MetaDrive: Composing Diverse Driving Scenarios for Generalizable Reinforcement Learning.

Reinforcement learning (RL) algorithms have been successfully applied to control tasks associated with unmanned aerial vehicles and robotics. In recent years, safe RL has been proposed to allow the safe execution of RL algorithms in industrial and mission-critical systems that operate in closed loops. However, if the system operating conditions change, such as when an unknown fault occurs, typical safe RL algorithms cannot adapt while retaining past knowledge. Continual RL algorithms have been proposed to address this issue. However, the impact of continual adaptation on the system’s safety is an understudied problem. In this paper, we study the intersection of safe and continual RL. First, we empirically demonstrate that a popular continual RL algorithm, elastic weight consolidation, does not satisfy safety constraints in nonlinear systems subject to varying operating conditions. Specifically, we study the MuJoCo HalfCheetah and Ant environments with velocity constraints and sudden joint loss non-stationarity. Then, we show that an agent trained using constrained policy optimization, a safe RL algorithm, experiences catastrophic forgetting in continual learning settings. With this in mind, we explore a simple reward-shaping method to ensure that elastic weight consolidation prioritizes remembering both safety and task performance for safety-constrained, nonlinear, and non-stationary systems.

Self-Preserving Genetic Algorithms (SPGA) combine the evolutionary strategy of a genetic algorithm with safety assurance methods commonly implemented in safe reinforcement learning (SRL), a branch of reinforcement learning (RL) that accounts for safety in the exploration and decision-making process of the agent. Safe learning approaches are especially important in safety-critical environments, where failure to account for the safety of the controlled system could result in the loss of millions of dollars in hardware or bodily harm to people working nearby, as is true of many cyber-physical systems. While SRL is a viable approach to safe learning, there are many challenges that must be taken into consideration when training agents, such as sample efficiency, stability, and exploration---an issue that is easily addressed by the evolutionary strategy of a genetic algorithm. By combining GAs with the safety mechanisms used with SRL, SPGA offers a safe learning alternative that is able to explore large areas of the solution space, addressing SRL's challenge of exploration. This work implements SPGA with both action masking and run time assurance safety strategies to evolve safe controllers for three types of discrete action space environments applicable to cyber physical systems (control, routing, and operations) and under various safety conditions. Training and testing evaluation metrics are compared with results from SRL trained controllers to validate results. SPGA and SRL controllers are trained across 5 random seeds and evaluated on 500 episodes to calculate average wall time to train, average expected return, and percentage of safe action evaluation metrics. SPGA achieves comparable reward and safety performance results with significantly improved training efficiency (55x faster on average), demonstrating the effectiveness of this safe learning approach.

Reinforcement learning (RL) approaches have recently been spotlighted for use in adaptive traffic-signal control on an area-wide level. Most researchers have employed multi-agent reinforcement learning (MARL) algorithms wherein each agent shares a holistic traffic state and cooperates with other agents to reach a common goal. However, MARL algorithms cannot guarantee a global optimal solution unless the actions of all agents are fully coordinated. The present study employs a RL algorithm that recognizes an entire traffic state and jointly controls all the traffic signals of multiple intersections. With this approach, a deep Q-network (DQN) that depends solely on traffic images is extended to overcome the curse of dimensionality that is associated with a large state and action space. Several front layers in a deep convolutional neural network (CNN) to approximate the true Q-function are shared by each intersection approach. Weight parameters connecting the last hidden layer to the output layer are fixed. The proposed methodology outperforms a fixed-signal operation, a fully actuated signal operation, a multi-agent RL control without coordination, and a multi-agent RL control with partial coordination.

This paper investigates the error accumulation problem in centralized training and decentralized execution (CTDE) policy‐based multi‐agent reinforcement learning (MARL) algorithms, which arises from local observation inaccuracies. To address this issue, we propose a novel MARL algorithm that incorporates imitation learning using local observations. Firstly, by analysing the multi‐agent proximal policy optimization algorithm and examining the problems arising when global states are replaced with local observations, it is proved that insufficient observations can lead to information loss, thereby introducing errors of advantage function, and it is demonstrated that the generalized advantage estimation method accumulates errors during the training process. Then, imitation learning is introduced and a novel training framework that combines reinforcement learning and imitation learning is proposed. During the reinforcement learning phase, an MARL agent trained with global observations acts as an expert. Subsequently, imitation learning is applied to train another agent that mimics the expert's decisions using only local observations. Finally, the effectiveness of this algorithm is verified in some commonly used multi‐agent environments, which demonstrates its superior performance compared to traditional multi‐agent reinforcement learning algorithms.

Many decision problems for complex systems that involve multiple decision makers can be formulated as a decentralized partially observable markov decision process (dec-POMDP) problem. Due to the computational difficulty with obtaining optimal policies, recent approaches to dec-POMDP often use a multi-agent reinforcement learning (MARL) algorithm. We propose a method to improve the existing cooperative MARL algorithms by adopting an imitation learning technique. For a reference policy in the imitation learning part, we use a centralized policy from a multi-agent MDP or a multi-agent POMDP model reduced from the original dec-POMDP model. In the proposed method, during the training process, we mix demonstrations from the reference policy by using a demonstration buffer. Demonstration samples from the buffer are used in the augmented policy gradient function for policy updates. We assess the performance of the proposed method for three well-known dec-POMDP benchmark problems -- Mars rover, co-operative box pushing, and dec-tiger. Experimental results indicate that augmenting the baseline MARL algorithm by mixing the demonstrations significantly improves the quality of policy solutions. With these results, we conclude that the imitation learning can enhance MARL algorithms and that policy solutions from MMDP and MPOMDP models are a reasonable reference policy to use in the proposed algorithm.

The widespread deployment of smart meters and communication technologies brings opportunities to improve the adaptability and flexibility of future manufacturing systems under growing complexity and uncertainty. As a result, data-driven approaches, especially Reinforcement Learning (RL) methods, are gaining wide attention in recent years. Although in the current literature, RL approaches show superiority over traditional methods in many applications, a comprehensive review and comparison of different RL methods and their use in discrete manufacturing system control under changing environments have not yet been established. This paper first provides a literature review of RL algorithms for decision-making in discrete manufacturing systems, and then systematically discusses the underlying mechanisms of four most commonly used RL algorithms. In addition, the performance of these RL algorithms is compared by solving a production control problem, which aims to maximize the profit under time-varying production costs. The comparison results show that the single-agent RL methods with discrete action space can provide satisfactory results in small-scale systems, while multi-agent RL algorithms are more suitable to solve problems in complex, large-scale systems to remit the curse of dimensionality. This research sheds light on the impacts of RL method selection on the search for high-quality solutions in manufacturing systems, which is envisioned to drive future research on RL-supported manufacturing systems.

Safe Chance Constrained Reinforcement Learning for Batch Process Optimization and Control

We propose a computationally efficient approach to safe reinforcement learning (RL) for frequency regulation in power systems with high levels of variable renewable energy resources. The approach draws on set-theoretic control techniques to craft a neural network-based control policy that is guaranteed to satisfy safety-critical state constraints, without needing to solve a model predictive control or projection problem in real time. By exploiting the properties of robust controlled-invariant polytopes, we construct a novel, closed-form "safety-filter" that enables end-to-end safe learning using any policy gradient-based RL algorithm. We then apply the safety filter in conjunction with the deep deterministic policy gradient (DDPG) algorithm to regulate frequency in a modified 9-bus power system, and show that the learned policy is more cost-effective than robust linear feedback control techniques while maintaining the same safety guarantee. We also show that the proposed paradigm outperforms DDPG augmented with constraint violation penalties.

Due to the advantages of a centralized critic to estimate the Q-function value and decentralized actors to optimize the agents’ policies, counterfactual multi-agent (COMA) stands out in most multi-agent reinforcement learning (MARL) algorithms. The sharing of policy parameters can improve sampling efficiency and learning effectiveness, but it may lead to a lack of policy diversity. Hence, to balance parameter sharing and diversity among agents in COMA has been a persistent research topic. In this paper, an effective training method for a COMA policy network based on a differential evolution (DE) algorithm is proposed, named DE-COMA. DE-COMA introduces individuals in a population as computational units to construct the policy network with operations such as mutation, crossover, and selection. The average return of DE-COMA is set as the fitness function, and the best individual of policy network will be chosen for the next generation. By maintaining better parameter sharing to enhance parameter diversity, multi-agent strategies will become more exploratory. To validate the effectiveness of DE-COMA, experiments were conducted in the StarCraft II environment with 2s_vs_1sc, 2s3z, 3m, and 8m battle scenarios. Experimental results demonstrate that DE-COMA significantly outperforms the traditional COMA and most other multi-agent reinforcement learning algorithms in terms of win rate and convergence speed.

Unmanned aerial vehicle (UAV) confrontation scenarios play a crucial role in the study of agent behavior selection and decision planning. Multi-agent reinforcement learning (MARL) algorithms serve as a universally effective method guiding agents toward appropriate action strategies. They determine subsequent actions based on the state of the agents and the environmental information that the agents receive. However, traditional MARL settings often result in one party agent consistently outperforming the other party due to superior strategies, or both agents reaching a strategic stalemate with no further improvement. To solve this issue, we propose a semi-static deep deterministic policy gradient algorithm based on MARL. This algorithm employs a centralized training and decentralized execution approach, dynamically adjusting the training intensity based on the comparative strengths and weaknesses of both agents’ strategies. Experimental results show that during the training process, the strategy of the winning team drives the losing team’s strategy to upgrade continuously, and the relationship between the winning team and the losing team keeps changing, thus achieving mutual improvement of the strategies of both teams. The semi-static reinforcement learning algorithm improves the win-loss relationship conversion by 8% and reduces the training time by 40% compared with the traditional reinforcement learning algorithm.

Spacecraft safe proximity, as a critical component of on-orbit servicing missions, primarily encounters the following two challenges: the partial observability of the environment surrounding the service spacecraft and the necessity to evade uncertain obstacles. A safe reinforcement learning algorithm based on a graph neural network is proposed to address the constrained Markov decision problem in partially observable scenarios for spacecraft safe proximity missions. A graph neural network mechanism is introduced to solve the problem of dynamic variations in the quantity and location of obstacles in the observation area of the service spacecraft. The graph attention network is used to facilitate the extraction of feature information from the graph structure, which is then utilized as input for the subsequent reinforcement learning algorithm. The Soft Actor–Critic–Lagrangian algorithm is adopted to deal with the problems of tuning reward function parameters and balancing safety and optimality. By introducing Lagrange multipliers, the constrained optimization problem is transformed into an unconstrained optimization problem. In order to verify the effectiveness of the algorithm proposed in this paper, a spacecraft safe proximity environment model with dynamic obstacles is constructed, and the GAT-SACL algorithm proposed in this paper is validated by the Monte Carlo shooting method. The results show that the GAT-SACL algorithm possess excellent exploratory characteristics and delivers significant advantages in balancing optimality and safety.

Conventional penalty function-based safe reinforcement learning (RL) algorithms often handle safety and policy difference constraints separately, lacking specialized mechanisms to address scenarios where both constraints are violated simultaneously. This limitation sometimes results in severe performance degradation, underscoring the need for further investigation. To address this issue, a simple yet efficient safe RL algorithm is proposed in this work called coupled penalty-based safe policy optimization (CPSPO). CPSPO operates within the first-order local policy search framework, ensuring ease of implementation in practice. To enhance behavior correction when both constraints are violated, CPSPO introduces coupled penalties, considering the cost and Kullback-Leibler (KL)-divergence constraint violation situation simultaneously, effectively improving both constraints handling. Moreover, instead of the conventional policy ratio clipping mechanism, CPSPO directly incorporates the KL-divergence constraint in the loss function design, providing a more intuitive and practical approach to proximal safe policy optimization. Extensive experiments demonstrate that CPSPO outperforms mainstream safe RL algorithms in reward optimization and safety constraint satisfaction.

Multiple unmanned aerial vehicle (multi-UAV) systems have gained significant attention in applications, such as aerial surveillance and search and rescue missions. With the recent development of state-of-the-art multiagent reinforcement learning (MARL) algorithms, it is possible to train multi-UAV systems in collaborative and competitive environments. However, the inherent vulnerabilities of multiagent systems pose significant privacy and security risks when deploying general and conventional MARL algorithms. The presence of even a single Byzantine adversary within the system can severely degrade the learning performance of UAV agents. This work proposes a Byzantine-resilient MARL algorithm that leverages a combination of geometric median consensus and a robust state update model to mitigate, or even eliminate, the influence of Byzantine attacks. To validate its effectiveness and feasibility, the authors include a multi-UAV threat model, provide a guarantee of robustness, and investigate key attack parameters for multiple UAV navigation scenarios. Results from the experiments show that the average rewards during a Byzantine attack increased by up to 60% for the cooperative navigation scenario compared with conventional MARL techniques. The learning rewards generated by the baseline algorithms could not converge during training under these attacks, while the proposed method effectively converged to an optimal solution, proving its viability and correctness.

Traditional multi-agent reinforcement learning (MARL) algorithms typically implement global parameter sharing across various types of heterogeneous agents without meticulously differentiating between different action semantics. This approach results in the action semantic conflict problem, which decreases the generalization ability of policy networks across heterogeneous types of agents and decreases the cooperation among agents in intricate scenarios. Conversely, completely independent agent parameters significantly escalate computational costs and training complexity. To address these challenges, we introduce an adaptive MARL algorithm named Generalized Action-Prediction Optimization (GAPO). First, we introduce the Generalized Action Space (GAS), which represents the union of all agent actions with distinct semantics. All agents first compute their unified representation in the GAS, and then generate their heterogeneous action policies with different available action masks. Second, in order to further improve cooperation between heterogeneous groups, we propose a Cross-Group Prediction (CGP) loss, which adaptively predicts the action policies of other groups by leveraging trajectory information. We integrate the GAPO into both value-based and policy-based MARL algorithms, giving rise to two practical algorithms: G-QMIX and G-MAPPO. Experimental results obtained within the SMAC, MPE, MAMuJoCo, and RPE environments demonstrate the superiority of G-QMIX and G-MAPPO over several state-of-the-art MARL methods, validating the effectiveness of our proposed adaptive generalized MARL approach.

This paper compares and investigates single-agent reinforcement learning (RL) algorithms on the simple and an extended taxi problem domain, and multiagent RL algorithms on a multiagent extension of the simple taxi problem domain we created. In particular, we extend the Policy Hill Climbing (PHC) and the Win or Learn Fast-PHC (WoLF-PHC) algorithms by combining them with the MAXQ hierarchical decomposition and investigate their efficiency. The results are very promising for the multiagent domain as they indicate that these two newly-created algorithms are the most efficient ones from the algorithms we compared.