Extended Event Graph Based Causal Tracing Analysis Of Complex Simulation System

With the application of Simulation technology in military area, as study object, system have been upgraded from system level to system-of-systems (SoS) level, simulation system becomes more and more complicated. In these simulations, there are numerous entities, complicated interactive relations, uncertain effect process. And simulation models are complex, multi-domain and multi-organization. Causal tracing analysis of simulation result only by subjective judgement has become very difficult. Therefore, when getting output data from complex simulation system, how to use causal tracing to gain causal explanation has become one of the most important and urgent problem in simulation system analysis.This thesis uses the SoS combat simulation as the application background. Because causality is multiplex, multi-levels and multi-views, causal tracing analysis is very difficult. Because causality is connotative, complicated and dynamic, building consistent priori causal model is hardly possible. To resolve them, a common, standard and automatic quantity causal tracing analysis theory is put forward. Facing complex simulation system whose characteristic data is event, this theory focuses on system behaviors and process. Based on simulation output data, it can explain simulation results in phenomena and rules by causal data analysis, information analysis and knowledge analysis at different abstraction levels .The contents are summarized into the following parts:(1) Facing the troubles in causal tracing analysis of complex simulation system such as SoS simulation, and referring to common causal tracing analysis theory in logic area, the output response model based causal tracing analysis theory is put forward. This theory builds consistent causal models via behaviors and relations data in simulation output, and makes the analysis process becoming quantificational, standard and more automatic.(2) Output response model (ORM) is the core of tracing analysis. Based on causality form, tracing requirements and characteristics of simulation, considering the event feature data, an output response model -- extended event graph (EEG) from event graph is proposed. The graph uses event node, logic node and causal edge to respectively describe behavior, logic and causal relation, and the causal relation is enriched with time parameter, probability parameter and enabled condition parameter. EEG model is described by formalization specification in three tracing analysis levels: element level, system level and combination level.(3) Simulation data is a bridge between simulation system and causal tracing analysis. To restrict data structure and content, simulation data model is defined. Then based on DEVS, simulation model is redesigned for traceability by adding relative interface to the model specification and simulator. The design is closed under tracing analysis levels.(4) After obtaining simulation data, the primary task is to build EEG. A building process from simulation data to I-Model and from I-Model to II-Model is proposed. In this process causal data is aggregated and abstracted to causal information gradually. And the process includes many model building algorithms, such as I-Model building algorithm, II-Model building algorithm and parameters aggregation algorithms. When models are built, a basal tracing structure is carried out to support the causal tracing process, and two basic causal tracing approaches are put forward based on this tracing structure. The two causal tracing approaches are explanation type tracing and validation type tracing, and they have those own tracing algorithms. (5) Causal knowledge comes from causal information, and it is purposeful and deep abstraction of causal information. Combined with EEG model and causal tracing approaches, many distillation algorithms for causal knowledge are suggested, such as possibility analysis, importance analysis, most probable path analysis, most important path analysis and reduction and compression of event set, etc. Abstract modeling is a process which needs the support by causal knowledge. If combining Fishwick's model abstraction theory with tracing analysis, more causal knowledge and more abundant causal explanation will be obtained. We use two abstract modeling methods—meta-modeling and qualitative discrete event simulation as instances to illustrate how to implement it.The main innovations of this thesis include: proposing the causal tracing analysis based on output response model to resolve tracing problems, carrying subsequent study about this and building a suit of theories and methods all for it; presenting multi-levels EEGs for output response model; proposing several approaches to build EEGs from data to I-Model and from I-Model to II-Model; giving explanation type and validation type causal tracing algorithms; suggesting several causal knowledge distillation algorithms, and introducing model abstraction theory to achieve more causal knowledge.The research of this dissertation focuses on a new theory of causal tracing analysis, and the theory supports causal explanation of simulation results. In analysis of complex simulation system, the theory can be use to understand the equipment operations, to evaluate the effectiveness of equipment objectively. It also may be applied to model validation and simulation validation, model extension and model reuse, model abstraction, and so on. Causal tracing analysis can improve user's ability to resolve problems. It is a kind of common theory, so it can afford a good reference to other areas.