perspective, remains a significant challenge when using RL to control real-world systems. In this paper, we first provide definitions of safety and stability for the RL system, and then combine the control barrier function (CBF) and control Lyapunov function (CLF) methods with the actor-critic method in RL to propose a Barrier-Lyapunov Actor-Critic (BLAC) framework which helps maintain the aforementioned safety and stability for the system. In this framework, CBF constraints for safety and CLF constraint for stability are constructed based on the data sampled from the replay buffer, and the augmented Lagrangian method is used to update the parameters of the RL-based controller. Furthermore, an additional backup controller is introduced in case the RL-based controller cannot provide valid control signals when safety and stability constraints cannot be satisfied simultaneously. Simulation results show that this framework yields a controller that can help the system approach the desired state and cause fewer violations of safety constraints compared to baseline algorithms.
An emerging field of sequential decision problems is safe Reinforcement Learning (RL), where the objective is to maximize the reward while obeying safety constraints. Being able to handle constraints is essential for deploying RL agents in real-world environments, where constraint violations can harm the agent and the environment. To this end, we propose a safe model-free RL algorithm with a novel multiplicative value function consisting of a safety critic and a reward critic. The safety critic predicts the probability of constraint violation and discounts the reward critic that only estimates constraint-free returns. By splitting responsibilities, we facilitate the learning task leading to increased sample efficiency. We integrate our approach into two popular RL algorithms, Proximal Policy Optimization and Soft Actor-Critic, and evaluate our method in four safety-focused environments, including classical RL benchmarks augmented with safety constraints and robot navigation tasks with images and raw Lidar scans as observations. Finally, we make the zero-shot sim-to-real transfer where a differential drive robot has to navigate through a cluttered room. Our code can be found at https://github.com/nikeke19/Safe-Mult-RL.
Safe Reinforcement learning (Safe RL) aims at learning optimal policies while staying safe. A popular solution to Safe RL is shielding, which uses a logical safety specification to prevent an RL agent from taking unsafe actions. However, traditional shielding techniques are difficult to integrate with continuous, end-to-end deep RL methods. To this end, we introduce Probabilistic Logic Policy Gradient (PLPG). PLPG is a model-based Safe RL technique that uses probabilistic logic programming to model logical safety constraints as differentiable functions. Therefore, PLPG can be seamlessly applied to any policy gradient algorithm while still providing the same convergence guarantees. In our experiments, we show that PLPG learns safer and more rewarding policies compared to other state-of-the-art shielding techniques.
We study optimality for the safety-constrained Markov decision process which is the underlying framework for safe reinforcement learning. Specifically, we consider a constrained Markov decision process (with finite states and finite actions) where the goal of the decision maker is to reach a target set while avoiding an unsafe set(s) with certain probabilistic guarantees. Therefore the underlying Markov chain for any control policy will be multichain since by definition there exists a target set and an unsafe set. The decision maker also has to be optimal (with respect to a cost function) while navigating to the target set. This gives rise to a multi-objective optimization problem. We highlight the fact that Bellman's principle of optimality may not hold for constrained Markov decision problems with an underlying multichain structure (as shown by the counterexample due to Haviv. We resolve the counterexample by formulating the aforementioned multi-objective optimization problem as a zero-sum game and thereafter construct an asynchronous value iteration scheme for the Lagrangian (similar to Shapley's algorithm). Finally, we consider the reinforcement learning problem for the same and construct a modified $Q$-learning algorithm for learning the Lagrangian from data. We also provide a lower bound on the number of iterations required for learning the Lagrangian and corresponding error bounds.
Safe reinforcement learning (RL) trains a constraint satisfaction policy by interacting with the environment. We aim to tackle a more challenging problem: learning a safe policy from an offline dataset. We study the offline safe RL problem from a novel multi-objective optimization perspective and propose the $\epsilon$-reducible concept to characterize problem difficulties. The inherent trade-offs between safety and task performance inspire us to propose the constrained decision transformer (CDT) approach, which can dynamically adjust the trade-offs during deployment. Extensive experiments show the advantages of the proposed method in learning an adaptive, safe, robust, and high-reward policy. CDT outperforms its variants and strong offline safe RL baselines by a large margin with the same hyperparameters across all tasks, while keeping the zero-shot adaptation capability to different constraint thresholds, making our approach more suitable for real-world RL under constraints. The code is available at https://github.com/liuzuxin/OSRL.
In this paper, we investigate a novel safe reinforcement learning problem with step-wise violation constraints. Our problem differs from existing works in that we consider stricter step-wise violation constraints and do not assume the existence of safe actions, making our formulation more suitable for safety-critical applications which need to ensure safety in all decision steps and may not always possess safe actions, e.g., robot control and autonomous driving. We propose a novel algorithm SUCBVI, which guarantees $\widetilde{O}(\sqrt{ST})$ step-wise violation and $\widetilde{O}(\sqrt{H^3SAT})$ regret. Lower bounds are provided to validate the optimality in both violation and regret performance with respect to $S$ and $T$. Moreover, we further study a novel safe reward-free exploration problem with step-wise violation constraints. For this problem, we design an $(\varepsilon,\delta)$-PAC algorithm SRF-UCRL, which achieves nearly state-of-the-art sample complexity $\widetilde{O}((\frac{S^2AH^2}{\varepsilon}+\frac{H^4SA}{\varepsilon^2})(\log(\frac{1}{\delta})+S))$, and guarantees $\widetilde{O}(\sqrt{ST})$ violation during the exploration. The experimental results demonstrate the superiority of our algorithms in safety performance, and corroborate our theoretical results.
In many applications of Reinforcement Learning (RL), it is critically important that the algorithm performs safely, such that instantaneous hard constraints are satisfied at each step, and unsafe states and actions are avoided. However, existing algorithms for ''safe'' RL are often designed under constraints that either require expected cumulative costs to be bounded or assume all states are safe. Thus, such algorithms could violate instantaneous hard constraints and traverse unsafe states (and actions) in practice. Therefore, in this paper, we develop the first near-optimal safe RL algorithm for episodic Markov Decision Processes with unsafe states and actions under instantaneous hard constraints and the linear mixture model. It not only achieves a regret $\tilde{O}(\frac{d H^3 \sqrt{dK}}{\Delta_c})$ that tightly matches the state-of-the-art regret in the setting with only unsafe actions and nearly matches that in the unconstrained setting, but is also safe at each step, where $d$ is the feature-mapping dimension, $K$ is the number of episodes, $H$ is the number of steps in each episode, and $\Delta_c$ is a safety-related parameter. We also provide a lower bound $\tilde{\Omega}(\max\{dH \sqrt{K}, \frac{H}{\Delta_c^2}\})$, which indicates that the dependency on $\Delta_c$ is necessary. Further, both our algorithm design and regret analysis involve several novel ideas, which may be of independent interest.
Safety has been recognized as the central obstacle to preventing the use of reinforcement learning (RL) for real-world applications. Different methods have been developed to deal with safety concerns in RL. However, learning reliable RL-based solutions usually require a large number of interactions with the environment. Likewise, how to improve the learning efficiency, specifically, how to utilize transfer learning for safe reinforcement learning, has not been well studied. In this work, we propose an adaptive aggregation framework for safety-critical control. Our method comprises two key techniques: 1) we learn to transfer the safety knowledge by aggregating the multiple source tasks and a target task through the attention network; 2) we separate the goal of improving task performance and reducing constraint violations by utilizing a safeguard. Experiment results demonstrate that our algorithm can achieve fewer safety violations while showing better data efficiency compared with several baselines.
Despite the tremendous success of Reinforcement Learning (RL) algorithms in simulation environments, applying RL to real-world applications still faces many challenges. A major concern is safety, in another word, constraint satisfaction. State-wise constraints are one of the most common constraints in real-world applications and one of the most challenging constraints in Safe RL. Enforcing state-wise constraints is necessary and essential to many challenging tasks such as autonomous driving, robot manipulation. This paper provides a comprehensive review of existing approaches that address state-wise constraints in RL. Under the framework of State-wise Constrained Markov Decision Process (SCMDP), we will discuss the connections, differences, and trade-offs of existing approaches in terms of (i) safety guarantee and scalability, (ii) safety and reward performance, and (iii) safety after convergence and during training. We also summarize limitations of current methods and discuss potential future directions.