Research On Key Technologies Of CGF Tactical Task Planning Behavior Modelling

Computer Generated Forces(CGF)tactical mission planning is a series of planning reasoning and selection decision activities carried out by CGF to achieve its operational objectives in the combat simulation system.It has been the focus and difficulty in building a real and reliable combat simulation system.However,it's very difficult to acquire domain knowledge in modeling the behavior in the current CGF tactical mission planning.Besides,the uncertainty of the confrontation environment poses a significant challenge to model the agent's behavior,the requirement of which can't be satisfied by a simple model.In view of these shortcomings,this thesis aims to study the general framework of tactical task planning behavior modeling in CGF,focusing on the key technologies of tactical mission planning behavior modeling.The main contributions and innovations of the thesis are as follows:(1)By analyzing and studying the difficulty in the uncertain planning process,we design a general CGF tactical mission planning behavior modeling for complex dynamic battlefield environment,to satisfy the requirements of CGF tactical mission planning behavior modeling in the uncertain confrontation environment.In the actual battlefield environment,the mission planning process faces difficulties and uncertainties caused by the uncertain environment,opponent behavior,and action effects.In order to improve the adaptability of planning algorithms to the uncertain environment and ensure the accuracy and effectiveness of planning results,this chapter designs a CGF tactical task planning behavior modeling framework under a dynamic battlefield environment.By comparison with other mission planning in the combat simulation,the characteristics of CGF tactical mission planning behavior modeling problem are first analyzed.The modeling requirements are defined from three aspects: perception information,opponent judgment,and model representation.On the basis of the hierarchical task network planning method,a general CGF tactical task planning behavior modeling framework in the uncertain battlefield environment is further proposed.From the three perspectives of environmental information processing,opponent behavior analysis,and planning result prediction,the components and decision-making process of the framework are detailedly described.Finally,the three key technology modules including the state assessment methods in partially observable environments,adversary behavior modeling methods for hierarchical task networks,and state assessment methods for action prediction are analyzed.Furthermore,their relationships and positioning in the framework are discussed to formulate the requirements and boundaries for the subsequent research.(2)From the perspective of environmental information,two methods,history information based single state generation,and fuzzy theory based single state generation are respectively proposed under the conditions of complete and incomplete history information.The methods are further combined with the adversarial hierarchical task network repairing(AHTNR)algorithm to solve the planning problems in the partially observable environments.In the actual battlefield,due to the fog of the war and the limited scope of perception,the environmental information that can be observed by two sides is often incomplete.In order to effectively solve the state uncertainty problem caused by the incomplete observation information,two state generation methods are proposed.First,the partial observable environment is described in the format of the “state-action” pair and the process of change of the state action is analyzed.Second,on the basis of the belief state built from environmental information set,a single state generation method based on historical information and single belief state generation method based on fuzzy inference are respectively proposed under the conditions of complete and incomplete historical information.Additionally,the two state generation methods are combined with the AHTNR algorithm to form the optimized hierarchical task network planning algorithms that are suitable for partially observable environments.Finally,a partially observable confrontation environment with scouting action and sensing range is constructed and verifies the effectiveness and efficiency of the two proposed planning algorithms.(3)From the perspective of the opponent's behavior,an implicit opponent modeling method for the adversarial hierarchical task network(AHTN)is proposed.A particle filter based strategy matching algorithm is used to model the opponent's behavior,and combined with the AHTN to form a game adversarial hierarchical task network algorithm,which can generate the more targeted plan.The uncertainty of the opponent's behavior is an influential factor in the planning process.In order to ensure the plan is forward-looking,the prediction of the opponent's behavior should be included in the planning process.In the existing intelligent planning algorithm,the opponent modeling is usually simplified and assumed to improve the efficiency of single-step planning.However,the diversity and dynamics of the opponent model are neglected.In order to effectively predict the opponent's behavior,an implicit opponent modeling method based on strategy matching is proposed.First,a particle filter-based strategy matching method is present.Based on the given opponent strategy set,the number of particles representing different strategies is iteratively updated based on the matching degree between their estimated observations and the actual observations until the conditions for obtaining the opponent's stable strategy are met.Second,the strategy matching algorithm is used to construct the implicit opponent model and predict the opponent's next action.Furthermore,a game adversarial hierarchical task network algorithm is proposed by combining the implicit opponent modeling method with the hierarchical task network.Finally,the proposed algorithm is verified under different confrontation environments and different opponent settings.(4)From the perspective of planning result prediction,a dynamic hierarchical evaluation network method is proposed by optimizing the design of evaluation index and weight,and the outcome is predicted based on the multi-scale convolutional neural network to evaluate the effect of the planned action from the overall situation.The evaluation of action effects is to evaluate the execution effect of the current action by predicting and reasoning the changes in the future situation and provide a basis for planning.Considering the impact of environmental dynamics on evaluation,a dynamic hierarchical evaluation network method is proposed.A hierarchical evaluation network is constructed to manage the factors and their correlations and dynamically calculated the weights of factors to achieve the dynamic assessment of the situation.In order to solve the problem of difficulty in extracting evaluation factors and complicated weight design,an offline situation assessment method based on win-and-loss prediction is further proposed.The battlefield environment information at a moment is abstracted into a multi-dimensional information set.Besides,the situation at different moments and the corresponding winning and losing results are extracted,and a multi-time state data set and a single-time state data set are constructed.The Goog Le Net-based multi-scale convolutional neural network model is used to train the data set to accurately predict the outcome of the game.