| In recent years,with the rapid development of computer simulation technology,combat simulation training with the characteristics of low risk,low consumption,controllable process and repeatable experiment has become an important form of military training for military powers.As the supporting platform for simulation training,the combat simulation system simulates the military equipment entities and their behavior of the enemy,us and neutral parties in the real world through the virtual computer generated forces(CGF).The fidelity and intelligence of the CGF behavior model directly affect the credibility of the simulation training effect.The goal of CGF decision behavior modeling is to generate a simulation model that can reflect the decision behavior laws of real forces,such as action planning,tactical selection and emergency disposal.Common CGF decision behavior modeling methods based on knowledge engineering,such as rule script,finite state machine and behavior tree,can be customized on demand and have strong readability.However,there is a "knowledge gap" between modelers and domain experts,which has long modeling cycle,low efficiency and poor generalization ability of generated models.Machine learning methods such as neural network can effectively improve the efficiency of model construction and enhance the adaptability of the model.However,the black box nature of learning methods makes the model poorly interpretable,which is difficult to be understood and checked by modelers and experts,and reduces the credibility of the model.In order to overcome the above shortcomings,based on the behavior tree,a cutting-edge and mainstream knowledge engineering method in the fields of games and robots in recent years,this paper designs an autonomous generation framework of CGF decision model combined with knowledge engineering and various learning methods,and carries out the research on the generation method of interpretable behavior tree decision model for CGF tactical application.In order to solve the problems of low efficiency and poor adaptability of artificial construction of behavior tree,on the one hand,an evolutionary behavior tree algorithm based on static constraints is proposed to efficiently generate behavior tree decision model according to expert domain knowledge.The algorithm designs abstract conditions and action nodes according to domain constraints,and designs behavior structure constraints to avoid generating invalid or redundant behavior tree individuals,reducing search space and improving learning efficiency.On the other hand,a behavior modeling algorithm of ART-Bev network model based on adaptive resonance theory is proposed for the autonomous generation of behavior tree decision model driven by expert example data.The algorithm learns the state action association rules in the form of "if-then" from the expert example data,and then transforms them into a behavior tree that can reflect the internal logic of expert behavior.In the simulation experiment stage,this paper carries out the experiment based on the independently developed military simulation platform TankSim.The results show that the proposed method can generate decision behavior models which meets the needs of users and can adapt to different maps on the basis of expert criterion knowledge;On the basis of the collected sample trajectory data,the proposed method can extract the decision behavior,and generate a behavior model similar to the behavior model generating the sample data.The two methods proposed in this paper can autonomously generate CGF decision behavior models that meet the experts’ expectations and users’ needs based on typical military simulation scenarios,reduce the workload of modelers and domain experts,improve the modeling efficiency of combat simulation decision behavior,and enhance the adaptability and diversity of behavior. |
