Dense Reward Research Articles

Impulsive maneuver strategy for multi-agent orbital pursuit-evasion game under sparse rewards

To address the subjectivity of dense reward designs for the orbital pursuit-evasion game with multiple optimization objectives, this paper proposes the reinforcement learning method with a hierarchical network structure to guide game strategies under sparse rewards. Initially, to overcome the convergence challenges in the reinforcement learning training process under sparse rewards, a hierarchical network structure is proposed based on the hindsight experience replay. Subsequently, considering the strict constraints imposed by orbital dynamics on spacecraft state space, the reachable domain method is introduced to refine the subgoal space in the hierarchical network, further facilitating the achievement of subgoals. Finally, by adopting the centralized training-layered execution approach, a complete multi-agent reinforcement learning method with the hierarchical network structure is established, enabling networks at each level to learn effectively in parallel within sparse reward environments. Numerical simulations indicate that, under the single-agent reinforcement learning framework, the proposed method exhibits superior stability in the late training stage and enhances exploration efficiency in the early stage by 38.89% to 55.56% to the baseline method. Under the multi-agent reinforcement learning framework, as the relative distance decreases, the subgoals generated by the hierarchical network transition from long-term to short-term, aligning with human behavioral logic.

A Deep Reinforcement Learning Method Based on a Transformer Model for the Flexible Job Shop Scheduling Problem

The flexible job shop scheduling problem (FJSSP), which can significantly enhance production efficiency, is a mathematical optimization problem widely applied in modern manufacturing industries. However, due to its NP-hard nature, finding an optimal solution for all scenarios within a reasonable time frame faces serious challenges. This paper proposes a solution that transforms the FJSSP into a Markov Decision Process (MDP) and employs deep reinforcement learning (DRL) techniques for resolution. First, we represent the state features of the scheduling environment using seven feature vectors and utilize a transformer encoder as a feature extraction module to effectively capture the relationships between state features and enhance representation capability. Second, based on the features of the jobs and machines, we design 16 composite dispatching rules from multiple dimensions, including the job completion rate, processing time, waiting time, and manufacturing resource utilization, to achieve flexible and efficient scheduling decisions. Furthermore, we project an intuitive and dense reward function with the objective of minimizing the total idle time of machines. Finally, to verify the performance and feasibility of the algorithm, we evaluate the proposed policy model on the Brandimarte, Hurink, and Dauzere datasets. Our experimental results demonstrate that the proposed framework consistently outperforms traditional dispatching rules, surpasses metaheuristic methods on larger-scale instances, and exceeds the performance of existing DRL-based scheduling methods across most datasets.

Trajectory tracking control based on deep reinforcement learning and ensemble random network distillation for robotic manipulator

Abstract In general, the trajectory tracking of robotic manipulator is exceptionally challenging due to the complex and strongly coupled mechanical architecture. In this paper, precise track control of the robotic manipulator is formulated as a dense reward problem for reinforcement learning(RL). A deep RL(DRL) approach combining the soft actor-critic (SAC) algorithm and ensemble random network distillation (ERND) is proposed to address the tracking control problem for robotic manipulator. Firstly, an ERND model is designed, consisting of a module list of multiple RND models. Each RND model obtains the error by learning the target features and the predicted features of the environment. The resulting error serves as internal rewards that drive the robotic agent to explore unknown and unpredictable environmental states. The ensemble model obtains the total internal reward by summing the internal rewards of each RND model, thereby obtaining more accurately reflecting the characteristics of the manipulator in tracking control tasks and improving control performance. Secondly, combining the SAC algorithm with ERND facilitates more robust exploration capabilities in environments with input saturation and joint angle constraints, thereby enabling faster learning of effective policies and enhancing the performance and efficiency of robotic manipulator tracking control tasks. Finally, the simulation results demonstrate that the robotic manipulator tracking control task is effectively completed in dense reward problems through the combination of the SAC algorithm and ERND.

A train trajectory optimization method based on the safety reinforcement learning with a relaxed dynamic reward

Train trajectory optimization (TTO) is an effective way to address energy consumption in rail transit. Reinforcement learning (RL), an excellent optimization method, has been used to solve TTO problems. Although traditional RL algorithms use penalty functions to restrict the random exploration behavior of agents, they cannot fully guarantee the safety of the process and results. This paper proposes a proximal policy optimization based safety reinforcement learning framework (S-PPO) for the train trajectory optimization, including a safe action rechoosing mechanism (SARM) and a relaxed dynamic reward mechanism (RDRM) combining a relaxed sparse reward and a dynamic dense reward. SARM guarantees that the new states generated by the agent consistently adhere to the environmental security constraints, thereby enhancing sampling efficiency and facilitating algorithm convergence. RDRM makes it easier for agents to obtain successful samples by relaxing time constraints, which also offers a better balance between exploration and exploitation. The experimental results show that S-PPO can significantly improve performance and obtain better train operation trajectories than soft constraint methods, and the convergence process is smoother. Finally, it was demonstrated that S-PPO exhibits good adaptability across various speed limit tracks.

Human-Inspired Meta-Reinforcement Learning Using Bayesian Knowledge and Enhanced Deep Q-Network

Over the last decades, there has been growing interest in research in multiple and interdisciplinary fields of human-AI computing. In particular, approaches integrating human’s perspective and design with reinforcement learning (RL) have received more attention. However, the current research on RL may need to consider its enhancement from human-inspired approaches further. In this work, we focus on enabling a meta-reinforcement learning (meta-RL) agent to achieve adaptation and generalization, according to modeling Markov Decision Processes (MDP) using Bayesian knowledge and analysis. By introducing a novel framework called human-inspired meta-RL (HMRL), we incorporate the agent performing resilient actions to leverage the dynamic dense reward based on the knowledge and prediction of a Bayesian analysis. The proposed framework can make the agent learn generalization and prevent the agent from failing catastrophically. The experimental results show that our approach helps the agent reduce computational costs with learning adaptation. In addition to the system design, we have also extended further algorithmic improvement based on learning within a deep Q-network (DQN) implementations for more complicated future tasks, which compared replay buffers to possibly enhance the optimization process. Finally, we conclude and anticipate that integrating human-inspired meta-RL can enable learning more formulations relating to robustness and scalability, leading to promising directions and more complex AI goals in the future.

Adaptive control for circulating cooling water system using deep reinforcement learning.

Due to the complex internal working process of circulating cooling water systems, most traditional control methods struggle to achieve stable and precise control. Therefore, this paper presents a novel adaptive control structure for the Twin Delayed Deep Deterministic Policy Gradient algorithm, which is based on a reference trajectory model (TD3-RTM). The structure is based on the Markov decision process of the recirculating cooling water system. Initially, the TD3 algorithm is employed to construct a deep reinforcement learning agent. Subsequently, a state space is selected, and a dense reward function is designed, considering the multivariable characteristics of the recirculating cooling water system. The agent updates its network based on different reward values obtained through interactions with the system, thereby gradually aligning the action values with the optimal policy. The TD3-RTM method introduces a reference trajectory model to accelerate the convergence speed of the agent and reduce oscillations and instability in the control system. Subsequently, simulation experiments were conducted in MATLAB/Simulink. The results show that compared to PID, fuzzy PID, DDPG and TD3, the TD3-RTM method improved the transient time in the flow loop by 6.09s, 5.29s, 0.57s, and 0.77s, respectively, and the Integral of Absolute Error(IAE) indexes decreased by 710.54, 335.1, 135.97, and 89.96, respectively, and the transient time in the temperature loop improved by 25.84s, 13.65s, 15.05s, and 0.81s, and the IAE metrics were reduced by 143.9, 59.13, 31.79, and 1.77, respectively. In addition, the overshooting of the TD3-RTM method in the flow loop was reduced by 17.64, 7.79, and 1.29 per cent, respectively, in comparison with the PID, the fuzzy PID, and the TD3.

Deep Model-Based Reinforcement Learning for Predictive Control of Robotic Systems with Dense and Sparse Rewards

Sparse rewards and sample efficiency are open areas of research in the field of reinforcement learning. These problems are especially important when considering applications of reinforcement learning to robotics and other cyber-physical systems. This is so because in these domains many tasks are goal-based and naturally expressed with binary successes and failures, action spaces are large and continuous, and real interactions with the environment are limited. In this work, we propose Deep Value-and-Predictive-Model Control (DVPMC), a model-based predictive reinforcement learning algorithm for continuous control that uses system identification, value function approximation and sampling-based optimization to select actions. The algorithm is evaluated on a dense reward and a sparse reward task. We show that it can match the performance of a predictive control approach to the dense reward problem, and outperforms model-free and model-based learning algorithms on the sparse reward task on the metrics of sample efficiency and performance. We verify the performance of an agent trained in simulation using DVPMC on a real robot playing the reach-avoid game. Video of the experiment can be found here: https://youtu.be/0Q274kcfn4c.

Autonomous navigation of catheters and guidewires in mechanical thrombectomy using inverse reinforcement learning

PurposeAutonomous navigation of catheters and guidewires can enhance endovascular surgery safety and efficacy, reducing procedure times and operator radiation exposure. Integrating tele-operated robotics could widen access to time-sensitive emergency procedures like mechanical thrombectomy (MT). Reinforcement learning (RL) shows potential in endovascular navigation, yet its application encounters challenges without a reward signal. This study explores the viability of autonomous guidewire navigation in MT vasculature using inverse reinforcement learning (IRL) to leverage expert demonstrations.MethodsEmploying the Simulation Open Framework Architecture (SOFA), this study established a simulation-based training and evaluation environment for MT navigation. We used IRL to infer reward functions from expert behaviour when navigating a guidewire and catheter. We utilized the soft actor-critic algorithm to train models with various reward functions and compared their performance in silico.ResultsWe demonstrated feasibility of navigation using IRL. When evaluating single- versus dual-device (i.e. guidewire versus catheter and guidewire) tracking, both methods achieved high success rates of 95% and 96%, respectively. Dual tracking, however, utilized both devices mimicking an expert. A success rate of 100% and procedure time of 22.6 s were obtained when training with a reward function obtained through ‘reward shaping’. This outperformed a dense reward function (96%, 24.9 s) and an IRL-derived reward function (48%, 59.2 s).ConclusionsWe have contributed to the advancement of autonomous endovascular intervention navigation, particularly MT, by effectively employing IRL based on demonstrator expertise. The results underscore the potential of using reward shaping to efficiently train models, offering a promising avenue for enhancing the accessibility and precision of MT procedures. We envisage that future research can extend our methodology to diverse anatomical structures to enhance generalizability.

Optimization of 2D Irregular Packing: Deep Reinforcement Learning with Dense Reward

Optimization of 2D Irregular Packing: Deep Reinforcement Learning with Dense Reward

Unveiling the Significance of Toddler-Inspired Reward Transition in Goal-Oriented Reinforcement Learning

Toddlers evolve from free exploration with sparse feedback to exploiting prior experiences for goal-directed learning with denser rewards. Drawing inspiration from this Toddler-Inspired Reward Transition, we set out to explore the implications of varying reward transitions when incorporated into Reinforcement Learning (RL) tasks. Central to our inquiry is the transition from sparse to potential-based dense rewards, which share optimal strategies regardless of reward changes. Through various experiments, including those in egocentric navigation and robotic arm manipulation tasks, we found that proper reward transitions significantly influence sample efficiency and success rates. Of particular note is the efficacy of the toddler-inspired Sparse-to-Dense (S2D) transition. Beyond these performance metrics, using Cross-Density Visualizer technique, we observed that transitions, especially the S2D, smooth the policy loss landscape, promoting wide minima that enhance generalization in RL models.

Learning strategies for underwater robot autonomous manipulation control

Autonomous manipulation operations represent the high intelligent coordination from robotic vision and control, it is also a symbol of the advances of robotic intelligence. The limitations of visual sensing and the increasingly complex experimental conditions make autonomous manipulation operations more difficult, particularly for deep reinforcement learning methods, which can enhance robotic control intelligence but require a lot of training process. Due to the high-dimensional continuous state space and continuous action space characteristics of underwater operations, this paper adopts a policy-based reinforcement learning method as the foundational approach. To address the issues of instability and low convergence efficiency in traditional policy-based reinforcement learning algorithms during the learning process, this paper proposes a novel policy learning method. This method adopts the Proximal Policy Optimization algorithm (PPOClip) and optimizes it through an actor-critic network. The aim is to improve the stability and effectiveness of convergence in the learning process. In the underwater training environment, a new reward shaping scheme has been designed to address the issue of reward sparsity during the training process. The manually crafted dense reward function is utilized as attractive and repulsive potential functions for goal manipulation and obstacle avoidance. On the highly complex underwater manipulation and training environment, transferred learning algorithm has been established to reduce the training times and compensate the differences between the simulation and experiment. Simulations and tank experiments have verified the performance of the proposed strategy learning method.

Enhancing variational quantum state diagonalization using reinforcement learning techniques

The variational quantum algorithms are crucial for the application of NISQ computers. Such algorithms require short quantum circuits, which are more amenable to implementation on near-term hardware, and many such methods have been developed. One of particular interest is the so-called variational quantum state diagonalization method, which constitutes an important algorithmic subroutine and can be used directly to work with data encoded in quantum states. In particular, it can be applied to discern the features of quantum states, such as entanglement properties of a system, or in quantum machine learning algorithms. In this work, we tackle the problem of designing a very shallow quantum circuit, required in the quantum state diagonalization task, by utilizing reinforcement learning (RL). We use a novel encoding method for the RL-state, a dense reward function, and an ε-greedy policy to achieve this. We demonstrate that the circuits proposed by the RL methods are shallower than the standard variational quantum state diagonalization algorithm and thus can be used in situations where hardware capabilities limit the depth of quantum circuits. The methods we propose in the paper can be readily adapted to address a wide range of variational quantum algorithms.

Performance Analysis of Different Reward Functions in Reinforcement Learning for the Scheduling of Modular Automotive Production Systems

Conventional, linear production lines struggle with the new flexibility requirements of the automotive market. Modular production has the potential to radically improve the production flexibility. However, scheduling modular production systems is still an open research question. Reinforcement learning (RL) is a form of artificial intelligence that shows great potential in scheduling complex modular production systems. Nonetheless, the performance of RL agents heavily depends on their reward function. Designing an optimal reward function is highly complex. This paper addresses this research gap and systematically compares six different reward functions for a variety of different modular production scenarios. In addition, a new learning method using resets is proposed and its performance is compared with the standard learning approach. The results suggest that dense reward functions perform better than a sparse one. However, there are major case-to-case discrepancies. The proposed learning method outperforms the standard learning method by 7 % on average. Nonetheless, the performance difference between different scenarios is larger than with the standard approach.

Informative Trajectory Planning Using Reinforcement Learning for Minimum-Time Exploration of Spatiotemporal Fields.

This article studies the informative trajectory planning problem of an autonomous vehicle for field exploration. In contrast to existing works concerned with maximizing the amount of information about spatial fields, this work considers efficient exploration of spatiotemporal fields with unknown distributions and seeks minimum-time trajectories of the vehicle while respecting a cumulative information constraint. In this work, upon adopting the observability constant as an information measure for expressing the cumulative information constraint, the existence of a minimum-time trajectory is proven under mild conditions. Given the spatiotemporal nature, the problem is modeled as a Markov decision process (MDP), for which a reinforcement learning (RL) algorithm is proposed to learn a continuous planning policy. To accelerate the policy learning, we design a new reward function by leveraging field approximations, which is demonstrated to yield dense rewards. Simulations show that the learned policy can steer the vehicle to achieve an efficient exploration, and it outperforms the commonly-used coverage planning method in terms of exploration time for sufficient cumulative information.

Decomposing user-defined tasks in a reinforcement learning setup using TextWorld.

The current paper proposes a hierarchical reinforcement learning (HRL) method to decompose a complex task into simpler sub-tasks and leverage those to improve the training of an autonomous agent in a simulated environment. For practical reasons (i.e., illustrating purposes, easy implementation, user-friendly interface, and useful functionalities), we employ two Python frameworks called TextWorld and MiniGrid. MiniGrid functions as a 2D simulated representation of the real environment, while TextWorld functions as a high-level abstraction of this simulated environment. Training on this abstraction disentangles manipulation from navigation actions and allows us to design a dense reward function instead of a sparse reward function for the lower-level environment, which, as we show, improves the performance of training. Formal methods are utilized throughout the paper to establish that our algorithm is not prevented from deriving solutions.

DisTop: Discovering a Topological Representation to Learn Diverse and Rewarding Skills

An efficient way for a deep reinforcement learning (DRL) agent to explore can be to learn a set of skills that achieves a uniform distribution of states. Following this, we introduce DisTop, a new model that simultaneously learns diverse skills and focuses on improving rewarding skills. DisTop progressively builds a discrete topology of the environment using an unsupervised contrastive loss, a growing network and a goal-conditioned policy. Using this topology, a state-independent hierarchical policy can select where the agent has to keep discovering skills in the state space and explicitly forget skills unrelated to tasks. In turn, the new set of visited states allows an improved learnt representation and the learning loop continues. Our experiments emphasize that DisTop is agnostic to the ground state representation and that the agent can discover the topology of its environment whether the states are high-dimensional binary data, images, or proprioceptive inputs. We demonstrate that this paradigm is competitive on MuJoCo benchmarks with state-of-the-art algorithms on both single-task dense rewards and diverse skill discovery. By combining these two aspects, we show that DisTop outperforms a state-of-the-art hierarchical reinforcement learning (HRL) algorithm when rewards are sparse. We believe DisTop opens new perspectives by showing that bottom-up skill discovery combined with representation learning can tackle different hard reward settings.

Reinforcement learning to achieve real-time control of triple inverted pendulum

This work uses reinforcement learning (RL) to achieve the first-ever data-driven real-time control of an actual, not simulated, triple inverted pendulum (TIP) in a model-free way. A swing-up control task for the TIP is formulated as a Markov decision process with a dense reward function, then conducted in real time by using a model-free RL approach. To increase the sample efficiency of learning, a structure-aware virtual experience replay (VER) method is proposed; it works together with an off-policy actor-critic algorithm. The VER exploits the geometrically-symmetric property of TIPs to create virtual sample trajectories from measured ones, then uses the resulting multifold augmented dataset to effectively train actor and critic networks during the learning process. These structure-infused training data serve to obtain additional information and hence increase the convergence speed of network learning. We combine the proposed VER with a state-of-the-art actor-critic algorithm, and then validate its effectiveness through numerical simulations. Notably, the inclusion of VER amplifies computational efficiency, slashing the requisite trials, training steps, and overall duration by approximately 66.67%. Finally experiments demonstrate the real-time control capability of the proposed approach on an actual TIP system.

A distributed multi-vehicle pursuit scheme: generative multi-adversarial reinforcement learning

Multi-vehicle pursuit (MVP) is one of the most challenging problems for intelligent traffic management systems due to multi-source heterogeneous data and its mission nature. While many reinforcement learning (RL) algorithms have shown promising abilities for MVP in structured grid-pattern roads, their lack of dynamic and effective traffic awareness limits pursuing efficiency. The sparse reward of pursuing tasks still hinders the optimization of these RL algorithms. Therefore, this paper proposes a distributed generative multi-adversarial RL for MVP (DGMARL-MVP) in urban traffic scenes. In DGMARL-MVP, a generative multi-adversarial network is designed to improve the Bellman equation by generating the potential dense reward, thereby properly guiding strategy optimization of distributed multi-agent RL. Moreover, a graph neural network-based intersecting cognition is proposed to extract integrated features of traffic situations and relationships among agents from multi-source heterogeneous data. These integrated and comprehensive traffic features are used to assist RL decision-making and improve pursuing efficiency. Extensive experimental results show that the DGMARL-MVP can reduce the pursuit time by 5.47% compared with proximal policy optimization and improve the pursuing average success rate up to 85.67%. Codes are open-sourced in Github.

Bi-Dueling DQN Enhanced Two-Stage Scheduling for Augmented Surveillance in Smart EMS

Safety production surveillance is of great significance to industrial operation management. While augmented intelligence of things is demonstrating tremendous potential in industrial applications, the analyzed information offers lots of benefits to the higher-level planning in the enterprise management systems, to further improve the operational efficiency. In this study, a video surveillance system with augmented intelligence of things is considered as a promising solution to enhance the operational efficiency for enterprises. However, the challenge is to process the surveillance video streams as soon as possible without ignoring any emergencies. This issue can be formulated as a two-stage scheduling problem, which is an NP-hard problem that can be integrated with higher-level enterprise systems for operational efficiency improvement. An improved Deep Q-Network (DQN) model with a newly designed prioritized replay scheme, named Bi-Dueling DQN with Prioritized Replay (Bi-DPR), is proposed to solve this two-stage scheduling problem in smart enterprise management system. A dense reward function based on a concrete state representation is designed to tackle the sparse reward challenge and to speed up the convergence in actual large-scale task scheduling process. A prioritized replay scheme is then developed to improve the sampling efficiency, so as to effectively reduce the training time in Deep Reinforcement Learning (DRL) for the optimal two-stage scheduling. The experiment results demonstrated that the proposed approach is able to provide an efficient scheduling policy to resolve the two-stage scheduling problem, while at the same time offer insight information to improve the performance of higher-level smart enterprise management system.

A partially observable multi-ship collision avoidance decision-making model based on deep reinforcement learning

Unmanned ships have drawn widespread attention for their potential to enhance navigational safety, minimize human errors, and improve shipping efficiency. Nevertheless, the complexity and uncertainty of mixed obstacle environments present significant challenges to developing unmanned ships, particularly in collision avoidance decision-making. This paper proposes a new model using the Partially Observable Markov Decision Process (POMDP) to construct a collision avoidance decision-making model in mixed obstacle environments for autonomous ships, which can address the environment’s complexity and uncertainty and improve decision accuracy. An image-state observation method is proposed as images can provide more accurate, rich, and reliable information. A dense reward function is designed to address the issue of sparse rewards in fitting the algorithm. The Proximal Policy Optimization (PPO) algorithm is utilized for model training. Based on this, a route guidance method called the PPO for POMDP with guidelines under dense reward (G-IPOMDP-PPO) is proposed, which can improve training efficiency. Simulations are conducted in various mixed obstacle environments and compared with conventional algorithms. The results show that the proposed model can safely and efficiently make collision avoidance decisions in complex and uncertain environments. This research provides a new solution and theoretical foundation for developing autonomous ships and can be extended to achieving dynamic interactive collision avoidance in mixed obstacle environments.