【作者】 张玉梅；

【导师】 阮晓钢；

【作者基本信息】 北京工业大学 ， 模式识别与智能系统， 2005， 硕士

【摘要】 生物的诸多技能是在个体发育过程中逐渐发展形成的,是生物的一种认知行为。本文从模拟人或动物的技能学习角度出发,构造了一种面向低级认知行为的结构可生长的认知模型,并将其应用于运动平衡技能获取的过程之中,实现了对倒立摆的平衡控制,取得的主要研究成果如下: （1） 本文依据神经生理学部分研究成果,构造了一种结构可生长的技能认知模型(Cognitive Model with Growing Structure,简称CM_gs)。CM_gs具有模拟生物系统反射弧的简单结构RC,包括传入神经、中枢神经和传出神经,尤其是其中枢神经工作域的网络结构与神经元数量可以增长。CM_gs的认知算法CA包括工作算法WA和组织算法OA:WA通过自组织特征映射的竞争机制,对输入刺激进行模式识别并输出对该刺激的反应;自组织算法OA,要对输入刺激进行自动的模式分类学习,使中枢神经网络自发地找到适合的结构和规模,还要通过对刺激进行锐化或钝化,使每一个神经元所代表区域自组织地逼近其正确响应。（2） CM_gs的认知算法CA中,组织算法OA是实现自组织技能获取的关键步骤。本文结合细胞生长结构算法（Growing Cell Structure,简称GCS）构成生长算法GRA,在神经中枢工作域实现对刺激信号的模式分类,通过新神经元的不断生长,自组织地进行演化。同时CM_gs采用强化Hebb突触修饰的无监督学习机制RHA,实现不同的神经元以最佳方式响应不同性质的信号刺激。而评价机构EA为GRA和RHA提供指导信息,是二者的基础。（3） 本文将CM_gs的结构与算法整合,提出了认知模型在技能学习中的实现方案。CM_gs在与对象或环境的交互过程中,通过“行动—评价—改进”的方式,实现自组织的技能学习。文中进行了仿真实验研究表明,采用CM_gs能够有效地习得针对开环不稳定的二阶系统的控制技能。文中还将CM_gs和模糊控制方案相比较,模糊控制需要事先对经验进行必要的总结,而CM_gs可以在无任何先验知识的情况下对控制技能进行自发的学习,并找到适合的网络结构。（4） 本文将CM_gs应用于运动平衡控制技能学习之中,构成了面向运动平衡技能的认知模型（Cognitive Model to Motor Balance Skill,简称CMMBS）,模型在学习过程中通过任务完成情况的评估,来指导神经中枢网络的生长与神经元联结权值的修正。本文以运动平衡控制的抽象物理模型——倒立摆为对象,应用CMMBS实现倒立摆的自学习控制,仿真实验表明:CMMBS在自治地与环境的交互作用中,可以通过神经系统自身的发育,自组织的发展运动平衡控制技能。更多还原

【Abstract】 Many skills of the biology are acquired gradually as the individual grows up, which is called as cognitive skill acquisition. It is the goal to understand and simulate the cognitive behaviors of natural life and endue these to artificial life in this thesis. A cognitive model with growing structure is presented here, and it is used in the application of motor balance skill learning and the model can control the inverted pendulum availably. The achievements in this thesis can be summarized as follows: （1） According to the achievements of neurophysiology, a cognitive model with growing structure (CM_gs) is proposed in this paper. CM_gs has reflex-like construction （RC） which is similar to the reflex arc of the biologic nerve, and RC is composed by afferent nerve, nerve center and efferent nerve. Especially, the neurons size and their topology can grow constantly. The cognitive algorithm （CA） includes work algorithm （WA） and organizing algorithm （OA）. WA helps the input pattern to finding its correct response region by the competitive rules and makes the right response. OA conducts the pattern classification learning of the input stimulus and makes the network of nerve center to form the suitable size and topology. In the other hand, OA needs to make the neurons in difference fields to respond to the different stimulus with the best value through Sensitization and Habituation. （2） In CA of CM_gs, OA is the key component for the skill acquisition. OA adopts a growing algorithm （GRA） from reference to Growing Cell Structures in order to perform the pattern classification in the nerve center. This growing mechanism is able to be evolved through the continuous growing of the new neuron. Reinforcement Hebbian synaptic modification algorithm （RHA） is used as the self-learning method in CM_gs to make the neurons in difference fields to respond to the different stimulus in the best way. EA presents the constructive information for GRA and RHA. （3） Implementation scheme of CM_gs for the skill learning is proposed in this issue. CM_gs can interact autonomously with the environment and develop the motor skill by the growing manners of neural system itself through the way of “action –evaluation -reinforcement”. An emulate experiment is presented to show the implement process of the skill learning by CM_gs. The results are shown that CM_gs can control a second order system whose open loop response is not stabile. And a comparison with the fuzzy control and CM_gs is maked in the paper. It is shown that Fuzzy control needs to summarize the declarative knowledge before the controlling, but CM_gs can form the spontaneous learning process to achieve the suitable network without any transcendental knowledge. （4） A cognitive model to motor balance skill （CMMBS） is provided, which is based on CM_gs, to achieve the acquisition of motor balance skill. CMMBS can更多还原