Research On Modeling And Inference Of CGF-Oriented Intention Recognition

Creating Computer Generated Force(CGF)is one of the frontier and key techniques in combat simulation domain.To solve the problems that the CGFs in existing systems have poor operating ability and unhumanlike behaviors,we need to improve the cognition level of the CGFs.Intention recognition is an important part of cognition behaviors,CGFs who recognize intentions can “know others as well as themselves” as human beings.In others words,they are able to identify the intentions of opponents and make counter decisions efficiently.Thus,CGFs oritented intention recognition is siginificant to make CGFs more intelligent and human-like.In this paper,the research background,siginificance,state of the art of theories and applications are introduced first.Then,the relations between intention recognition and related theories including automatic planning,parameter estimation and reasoning under uncertainty are analyzied.The universe research framework of intention recognition is also proposed,where the recognizing problem is decomposed into three parts: a)modeling domain problem;b)obtaining behavior parameters;c)intention inference.In the main part of this paper,we combine the research framework and applications,and propose different models and corresponding inferring algorithms for four scenarios in combat simulation.First,to recognize the destination of a maneuvering agent,we propose a Semi-Markov Decision Model(SMDM),and present the maneuvering process on a grid-based map formally by the SMDM.Since the esitimating results of the standard particle filter(SPF)have large variances,we apply Rao-Blackwellised Particle Filter(RBPF)in our paper.In experiments,we design a scenariso where a scouting soldier moves to a predefined destination and gets away from a patrolling vehicle,and use SMDM and RBPF to recognize the destination.The results show that,the SMDM can recognize destinations even when the destinations are changeable and the observations are partially missing.Additionally,it outperforms an extension of the Hidden Markov Model(HMM)in terms of recognition metrics.In the approximate inference,when the number of particles is sufficient,the RBPF can get results with smaller variances and consume less time at the same time than the SPF;when the number of particles is insufficient,the average failure rate of RBPF is lower than that of the SPF.Second,to recognize the multi-agent joint goal,we propose a decentralized partially observable Markov decision model(Dec-POMDM)and use it to model the joint goal recognition problem.In the framework of Dec-POMDM,we make use of the existing multi-agent reinforment learning algorithm to estimate the optimal policies of agents,and exploit a modified marginal filer(MF)to infer the joint goal in the large discrete state space.Our modeling method does not need detailed information of cooperation,and the policy estimating algorithm does not need training dataset and environment information.In experiments,we design a modified capture problem,and use the Dec-POMDM,cooperative co-learning algorithm and MF to recognize the joint goal.The results show that our methods perform well no matter goals are changed or not.The Dec-POMDM outperforms the HMM and the MF is more suitable than the SPF in large discrete state space.Third,to recognize the task plan of an agent,we propose a Logical Hidden Semi-Markov Model(LHSMM).Based on the LHSMM,tactical task plan of a UAV is presented by a graph and the duration of a UAV task is model by discrete Coxian distribution explictly.Additionally,the plan inferring process base on maximum likelihood estimatin is also given.To sovle the evaluation problem in the inference,a logical forward algorithm with duration(LFAD)is also proposed.The experiment results show that our LHSMM and LFDA can recognize the task plan of opponents as well as actions and their targets and parameters.Additionally,we also show LHSMM outperforms the logical hidden Markov model,which proves the effectiveness of modeling duration explicitly.Forth,to recognize the team intention,we propose a logical hierarchical hidden semi-Markov Model(LHHSMM).The LHHSMM can model the team intention efficiently in complex scenarios.To solve the inference problem of LHHSMM,we also propose a logical particl filter(LPF)by comblining the SPF and definitions of LHHSMM.In experiments,we design a scenario where a team cooperately attack a target,and use the LHHSMM and the SPF to recognize the team's target and cooperation mode.The results show that the LHHSMM and LPF can recognize the team intention no matter it is changed or not,and our LHHSMM outperforms a modified logical hierarchical hidden Markov Model,especially when the team intention varies.Additionally,we also prove the modeling duration of team intention explicitly can improve the recognition performances.At last,we draw a conclusion of our research in this paper,and present the problems and theories which should be concerned on in the future.