Future unmanned battles desperately require intelli-gent combat policies,and multi-agent reinforcement learning offers a promising solution.However,due to the complexity of combat operations and large size of the comb...Future unmanned battles desperately require intelli-gent combat policies,and multi-agent reinforcement learning offers a promising solution.However,due to the complexity of combat operations and large size of the combat group,this task suffers from credit assignment problem more than other rein-forcement learning tasks.This study uses reward shaping to relieve the credit assignment problem and improve policy train-ing for the new generation of large-scale unmanned combat operations.We first prove that multiple reward shaping func-tions would not change the Nash Equilibrium in stochastic games,providing theoretical support for their use.According to the characteristics of combat operations,we propose tactical reward shaping(TRS)that comprises maneuver shaping advice and threat assessment-based attack shaping advice.Then,we investigate the effects of different types and combinations of shaping advice on combat policies through experiments.The results show that TRS improves both the efficiency and attack accuracy of combat policies,with the combination of maneuver reward shaping advice and ally-focused attack shaping advice achieving the best performance compared with that of the base-line strategy.展开更多
In the evolutionary game of the same task for groups,the changes in game rules,personal interests,the crowd size,and external supervision cause uncertain effects on individual decision-making and game results.In the M...In the evolutionary game of the same task for groups,the changes in game rules,personal interests,the crowd size,and external supervision cause uncertain effects on individual decision-making and game results.In the Markov decision framework,a single-task multi-decision evolutionary game model based on multi-agent reinforcement learning is proposed to explore the evolutionary rules in the process of a game.The model can improve the result of a evolutionary game and facilitate the completion of the task.First,based on the multi-agent theory,to solve the existing problems in the original model,a negative feedback tax penalty mechanism is proposed to guide the strategy selection of individuals in the group.In addition,in order to evaluate the evolutionary game results of the group in the model,a calculation method of the group intelligence level is defined.Secondly,the Q-learning algorithm is used to improve the guiding effect of the negative feedback tax penalty mechanism.In the model,the selection strategy of the Q-learning algorithm is improved and a bounded rationality evolutionary game strategy is proposed based on the rule of evolutionary games and the consideration of the bounded rationality of individuals.Finally,simulation results show that the proposed model can effectively guide individuals to choose cooperation strategies which are beneficial to task completion and stability under different negative feedback factor values and different group sizes,so as to improve the group intelligence level.展开更多
In this paper, the reinforcement learning method for cooperative multi-agent systems(MAS) with incremental number of agents is studied. The existing multi-agent reinforcement learning approaches deal with the MAS with...In this paper, the reinforcement learning method for cooperative multi-agent systems(MAS) with incremental number of agents is studied. The existing multi-agent reinforcement learning approaches deal with the MAS with a specific number of agents, and can learn well-performed policies. However, if there is an increasing number of agents, the previously learned in may not perform well in the current scenario. The new agents need to learn from scratch to find optimal policies with others,which may slow down the learning speed of the whole team. To solve that problem, in this paper, we propose a new algorithm to take full advantage of the historical knowledge which was learned before, and transfer it from the previous agents to the new agents. Since the previous agents have been trained well in the source environment, they are treated as teacher agents in the target environment. Correspondingly, the new agents are called student agents. To enable the student agents to learn from the teacher agents, we first modify the input nodes of the networks for teacher agents to adapt to the current environment. Then, the teacher agents take the observations of the student agents as input, and output the advised actions and values as supervising information. Finally, the student agents combine the reward from the environment and the supervising information from the teacher agents, and learn the optimal policies with modified loss functions. By taking full advantage of the knowledge of teacher agents, the search space for the student agents will be reduced significantly, which can accelerate the learning speed of the holistic system. The proposed algorithm is verified in some multi-agent simulation environments, and its efficiency has been demonstrated by the experiment results.展开更多
To solve the problem of multi-target hunting by an unmanned surface vehicle(USV)fleet,a hunting algorithm based on multi-agent reinforcement learning is proposed.Firstly,the hunting environment and kinematic model wit...To solve the problem of multi-target hunting by an unmanned surface vehicle(USV)fleet,a hunting algorithm based on multi-agent reinforcement learning is proposed.Firstly,the hunting environment and kinematic model without boundary constraints are built,and the criteria for successful target capture are given.Then,the cooperative hunting problem of a USV fleet is modeled as a decentralized partially observable Markov decision process(Dec-POMDP),and a distributed partially observable multitarget hunting Proximal Policy Optimization(DPOMH-PPO)algorithm applicable to USVs is proposed.In addition,an observation model,a reward function and the action space applicable to multi-target hunting tasks are designed.To deal with the dynamic change of observational feature dimension input by partially observable systems,a feature embedding block is proposed.By combining the two feature compression methods of column-wise max pooling(CMP)and column-wise average-pooling(CAP),observational feature encoding is established.Finally,the centralized training and decentralized execution framework is adopted to complete the training of hunting strategy.Each USV in the fleet shares the same policy and perform actions independently.Simulation experiments have verified the effectiveness of the DPOMH-PPO algorithm in the test scenarios with different numbers of USVs.Moreover,the advantages of the proposed model are comprehensively analyzed from the aspects of algorithm performance,migration effect in task scenarios and self-organization capability after being damaged,the potential deployment and application of DPOMH-PPO in the real environment is verified.展开更多
In general-sum games, taking all agent＇s collective rationality into account, we define agents＇ global objective, and propose a novel multi-agent reinforcement learning（RL） algorithm based on global policy. In eac...In general-sum games, taking all agent＇s collective rationality into account, we define agents＇ global objective, and propose a novel multi-agent reinforcement learning（RL） algorithm based on global policy. In each learning step, all agents commit to select the global policy to achieve the global goal. We prove this learning algorithm converges given certain restrictions on stage games of learned Q values, and show that it has quite lower computation time complexity than already developed multi-agent learning algorithms for general-sum games. An example is analyzed to show the algorithm＇ s merits.展开更多
Mobile CrowdSensing(MCS)is a promising sensing paradigm that recruits users to cooperatively perform sensing tasks.Recently,unmanned aerial vehicles(UAVs)as the powerful sensing devices are used to replace user partic...Mobile CrowdSensing(MCS)is a promising sensing paradigm that recruits users to cooperatively perform sensing tasks.Recently,unmanned aerial vehicles(UAVs)as the powerful sensing devices are used to replace user participation and carry out some special tasks,such as epidemic monitoring and earthquakes rescue.In this paper,we focus on scheduling UAVs to sense the task Point-of-Interests(PoIs)with different frequency coverage requirements.To accomplish the sensing task,the scheduling strategy needs to consider the coverage requirement,geographic fairness and energy charging simultaneously.We consider the complex interaction among UAVs and propose a grouping multi-agent deep reinforcement learning approach(G-MADDPG)to schedule UAVs distributively.G-MADDPG groups all UAVs into some teams by a distance-based clustering algorithm(DCA),then it regards each team as an agent.In this way,G-MADDPG solves the problem that the training time of traditional MADDPG is too long to converge when the number of UAVs is large,and the trade-off between training time and result accuracy could be controlled flexibly by adjusting the number of teams.Extensive simulation results show that our scheduling strategy has better performance compared with three baselines and is flexible in balancing training time and result accuracy.展开更多
Missile interception problem can be regarded as a two-person zero-sum differential games problem,which depends on the solution of Hamilton-Jacobi-Isaacs(HJI)equa-tion.It has been proved impossible to obtain a closed-f...Missile interception problem can be regarded as a two-person zero-sum differential games problem,which depends on the solution of Hamilton-Jacobi-Isaacs(HJI)equa-tion.It has been proved impossible to obtain a closed-form solu-tion due to the nonlinearity of HJI equation,and many iterative algorithms are proposed to solve the HJI equation.Simultane-ous policy updating algorithm(SPUA)is an effective algorithm for solving HJI equation,but it is an on-policy integral reinforce-ment learning(IRL).For online implementation of SPUA,the dis-turbance signals need to be adjustable,which is unrealistic.In this paper,an off-policy IRL algorithm based on SPUA is pro-posed without making use of any knowledge of the systems dynamics.Then,a neural-network based online adaptive critic implementation scheme of the off-policy IRL algorithm is pre-sented.Based on the online off-policy IRL method,a computa-tional intelligence interception guidance(CIIG)law is developed for intercepting high-maneuvering target.As a model-free method,intercepting targets can be achieved through measur-ing system data online.The effectiveness of the CIIG is verified through two missile and target engagement scenarios.展开更多
In real-time strategy(RTS)games,the ability of recognizing other players’goals is important for creating artifical intelligence(AI)players.However,most current goal recognition methods do not take the player’s decep...In real-time strategy(RTS)games,the ability of recognizing other players’goals is important for creating artifical intelligence(AI)players.However,most current goal recognition methods do not take the player’s deceptive behavior into account which often occurs in RTS game scenarios,resulting in poor recognition results.In order to solve this problem,this paper proposes goal recognition for deceptive agent,which is an extended goal recognition method applying the deductive reason method(from general to special)to model the deceptive agent’s behavioral strategy.First of all,the general deceptive behavior model is proposed to abstract features of deception,and then these features are applied to construct a behavior strategy that best matches the deceiver’s historical behavior data by the inverse reinforcement learning(IRL)method.Final,to interfere with the deceptive behavior implementation,we construct a game model to describe the confrontation scenario and the most effective interference measures.展开更多
As an important mechanism in multi-agent interaction,communication can make agents form complex team relationships rather than constitute a simple set of multiple independent agents.However,the existing communication ...As an important mechanism in multi-agent interaction,communication can make agents form complex team relationships rather than constitute a simple set of multiple independent agents.However,the existing communication schemes can bring much timing redundancy and irrelevant messages,which seriously affects their practical application.To solve this problem,this paper proposes a targeted multiagent communication algorithm based on state control(SCTC).The SCTC uses a gating mechanism based on state control to reduce the timing redundancy of communication between agents and determines the interaction relationship between agents and the importance weight of a communication message through a series connection of hard-and self-attention mechanisms,realizing targeted communication message processing.In addition,by minimizing the difference between the fusion message generated from a real communication message of each agent and a fusion message generated from the buffered message,the correctness of the final action choice of the agent is ensured.Our evaluation using a challenging set of Star Craft II benchmarks indicates that the SCTC can significantly improve the learning performance and reduce the communication overhead between agents,thus ensuring better cooperation between agents.展开更多
This paper investigates impulsive orbital attack-defense(AD)games under multiple constraints and victory conditions,involving three spacecraft:attacker,target,and defender.In the AD scenario,the attacker aims to breac...This paper investigates impulsive orbital attack-defense(AD)games under multiple constraints and victory conditions,involving three spacecraft:attacker,target,and defender.In the AD scenario,the attacker aims to breach the defender's interception to rendezvous with the target,while the defender seeks to protect the target by blocking or actively pursuing the attacker.Four different maneuvering constraints and five potential game outcomes are incorporated to more accurately model AD game problems and increase complexity,thereby reducing the effectiveness of traditional methods such as differential games and game-tree searches.To address these challenges,this study proposes a multiagent deep reinforcement learning solution with variable reward functions.Two attack strategies,Direct attack(DA)and Bypass attack(BA),are developed for the attacker,each focusing on different mission priorities.Similarly,two defense strategies,Direct interdiction(DI)and Collinear interdiction(CI),are designed for the defender,each optimizing specific defensive actions through tailored reward functions.Each reward function incorporates both process rewards(e.g.,distance and angle)and outcome rewards,derived from physical principles and validated via geometric analysis.Extensive simulations of four strategy confrontations demonstrate average defensive success rates of 75%for DI vs.DA,40%for DI vs.BA,80%for CI vs.DA,and 70%for CI vs.BA.Results indicate that CI outperforms DI for defenders,while BA outperforms DA for attackers.Moreover,defenders achieve their objectives more effectively under identical maneuvering capabilities.Trajectory evolution analyses further illustrate the effectiveness of the proposed variable reward function-driven strategies.These strategies and analyses offer valuable guidance for practical orbital defense scenarios and lay a foundation for future multi-agent game research.展开更多
This paper comprehensively explores the impulsive on-orbit inspection game problem utilizing reinforcement learning and game training methods.The purpose of the spacecraft is to inspect the entire surface of a non-coo...This paper comprehensively explores the impulsive on-orbit inspection game problem utilizing reinforcement learning and game training methods.The purpose of the spacecraft is to inspect the entire surface of a non-cooperative target with active maneuverability in front lighting.First,the impulsive orbital game problem is formulated as a turn-based sequential game problem.Second,several typical relative orbit transfers are encapsulated into modules to construct a parameterized action space containing discrete modules and continuous parameters,and multi-pass deep Q-networks(MPDQN)algorithm is used to implement autonomous decision-making.Then,a curriculum learning method is used to gradually increase the difficulty of the training scenario.The backtracking proportional self-play training framework is used to enhance the agent’s ability to defeat inconsistent strategies by building a pool of opponents.The behavior variations of the agents during training indicate that the intelligent game system gradually evolves towards an equilibrium situation.The restraint relations between the agents show that the agents steadily improve the strategy.The influence of various factors on game results is tested.展开更多
The multi-agent system is the optimal solution to complex intelligent problems. In accordance with the game theory, the concept of loyalty is introduced to analyze the relationship between agents' individual incom...The multi-agent system is the optimal solution to complex intelligent problems. In accordance with the game theory, the concept of loyalty is introduced to analyze the relationship between agents' individual income and global benefits and build the logical architecture of the multi-agent system. Besides, to verify the feasibility of the method, the cyclic neural network is optimized, the bi-directional coordination network is built as the training network for deep learning, and specific training scenes are simulated as the training background. After a certain number of training iterations, the model can learn simple strategies autonomously. Also,as the training time increases, the complexity of learning strategies rises gradually. Strategies such as obstacle avoidance, firepower distribution and collaborative cover are adopted to demonstrate the achievability of the model. The model is verified to be realizable by the examples of obstacle avoidance, fire distribution and cooperative cover. Under the same resource background, the model exhibits better convergence than other deep learning training networks, and it is not easy to fall into the local endless loop.Furthermore, the ability of the learning strategy is stronger than that of the training model based on rules, which is of great practical values.展开更多
文摘Future unmanned battles desperately require intelli-gent combat policies,and multi-agent reinforcement learning offers a promising solution.However,due to the complexity of combat operations and large size of the combat group,this task suffers from credit assignment problem more than other rein-forcement learning tasks.This study uses reward shaping to relieve the credit assignment problem and improve policy train-ing for the new generation of large-scale unmanned combat operations.We first prove that multiple reward shaping func-tions would not change the Nash Equilibrium in stochastic games,providing theoretical support for their use.According to the characteristics of combat operations,we propose tactical reward shaping(TRS)that comprises maneuver shaping advice and threat assessment-based attack shaping advice.Then,we investigate the effects of different types and combinations of shaping advice on combat policies through experiments.The results show that TRS improves both the efficiency and attack accuracy of combat policies,with the combination of maneuver reward shaping advice and ally-focused attack shaping advice achieving the best performance compared with that of the base-line strategy.
基金supported by the National Key R&D Program of China(2017YFB1400105).
文摘In the evolutionary game of the same task for groups,the changes in game rules,personal interests,the crowd size,and external supervision cause uncertain effects on individual decision-making and game results.In the Markov decision framework,a single-task multi-decision evolutionary game model based on multi-agent reinforcement learning is proposed to explore the evolutionary rules in the process of a game.The model can improve the result of a evolutionary game and facilitate the completion of the task.First,based on the multi-agent theory,to solve the existing problems in the original model,a negative feedback tax penalty mechanism is proposed to guide the strategy selection of individuals in the group.In addition,in order to evaluate the evolutionary game results of the group in the model,a calculation method of the group intelligence level is defined.Secondly,the Q-learning algorithm is used to improve the guiding effect of the negative feedback tax penalty mechanism.In the model,the selection strategy of the Q-learning algorithm is improved and a bounded rationality evolutionary game strategy is proposed based on the rule of evolutionary games and the consideration of the bounded rationality of individuals.Finally,simulation results show that the proposed model can effectively guide individuals to choose cooperation strategies which are beneficial to task completion and stability under different negative feedback factor values and different group sizes,so as to improve the group intelligence level.
基金supported by the National Key R&D Program of China (2018AAA0101400)the National Natural Science Foundation of China (62173251+3 种基金61921004U1713209)the Natural Science Foundation of Jiangsu Province of China (BK20202006)the Guangdong Provincial Key Laboratory of Intelligent Decision and Cooperative Control。
文摘In this paper, the reinforcement learning method for cooperative multi-agent systems(MAS) with incremental number of agents is studied. The existing multi-agent reinforcement learning approaches deal with the MAS with a specific number of agents, and can learn well-performed policies. However, if there is an increasing number of agents, the previously learned in may not perform well in the current scenario. The new agents need to learn from scratch to find optimal policies with others,which may slow down the learning speed of the whole team. To solve that problem, in this paper, we propose a new algorithm to take full advantage of the historical knowledge which was learned before, and transfer it from the previous agents to the new agents. Since the previous agents have been trained well in the source environment, they are treated as teacher agents in the target environment. Correspondingly, the new agents are called student agents. To enable the student agents to learn from the teacher agents, we first modify the input nodes of the networks for teacher agents to adapt to the current environment. Then, the teacher agents take the observations of the student agents as input, and output the advised actions and values as supervising information. Finally, the student agents combine the reward from the environment and the supervising information from the teacher agents, and learn the optimal policies with modified loss functions. By taking full advantage of the knowledge of teacher agents, the search space for the student agents will be reduced significantly, which can accelerate the learning speed of the holistic system. The proposed algorithm is verified in some multi-agent simulation environments, and its efficiency has been demonstrated by the experiment results.
基金financial support from National Natural Science Foundation of China(Grant No.61601491)Natural Science Foundation of Hubei Province,China(Grant No.2018CFC865)Military Research Project of China(-Grant No.YJ2020B117)。
文摘To solve the problem of multi-target hunting by an unmanned surface vehicle(USV)fleet,a hunting algorithm based on multi-agent reinforcement learning is proposed.Firstly,the hunting environment and kinematic model without boundary constraints are built,and the criteria for successful target capture are given.Then,the cooperative hunting problem of a USV fleet is modeled as a decentralized partially observable Markov decision process(Dec-POMDP),and a distributed partially observable multitarget hunting Proximal Policy Optimization(DPOMH-PPO)algorithm applicable to USVs is proposed.In addition,an observation model,a reward function and the action space applicable to multi-target hunting tasks are designed.To deal with the dynamic change of observational feature dimension input by partially observable systems,a feature embedding block is proposed.By combining the two feature compression methods of column-wise max pooling(CMP)and column-wise average-pooling(CAP),observational feature encoding is established.Finally,the centralized training and decentralized execution framework is adopted to complete the training of hunting strategy.Each USV in the fleet shares the same policy and perform actions independently.Simulation experiments have verified the effectiveness of the DPOMH-PPO algorithm in the test scenarios with different numbers of USVs.Moreover,the advantages of the proposed model are comprehensively analyzed from the aspects of algorithm performance,migration effect in task scenarios and self-organization capability after being damaged,the potential deployment and application of DPOMH-PPO in the real environment is verified.
文摘In general-sum games, taking all agent＇s collective rationality into account, we define agents＇ global objective, and propose a novel multi-agent reinforcement learning（RL） algorithm based on global policy. In each learning step, all agents commit to select the global policy to achieve the global goal. We prove this learning algorithm converges given certain restrictions on stage games of learned Q values, and show that it has quite lower computation time complexity than already developed multi-agent learning algorithms for general-sum games. An example is analyzed to show the algorithm＇ s merits.
基金supported by the Innovation Capacity Construction Project of Jilin Development and Reform Commission(2020C017-2)Science and Technology Development Plan Project of Jilin Province(20210201082GX)。
文摘Mobile CrowdSensing(MCS)is a promising sensing paradigm that recruits users to cooperatively perform sensing tasks.Recently,unmanned aerial vehicles(UAVs)as the powerful sensing devices are used to replace user participation and carry out some special tasks,such as epidemic monitoring and earthquakes rescue.In this paper,we focus on scheduling UAVs to sense the task Point-of-Interests(PoIs)with different frequency coverage requirements.To accomplish the sensing task,the scheduling strategy needs to consider the coverage requirement,geographic fairness and energy charging simultaneously.We consider the complex interaction among UAVs and propose a grouping multi-agent deep reinforcement learning approach(G-MADDPG)to schedule UAVs distributively.G-MADDPG groups all UAVs into some teams by a distance-based clustering algorithm(DCA),then it regards each team as an agent.In this way,G-MADDPG solves the problem that the training time of traditional MADDPG is too long to converge when the number of UAVs is large,and the trade-off between training time and result accuracy could be controlled flexibly by adjusting the number of teams.Extensive simulation results show that our scheduling strategy has better performance compared with three baselines and is flexible in balancing training time and result accuracy.
文摘Missile interception problem can be regarded as a two-person zero-sum differential games problem,which depends on the solution of Hamilton-Jacobi-Isaacs(HJI)equa-tion.It has been proved impossible to obtain a closed-form solu-tion due to the nonlinearity of HJI equation,and many iterative algorithms are proposed to solve the HJI equation.Simultane-ous policy updating algorithm(SPUA)is an effective algorithm for solving HJI equation,but it is an on-policy integral reinforce-ment learning(IRL).For online implementation of SPUA,the dis-turbance signals need to be adjustable,which is unrealistic.In this paper,an off-policy IRL algorithm based on SPUA is pro-posed without making use of any knowledge of the systems dynamics.Then,a neural-network based online adaptive critic implementation scheme of the off-policy IRL algorithm is pre-sented.Based on the online off-policy IRL method,a computa-tional intelligence interception guidance(CIIG)law is developed for intercepting high-maneuvering target.As a model-free method,intercepting targets can be achieved through measur-ing system data online.The effectiveness of the CIIG is verified through two missile and target engagement scenarios.
文摘In real-time strategy(RTS)games,the ability of recognizing other players’goals is important for creating artifical intelligence(AI)players.However,most current goal recognition methods do not take the player’s deceptive behavior into account which often occurs in RTS game scenarios,resulting in poor recognition results.In order to solve this problem,this paper proposes goal recognition for deceptive agent,which is an extended goal recognition method applying the deductive reason method(from general to special)to model the deceptive agent’s behavioral strategy.First of all,the general deceptive behavior model is proposed to abstract features of deception,and then these features are applied to construct a behavior strategy that best matches the deceiver’s historical behavior data by the inverse reinforcement learning(IRL)method.Final,to interfere with the deceptive behavior implementation,we construct a game model to describe the confrontation scenario and the most effective interference measures.
文摘As an important mechanism in multi-agent interaction,communication can make agents form complex team relationships rather than constitute a simple set of multiple independent agents.However,the existing communication schemes can bring much timing redundancy and irrelevant messages,which seriously affects their practical application.To solve this problem,this paper proposes a targeted multiagent communication algorithm based on state control(SCTC).The SCTC uses a gating mechanism based on state control to reduce the timing redundancy of communication between agents and determines the interaction relationship between agents and the importance weight of a communication message through a series connection of hard-and self-attention mechanisms,realizing targeted communication message processing.In addition,by minimizing the difference between the fusion message generated from a real communication message of each agent and a fusion message generated from the buffered message,the correctness of the final action choice of the agent is ensured.Our evaluation using a challenging set of Star Craft II benchmarks indicates that the SCTC can significantly improve the learning performance and reduce the communication overhead between agents,thus ensuring better cooperation between agents.
基金supported by National Key R&D Program of China:Gravitational Wave Detection Project(Grant Nos.2021YFC22026,2021YFC2202601,2021YFC2202603)National Natural Science Foundation of China(Grant Nos.12172288 and 12472046)。
文摘This paper investigates impulsive orbital attack-defense(AD)games under multiple constraints and victory conditions,involving three spacecraft:attacker,target,and defender.In the AD scenario,the attacker aims to breach the defender's interception to rendezvous with the target,while the defender seeks to protect the target by blocking or actively pursuing the attacker.Four different maneuvering constraints and five potential game outcomes are incorporated to more accurately model AD game problems and increase complexity,thereby reducing the effectiveness of traditional methods such as differential games and game-tree searches.To address these challenges,this study proposes a multiagent deep reinforcement learning solution with variable reward functions.Two attack strategies,Direct attack(DA)and Bypass attack(BA),are developed for the attacker,each focusing on different mission priorities.Similarly,two defense strategies,Direct interdiction(DI)and Collinear interdiction(CI),are designed for the defender,each optimizing specific defensive actions through tailored reward functions.Each reward function incorporates both process rewards(e.g.,distance and angle)and outcome rewards,derived from physical principles and validated via geometric analysis.Extensive simulations of four strategy confrontations demonstrate average defensive success rates of 75%for DI vs.DA,40%for DI vs.BA,80%for CI vs.DA,and 70%for CI vs.BA.Results indicate that CI outperforms DI for defenders,while BA outperforms DA for attackers.Moreover,defenders achieve their objectives more effectively under identical maneuvering capabilities.Trajectory evolution analyses further illustrate the effectiveness of the proposed variable reward function-driven strategies.These strategies and analyses offer valuable guidance for practical orbital defense scenarios and lay a foundation for future multi-agent game research.
文摘This paper comprehensively explores the impulsive on-orbit inspection game problem utilizing reinforcement learning and game training methods.The purpose of the spacecraft is to inspect the entire surface of a non-cooperative target with active maneuverability in front lighting.First,the impulsive orbital game problem is formulated as a turn-based sequential game problem.Second,several typical relative orbit transfers are encapsulated into modules to construct a parameterized action space containing discrete modules and continuous parameters,and multi-pass deep Q-networks(MPDQN)algorithm is used to implement autonomous decision-making.Then,a curriculum learning method is used to gradually increase the difficulty of the training scenario.The backtracking proportional self-play training framework is used to enhance the agent’s ability to defeat inconsistent strategies by building a pool of opponents.The behavior variations of the agents during training indicate that the intelligent game system gradually evolves towards an equilibrium situation.The restraint relations between the agents show that the agents steadily improve the strategy.The influence of various factors on game results is tested.
基金supported by the National Natural Science Foundation of China(61503407,61806219,61703426,61876189,61703412)the China Postdoctoral Science Foundation(2016 M602996)。
文摘The multi-agent system is the optimal solution to complex intelligent problems. In accordance with the game theory, the concept of loyalty is introduced to analyze the relationship between agents' individual income and global benefits and build the logical architecture of the multi-agent system. Besides, to verify the feasibility of the method, the cyclic neural network is optimized, the bi-directional coordination network is built as the training network for deep learning, and specific training scenes are simulated as the training background. After a certain number of training iterations, the model can learn simple strategies autonomously. Also,as the training time increases, the complexity of learning strategies rises gradually. Strategies such as obstacle avoidance, firepower distribution and collaborative cover are adopted to demonstrate the achievability of the model. The model is verified to be realizable by the examples of obstacle avoidance, fire distribution and cooperative cover. Under the same resource background, the model exhibits better convergence than other deep learning training networks, and it is not easy to fall into the local endless loop.Furthermore, the ability of the learning strategy is stronger than that of the training model based on rules, which is of great practical values.
