Traditional Reinforcement Learning Algorithms Research Articles

The Application of Residual Connection-Based State Normalization Method in GAIL

In the domain of reinforcement learning (RL), deriving efficacious state representations and maintaining algorithmic stability are crucial for optimal agent performance. However, the inherent dynamism of state representations often complicates the normalization procedure. To overcome these challenges, we present an innovative RL framework that integrates state normalization techniques with residual connections and incorporates attention mechanisms into generative adversarial imitation learning (GAIL). This combination not only enhances the expressive capability of state representations, thereby improving the agent’s accuracy in state recognition, but also significantly mitigates the common issues of gradient dissipation and explosion. Compared to traditional RL algorithms, GAIL combined with the residual connection-based state normalization method empowers the agent to markedly reduce the exploration duration such that feedback concerning rewards in the current state can be provided in real time. Empirical evaluations demonstrate the superior efficacy of this methodology across various RL environments.

A method for solving reconfiguration blueprints based on multi-agent reinforcement learning

Integrated modular avionics systems primarily achieve system fault tolerance by reconfiguring the system configuration blueprints. In the design of manual reconfiguration, the quality of reconfiguration blueprints is influenced by various unstable factors, leading to a certain degree of uncertainty. The effectiveness of reconfiguration blueprints depends on various factors, including load balancing, the impact of reconfiguration, and the time required for the process. Solving high-quality reconfiguration configuration blueprints can be regarded as a type of multi-objective optimization problem. Traditional algorithms have limitations in solving multi-objective optimization problems. Multi-Agent Reinforcement Learning (MARL) is an important branch in the field of machine learning. It enables the accomplishment of more complex tasks in dynamic real-world scenarios through interaction and decision-making. Combining Multi-Agent Reinforcement Learning algorithms with reconfiguration techniques and utilizing MARL methods to generate blueprints can optimize the quality of blueprints in multiple ways. In this paper, an Improved Value-Decomposition Networks (VDN) based on the average sequential cumulative reward is proposed. By refining the characteristics of the integrated modular avionics system, mathematical models are developed for both the system and the reconfiguration blueprint. The Improved VDN algorithm demonstrates superior convergence characteristics and optimization effects compared with traditional reinforcement learning algorithms such as Q-learning, Deep Q-learning Network (DQN), and VDN. This superiority has been confirmed through experiments involving single and continuous faults.

On Transforming Reinforcement Learning With Transformers: The Development Trajectory.

Transformers, originally devised for natural language processing (NLP), have also produced significant successes in computer vision (CV). Due to their strong expression power, researchers are investigating ways to deploy transformers for reinforcement learning (RL), and transformer-based models have manifested their potential in representative RL benchmarks. In this paper, we collect and dissect recent advances concerning the transformation of RL with transformers (transformer-based RL (TRL)) to explore the development trajectory and future trends of this field. We group the existing developments into two categories: architecture enhancements and trajectory optimizations, and examine the main applications of TRL in robotic manipulation, text-based games (TBGs), navigation, and autonomous driving. Architecture enhancement methods consider how to apply the powerful transformer structure to RL problems under the traditional RL framework, facilitating more precise modeling of agents and environments compared to traditional deep RL techniques. However, these methods are still limited by the inherent defects of traditional RL algorithms, such as bootstrapping and the "deadly triad". Trajectory optimization methods treat RL problems as sequence modeling problems and train a joint state-action model over entire trajectories under the behavior cloning framework; such approaches are able to extract policies from static datasets and fully use the long-sequence modeling capabilities of transformers. Given these advancements, the limitations and challenges in TRL are reviewed and proposals regarding future research directions are discussed. We hope that this survey can provide a detailed introduction to TRL and motivate future research in this rapidly developing field.

A reinforcement learning-based routing algorithm for large street networks

Evacuation planning and emergency routing systems are crucial in saving lives during disasters. Traditional emergency routing systems, despite their best efforts, often struggle to accurately capture the dynamic nature of flood conditions, road closures, and other real-time changes inherent in urban disaster logistics. This paper introduces the ReinforceRouting model, a novel approach to optimizing evacuation routes using reinforcement learning (RL). The model incorporates a unique RL environment that considers multiple criteria, such as traffic conditions, hazardous situations, and the availability of safe routes. The RL agent in this model learns optimal actions through interaction with the environment, receiving feedback in the form of rewards or penalties. The ReinforceRouting model excels in executing prompt and accurate route planning on large road networks, outperforming traditional RL algorithms and shortest-path-based algorithms. A higher safety score and episode reward of the model are demonstrated when compared to these classical methods. This innovative approach to disaster evacuation planning offers a promising avenue for enhancing the efficiency, safety, and reliability of emergency responses in dynamic urban environments.

Prioritized experience replay based deep distributional reinforcement learning for battery operation in microgrids

Reinforcement Learning (RL) provides a pathway for efficiently utilizing the battery storage in a microgrid. However, traditional value-based RL algorithms used in battery management focus on formulating the policies based on the reward expectation rather than its probability distribution. Hence the scheduling strategy is solely based on the expectation of the rewards rather than the distribution. This paper focuses on scheduling strategy based on probability distribution of the rewards which optimally reflects the uncertainties in the incoming dataset. Furthermore, the prioritized experience replay samples of the training experience are used to enhance the quality of the learning by reducing bias. The results are obtained with different variants of distributional RL algorithms like C51, Quantile Regression Deep Q-Network (QR-DQN), Fully Quantizable Function (FQF), Implicit Quantile Networks (IQN) and rainbow. Moreover, the results are compared with the traditional deep Q-learning algorithm with prioritized experienced replay. The convergence results on the training dataset are further analyzed by varying the action spaces, using randomized experience replay and without including the tariff-based action while enforcing the penalties for violating battery SoC limits. The best trained Q-network is tested with different load and PV profiles to obtain the battery operation and costs. The performance of the distributional RL algorithms is analyzed under different schemes of Time of Use (ToU) tariff. QR-DQN with prioritized experience replay has been found to be the best performing algorithm in terms of convergence on the training dataset, with least fluctuation in validation dataset and battery operations during different tariff regimes during the day.

Doubly constrained offline reinforcement learning for learning path recommendation

Learning path recommendation refers to the application of interactive recommendation systems in the field of education, aimed at optimizing learning outcomes while minimizing the workload of learners, teachers, and curriculum designers. Reinforcement Learning (RL) has proven effective in capturing and modeling the complex interactions among course activities, learner behaviors, and educational outcomes. Therefore, combining the two approaches presents endless possibilities for personalized education through the use of interactive recommendation systems in the education domain. However, traditional RL algorithms require extensive interaction with the environment during the training phase. Using unverified recommendation logic in interactions with actual students can give rise to unmanageable problems and hinder effective performance in an educational setting. This is because extrapolation introduces substantial evaluation errors that result in recommendations deviating significantly from the actual educational requirements. To address this limitation, we propose a novel method of offline reinforcement learning called Doubly Constrained deep Q-learning Network (DCQN). This method utilizes two generative models to fit existing student historical interaction data, which in turn, constrains the original policy network to generate new actions based on past interactions, avoiding the occurrence of overestimated actions and reducing extrapolation errors. Empirical results on demonstrate that this approach performs better than existing techniques across D4RL, i.e., datasets for deep data-driven reinforcement learning and real educational datasets.

Mean Field Multi-Agent Reinforcement Learning Method for Area Traffic Signal Control

Reinforcement learning is an effective method for adaptive traffic signal control in urban transportation networks. As the number of training rounds increases, the optimal control strategy is learned, and the learning capabilities of deep neural networks are further enhanced, thereby avoiding the limitations of traditional signal control methods. However, when faced with the sequential decision tasks of regional signal control, it encounters issues such as the curse of dimensionality and environmental non-stationarity. To address the limitations of traditional reinforcement learning algorithms applied to multiple intersections, the mean field theory is applied. This models the traffic signal control problem at multiple intersections within a region as interactions between individual intersections and the average effects of neighboring intersections. By decomposing the Q-function through bilateral estimation between the agent and its neighbors, this method reduces the complexity of interactions between agents while preserving global interactions between the agents. A traffic signal control model based on Mean Field Multi-Agent Reinforcement Learning (MFMARL) was constructed, containing two algorithms: Mean Field Q-Network Area Traffic Signal Control (MFQ-ATSC) and Mean Field Actor-Critic Network Area Traffic Signal Control (MFAC-ATSC). The model was validated using the SUMO simulation platform. The experimental results indicate that across different metrics, such as average speed, the mean field reinforcement learning method outperforms classical signal control methods and several existing approaches.

Safety Constraint-Guided Reinforcement Learning with Linear Temporal Logic

In the context of reinforcement learning (RL), ensuring both safety and performance is crucial, especially in real-world scenarios where mistakes can lead to severe consequences. This study aims to address this challenge by integrating temporal logic constraints into RL algorithms, thereby providing a formal mechanism for safety verification. We employ a combination of theoretical and empirical methods, including the use of temporal logic for formal verification and extensive simulations to validate our approach. Our results demonstrate that the proposed method not only maintains high levels of safety but also achieves comparable performance to traditional RL algorithms. Importantly, our approach fills a critical gap in the existing literature by offering a solution that is both mathematically rigorous and empirically validated. The study concludes that the integration of temporal logic into RL offers a promising avenue for developing algorithms that are both safe and efficient. This work lays the foundation for future research aimed at generalizing this approach to various complex systems and applications.

Optimization of Robotic Arm Grasping through Fractional-Order Deep Deterministic Policy Gradient Algorithm

With the rapid development of robotics technology, robotic arm grasping has gained significant attention in the fields of automation and artificial intelligence. In this study, we propose a fractional-order deep deterministic policy gradient (DDPG) algorithm for optimizing robotic arm grasping tasks. Traditional machine learning algorithms face challenges in handling continuous action spaces, while the DDPG algorithm effectively addresses this issue. In this research, we first review the background and challenges of robotic arm grasping and provide an overview of the application of traditional reinforcement learning algorithms in grasping tasks. Subsequently, we introduce the principles and fundamental ideas of the DDPG algorithm in detail, discussing its potential for optimizing robotic arm grasping. To further enhance the performance of robotic arm grasping, we propose an improved approach based on fractional-order control. Fractional-order control exhibits unique advantages in environmental dynamics modeling and grasp posture optimization, enhancing the robustness and adaptability of robotic arm grasping. Through a series of experiments, we validate the effectiveness and superiority of the fractional-order DDPG algorithm in robotic arm grasping tasks. Our algorithm achieves significant improvements in grasping success rate and stability compared to traditional methods. The experimental results demonstrate that the fractional-order DDPG algorithm is better equipped to handle control challenges in continuous action spaces and optimize the performance of robotic arm grasping tasks.

Data-Driven Load Frequency Control Based on Multi-Agent Reinforcement Learning With Attention Mechanism

With the massive penetration of renewable energy, traditional reinforcement learning algorithms suffer from slow convergence and area control error (ACE) in interconnected power systems. This paper proposes data-driven load frequency control (LFC) based on multi-agent reinforcement learning with attention mechanism in interconnected power systems. It can be divided into two phases; in the centralized training, the agents are trained by an experience replay mechanism; in the decentralized execution, the trained agent automatically regulates the generation power to control the load frequency by real-time access to the grid data in the area. The agent can selectively focus on specific information in the environment by introducing a criticism network with an attention mechanism. The attention mechanism can reduce the training time for reinforcement learning while improving control performance under disturbance. A novel reward function based on a cooperation mechanism is used to score the performance of agent, which can guide the reinforcement learning algorithm to reduce the ACE of each area simultaneously. The proposed method is validated by the IEEE three-area interconnected power system, and it is concluded that the method can reduce the ACE caused by load and renewable power disturbances, and greatly reduce the training time of the algorithm.

A Hybrid Online Off-Policy Reinforcement Learning Agent Framework Supported by Transformers.

Reinforcement learning (RL) is a powerful technique that allows agents to learn optimal decision-making policies through interactions with an environment. However, traditional RL algorithms suffer from several limitations such as the need for large amounts of data and long-term credit assignment, i.e. the problem of determining which actions actually produce a certain reward. Recently, Transformers have shown their capacity to address these constraints in this area of learning in an offline setting. This paper proposes a framework that uses Transformers to enhance the training of online off-policy RL agents and address the challenges described above through self-attention. The proposal introduces a hybrid agent with a mixed policy that combines an online off-policy agent with an offline Transformer agent using the Decision Transformer architecture. By sequentially exchanging the experience replay buffer between the agents, the agent's learning training efficiency is improved in the first iterations and so is the training of Transformer-based RL agents in situations with limited data availability or unknown environments.

Toward personalized decision making for autonomous vehicles: A constrained multi-objective reinforcement learning technique

Reinforcement learning promises to provide a state-of-the-art solution to the decision making problem of autonomous driving. Nonetheless, numerous real-world decision making problems involve balancing multiple conflicting or competing objectives. In addition, passengers may typically prefer to explore diversified driving modes through their specific preferences (i.e., relative importance of different objectives). Taking into account these demands, traditional reinforcement learning algorithms with applications in personalized self-driving vehicles remain challenging. Consequently, here we present a novel constrained multi-objective reinforcement learning technique for personalized decision making in autonomous driving, with the goal of learning a single model for Pareto optimal policies across the space of all possible user preferences. Specifically, a nonlinear constraint incorporating a user-specified preference and a vectorized action–value function is introduced to ensure both diversity in learned decision behaviors and efficient alignment between the user-specified preference and the corresponding optimal policy. Additionally, a constrained multi-objective actor–critic approach is advanced to approximate the Pareto optimal policies for any user-specified preferences while adhering to the nonlinear constraint. Finally, the proposed personalized decision making scheme for autonomous driving is assessed in a highway on-ramp merging scenario with dynamic traffic flows. The results demonstrate the effectiveness of our method by comparing it with classical and state-of-the-art baselines.

Path Planning for Autonomous Vehicles in Unknown Dynamic Environment Based on Deep Reinforcement Learning

Autonomous vehicles can reduce labor power during cargo transportation, and then improve transportation efficiency, for example, the automated guided vehicle (AGV) in the warehouse can improve the operation efficiency. To overcome the limitations of traditional path planning algorithms in unknown environments, such as reliance on high-precision maps, lack of generalization ability, and obstacle avoidance capability, this study focuses on investigating the Deep Q-Network and its derivative algorithm to enhance network and algorithm structures. A new algorithm named APF-D3QNPER is proposed, which combines the action output method of artificial potential field (APF) with the Dueling Double Deep Q Network algorithm, and experience sample rewards are considered in the experience playback portion of the traditional Deep Reinforcement Learning (DRL) algorithm, which enhances the convergence ability of the traditional DRL algorithm. A long short-term memory (LSTM) network is added to the state feature extraction network part to improve its adaptability in unknown environments and enhance its spatiotemporal sensitivity to the environment. The APF-D3QNPER algorithm is compared with mainstream deep reinforcement learning algorithms and traditional path planning algorithms using a robot operating system and the Gazebo simulation platform by conducting experiments. The results demonstrate that the APF-D3QNPER algorithm exhibits excellent generalization abilities in the simulation environment, and the convergence speed, the loss value, the path planning time, and the path planning length of the APF-D3QNPER algorithm are all less than for other algorithms in diverse scenarios.

Hierarchical Maneuver Decision Method Based on PG-Option for UAV Pursuit-Evasion Game

Aiming at the autonomous decision-making problem in an Unmanned aerial vehicle (UAV) pursuit-evasion game, this paper proposes a hierarchical maneuver decision method based on the PG-option. Firstly, considering various situations of the relationship of both sides comprehensively, this paper designs four maneuver decision options: advantage game, quick escape, situation change and quick pursuit, and the four options are trained by Soft Actor-Critic (SAC) to obtain the corresponding meta-policy. In addition, to avoid high dimensions in the state space in the hierarchical model, this paper combines the policy gradient (PG) algorithm with the traditional hierarchical reinforcement learning algorithm based on the option. The PG algorithm is used to train the policy selector as the top-level strategy. Finally, to solve the problem of frequent switching of meta-policies, this paper sets the delay selection of the policy selector and introduces the expert experience to design the termination function of the meta-policies, which improves the flexibility of switching policies. Simulation experiments show that the PG-option algorithm has a good effect on UAV pursuit-evasion game and adapts to various environments by switching corresponding meta-policies according to current situation.

Asynchronous Federated Deep Reinforcement Learning-Based URLLC-Aware Computation Offloading in Space-Assisted Vehicular Networks

Space-assisted vehicular networks (SAVN) provide seamless coverage and on-demand data processing services for user vehicles (UVs). However, ultra-reliable and low-latency communication (URLLC) demands imposed by emerging vehicular applications are hard to be satisfied in SAVN by existing computation offloading techniques. Traditional deep reinforcement learning algorithms are unsuitable for highly dynamic SAVN due to the underutilization of environment observations. An AsynchronouS federaTed deep Q-learning (DQN)-basEd and URLLC-aware cOmputatIon offloaDing algorithm (ASTEROID) is presented in this paper to achieve throughput maximization considering the long-term URLLC constraints. Specifically, we first establish an extreme value theory-based URLLC constraint model. Second, the task offloading and computation resource allocation are decomposed by employing Lyapunov optimization. Finally, an asynchronous federated DQN-based (AF-DQN) algorithm is presented to address the UV-side task offloading problem. The server-side computation resource allocation is settled by an queue backlog-aware algorithm. Simulation results verify that ASTEROID achieves superior throughput and URLLC performances.

Comprehensive Ocean Information-Enabled AUV Motion Planning Based on Reinforcement Learning

Motion planning based on the reinforcement learning algorithms of the autonomous underwater vehicle (AUV) has shown great potential. Motion planning algorithms are primarily utilized for path planning and trajectory-tracking. However, prior studies have been confronted with some limitations. The time-varying ocean current affects algorithmic sampling and AUV motion and then leads to an overestimation error during path planning. In addition, the ocean current makes it easy to fall into local optima during trajectory planning. To address these problems, this paper presents a reinforcement learning-based motion planning algorithm with comprehensive ocean information (RLBMPA-COI). First, we introduce real ocean data to construct a time-varying ocean current motion model. Then, comprehensive ocean information and AUV motion position are introduced, and the objective function is optimized in the state-action value network to reduce overestimation errors. Finally, state transfer and reward functions are designed based on real ocean current data to achieve multi-objective path planning and adaptive event triggering in trajectorytracking to improve robustness and adaptability. The numerical simulation results show that the proposed algorithm has a better path planning ability and a more robust trajectory-tracking effect than those of traditional reinforcement learning algorithms.

A multi-agent reinforcement learning model for inventory transshipments under supply chain disruption

The COVID-19 pandemic has significantly disrupted global Supply Chains (SCs), emphasizing the importance of SC resilience, which refers to the ability of SCs to return to their original or more desirable state following disruptions. This study focuses on collaboration, a key component of SC resilience, and proposes a novel collaborative structure that incorporates a fictitious agent to manage inventory transshipment decisions between retailers in a centralized manner while maintaining the retailers’ autonomy in ordering. The proposed collaborative structure offers the following advantages from SC resilience and operational perspectives: (i) it facilitates decision synchronization for enhanced collaboration among retailers, and (ii) it allows retailers to collaborate without the need for information sharing, addressing the potential issue of information sharing reluctance. Additionally, this study employs non-stationary probability to capture the deeply uncertain nature of the ripple effect and the highly volatile customer demand caused by the pandemic. A new Reinforcement Learning (RL) algorithm is developed to handle non-stationary environments and to implement the proposed collaborative structure. Experimental results demonstrate that the proposed collaborative structure using the new RL algorithm achieves superior SC resilience compared with centralized inventory management systems with transshipment and decentralized inventory management systems without transshipment using traditional RL algorithms.

Toward Heterogeneous Environment: Lyapunov-Orientated ImpHetero Reinforcement Learning for Task Offloading

Task offloading combined with reinforcement learning (RL) is a promising research direction in edge computing. However, the intractability in the training of RL and the heterogeneity of network devices have hindered the application of RL in large-scale networks. Moreover, traditional RL algorithms lack mechanisms to share information effectively in a heterogeneous environment, which makes it more difficult for RL algorithms to converge due to the lack of global information. This article focuses on the task offloading problem in a heterogeneous environment. First, we give a formalized representation of the Lyapunov function to normalize both data and virtual energy queue operations. Subsequently, we jointly consider the computing rate and energy consumption in task offloading and then derive the optimization target leveraging Lyapunov optimization. A Deep Deterministic Policy Gradient(DDPG)-based multiple continuous variable decision model is proposed to make the optimal offloading decision in edge computing. Considering the heterogeneous environment, we improve Hetero Federated Learning (HFL) by introducing Kullback-Leibler (KL) divergence to accelerate the convergence of our DDPG based model. Experiments demonstrate that our algorithm accelerates the search for the optimal task offloading decision in heterogeneous environment.

Fast-Convergence Reinforcement Learning for Routing in LEO Satellite Networks.

Fast convergence routing is a critical issue for Low Earth Orbit (LEO) constellation networks because these networks have dynamic topology changes, and transmission requirements can vary over time. However, most of the previous research has focused on the Open Shortest Path First (OSPF) routing algorithm, which is not well-suited to handle the frequent changes in the link state of the LEO satellite network. In this regard, we propose a Fast-Convergence Reinforcement Learning Satellite Routing Algorithm (FRL-SR) for LEO satellite networks, where the satellite can quickly obtain the network link status and adjust its routing strategy accordingly. In FRL-SR, each satellite node is considered an agent, and the agent selects the appropriate port for packet forwarding based on its routing policy. Whenever the satellite network state changes, the agent sends "hello" packets to the neighboring nodes to update their routing policy. Compared to traditional reinforcement learning algorithms, FRL-SR can perceive network information faster and converge faster. Additionally, FRL-SR can mask the dynamics of the satellite network topology and adaptively adjust the forwarding strategy based on the link state. The experimental results demonstrate that the proposed FRL-SR algorithm outperforms the Dijkstra algorithm in the performance of average delay, packet arriving ratio, and network load balance.

Sliding-mode surface-based decentralized event-triggered control of partially unknown interconnected nonlinear systems via reinforcement learning

This paper studies a sliding-mode surface (SMS)-based decentralized event-triggered control problem for a class of partially unknown interconnected nonlinear systems within an identifier-critic neural networks framework. By assigning an especial cost function associated with SMS for each nominal subsystem, the original control issue is equivalently converted into designing a number of optimal controllers updating under an event-triggered manner, which considerably saves the communication resources. To derive these optimal event-based control policies, the corresponding Hamilton-Jacobi-Bellman equations are solved via the reinforcement learning algorithm. Different from traditional reinforcement learning algorithms, the identifier-critic network framework can obviate the requirement for the knowledge of system internal dynamics, and remove the errors generated by approximating actor networks. The weight vector in the critic network is updated through the gradient descent method and the experience replay technology, such that the persistence of excitation condition is relaxed. Under the proposed SMS-based decentralized control scheme, the considered system has a faster control response while minimizing the cost function. After that, according to the Lyapunov stability theory, all signals of the interconnected nonlinear systems are strictly proved to be bounded. Finally, the validity of the proposed control scheme is demonstrated via a simulation example.