【作者】 刘建国；

【导师】 卢炎生；

【作者基本信息】 华中科技大学 ， 计算机软件与理论， 2009， 博士

【摘要】 近年来,随着计算机科学、认知学、人工心理学等相关学科的发展,虚拟环境、虚拟人、虚拟智能主体、软件人等一系列与人工智能密切相关的几个概念逐步进入到研究者的视野,在一个虚拟的数字化环境中进行虚拟人智能仿真,是人工智能领域继虚拟人行为仿真研究之后的又一研究热点,包括：情感模型、虚拟人认知过程框架、情感对智能的影响、虚拟人行为规划以及虚拟人智能仿真环境等。情感模型是虚拟人高级智能仿真研究的关键技术。该模型以虚拟人的虚拟感受为基础,对虚拟人的需求、感受、动机、行为规划等做出了定义,并探讨了情感在其中的重要作用。虚拟人情感定义为虚拟人主观感受对于需求的满足程度。情感模型由以下几个部分组成,包括虚拟人属性、虚拟人需求模型、虚拟人感受、虚拟人情感价值评估以及情感意图。该模型接受虚拟人的虚拟视觉信息输入,以内部变量的形式产生虚拟感受,以该感受为基础产生情感。考虑到人类复杂的情感,以虚拟人的内部情感模型和外部情感模型来反映这种复杂性。其中内部模型为虚拟人的真实地体验,而外部模型则受到虚拟人其它因素的控制,如智能、个性化因素等,它负责把虚拟人的情感以一种合适的方式表现出来。虚拟人个性化特性则通过情感模型中的个性化参数来体现,包括：感受强度个性化参数、感受敏感性个性化参数、需求满足个性化参数等。情感模型最后输出情感意图。基于情感的虚拟人认知过程建模是实现虚拟人高级智能仿真的关键,动机、记忆和学习处于认知过程的核心。动机模型由情感意图、候选动机、分解动机、焦点动机、获取候选动机算法、动机分析以及动机评价等部分组成。其中情感意图从情感模型获得,通过情感价值评估算法最后得到焦点动机,并以此焦点动机展开行为规划。该认知结构以虚拟人情感为基础,构建了建立在情感模型基础之上的认知结构,和其他的研究者提出的认知过程相比,它更加注重虚拟人的自身体验、自身情感对包括感知、感受、记忆、动机、需求、行为规划、学习等在内的各个认知过程的影响。三阶段的记忆方法对虚拟人的智能决策起到重要作用,虚拟人的感受信息传递到瞬时记忆中进行短暂的存储,经过过滤和处理后的情感状态进入到短时记忆保存,一个动机的实现过程的长时间的情感状态进入长时记忆保存。而瞬时记忆、短时记忆以及长时记忆的情感状态更新的时间也不同。通过一个具体的算法实例,讨论了情感对智能的影响以及三阶段记忆方法的工作过程。由于改进的A*算法在虚拟人路径导航方面存在两个问题：其一是在对场景中地形的利用上,由于不加区分的额外代价将屏蔽掉一些有用地形,从而得不到一条满足最小代价的实际路径；其二是动态的场景中新出现的机会完全被忽略。基于情感的改进A*算法对上述两个方面进行了修正,实验表明,该算法在虚拟人路径导航方面具有优越性。虚拟人的学习能力对提高自主性具有重要意义,学习的最基本形式是条件反射。CRMBBN(Conditional Reflex Model Based Bayesian Network)是一种基于情感模型的采用贝叶斯网络来实现的一种经典条件反射模型。该模型模拟了自然智能体的最为基本的学习形式：经典条件反射,试图使虚拟人也具备这种最为基础的学习能力。CRMBBN通过抛弃反射行为,减小了问题域空间,把复杂、多样的反射行为独立出来,和非条件刺激感觉一起组成刺激-响应对,作为本能行为成为虚拟人的基本行为模式。另外,通过引入感觉,涵盖所有的非条件刺激感觉和条件刺激感觉,并且CRMBBN使用贝叶斯网络来学习条件反射,来表示条件刺激感觉和非条件刺激感觉之间的因果关系,从而使该模型能够根据环境的变化,模拟和环境变化相适应的多个条件反射,能够以更加接近自然智能体的方式模拟Agent的最基本的学习能力。通过试验表明该模型可以完成经典条件反射的几个典型实验比如：习得、遗忘、再习得等。基于情感模型的条件反射模拟是虚拟人学习的基础。在虚拟人智能仿真的实现方面,设计了一种基于BDI模型的Agent体系结构,该体系结构能够实现虚拟人的情感模型、动机模型,并且包含虚拟人的一些本能的反应,是一个适合虚拟人智能仿真的Agent体系结构。在理论研究的基础上,构建一个虚拟人智能仿真平台AlPlayer,并借助于AlPlayer原型系统,进一步验证情感模型、动机模型等理论成果。更多还原

【Abstract】 Recently, with the development of cognitive theories for computer science, artificial psychology as well as some relevant subjects, some concepts such as virtual environment, virtual human and software human and so on that are closely relevant to the artificial intelligence begin to emerge in the minds of researchers. The intelligence simulation with virtual humans in the virtual digitalized environment is considered as another research focus after the behavior simulation in the field of the artificial intelligence, this simulation with virtual human includes emotion model, the framework for cognitive process with virtual human, the influence of emotion to intelligence, behavior planning as well as the intelligent simulation environment and so on.Emotion model is always considered as the core part in the research of the advanced intelligent simulation with virtual human. This model is based on virtual sense, including the definition of virtual human demands, feelings, motivations, behavior planning, and the important role of emotions is discussed. Emotion is defined as a satisfaction level of the individual mood to the demand. Emotional model consists of several components, including the virtual human attributes, virtual human demand model, virtual human feelings, emotions and emotional intention of the valuation. The model accepts virtual human visual information input, in order to generate the internal variables in the form of a virtual experience, in order to generate the basis of emotions. Taking into account the complexity of human emotions, a virtual person’s internal emotional model and the external emotional models is defined to reflect this complexity. The internal model is a virtual human’s true feelings, while the external model is limited by other factors, the control of virtual humans, such as intelligence, personalization factor, it is responsible for the virtual human emotions in an appropriate method. Virtual human’s personalization features is considered through the personalization parameters in this model to reflect the individual, including:feelings of strength of individual parameters, the sensitivity of individual parameters of experience, demands to meet individual parameters. Emotion intention is the output of the model.Modeling for the Cognitive process based on the emotion with virtual human is critical in the course of the advanced intelligent simulation, and intention, memory and study are positioned in the core of this process. Motivation model consists of emotional intent, the candidate motivation, decomposition of motivation, focus motivation, obtaining the candidate motivation algorithms, motivation analysis, and motivation evaluation. The emotion intention is acquired from the emotion model. Focus motivation, finally gained by the evaluation arithmetic for the emotion value, is considered as the focus intention for the behavior planning. Memory models in three stages will play a critical role in the decision-making of intelligence with virtual human, by illustrating a concrete arithmetic case, the influence of the emotion to intelligence as well as the working process of the three-stage memory model is discussed particularly.There are two shortcomings always exist in the process of path-navigation by the improved A* arithmetic, one is the use of the terrain of the situation. For the inappropriate screening of the some useful terrains, a real path with the most cost-effective benefit could not be gained; the other is the full ignoring for new chances in the dynamic situation. The Improved A* arithmetic based on the emotion, however, compensates for the shortcomings stated above, therefore, the test based on this arithmetic has some superiority in the process of path-navigation with virtual human.The study capability with virtual human is of great significance to improve its autonomy. The basic style for that is the condition response. CRMBBN(Conditional Reflex Model Based Bayesian Network), founded on the emotion model and also acquired by the bayes network, is a classical model for condition response. This model can complete several typical experiments such as gain, oblivion, re-gain and so on that is showed by tests definitely, and therefore is a foundation to study for virtual human.With regard to the achievement of the intelligence simulation with virtual human, we have designed a structure of Agent system based on the BDI model. This model, helping us create models for emotion and intention with virtual human and, including some instinct response, therefore is a structure of Agent system that suitable to intelligence simulation with virtual human:an Agent structure based on the smart object (BSOAA). Based on the theory research, we construct a platform for intelligent simulation with virtual human---AIPlayer, and further verify the theory achievements such as emotion model and intention model based on this platform.更多还原