融合神经元激励机制的机器人情景学习与行为控制  

Robotic Episodic Learning and Behaviour Control Integrated with Neuron Stimulation Mechanism

作　　者：刘冬[1] 丛明[1] 高森[1] 韩晓东[1] 杜宇[2] 

机构地区：[1]大连理工大学机械工程学院,辽宁大连116024 [2]英属哥伦比亚大学机械工程学院

出　　处：《机器人》2014年第5期576-583,共8页Robot

基　　金：中央高校基本科研业务费专项资金资助项目(DUT14RC(3)097);国家863计划资助项目(2013AA040303)

摘　　要：针对不确定环境下机器人行为控制的维数灾难和感知混淆问题,引入神经元激励机制,提出一种情景记忆驱动的马尔可夫决策过程(EM-MDP)以实现机器人对环境经验自主学习,及多源不确定性条件下的行为控制.首先,构建情景记忆模型,并基于认知神经科学提出事件中状态神经元激活及组织机制.其次,基于自适应共振理论(ART)与稀疏分布记忆(SDM)通过Hebbian规则实现情景记忆的自主学习,采用神经元突触势能建立机器人行为控制策略,机器人能够评估过去的事件序列,预测当前状态并规划期望的行为.最后,实验结果验证,该模型框架与控制策略能够实现机器人在普遍场景中的行为控制目标.There are problems of curse of dimensionality and perceptual aliasing in robot behaviour control under uncer- tainty. To solve the problem, a framework called episodic memory-driving Markov decision process （EM-MDP） is proposed by introducing neuron stimulation mechanism, in order to achieve environmental experience self-learning and behaviour con- trol under multi-source uncertainty. Firstly, an episodic memory model is built, and an activation and organization mechanism of state neurons is proposed based on cognitive neuroscience. Secondly, self-learning of episodic memory is realized by u- tilizing adaptive resonance theory （ART） and sparse distributed memory （SDM） through Hebbian rules. A robot behaviour control strategy is established by neuron synaptic potential. Robot can evaluate the past events sequence, predict the current state and plan the desired behaviour. Finally, the experimental results show that the model and control strategy can achieve the objectives of robot behaviour control in universal scenes.

关 键 词：移动机器人 情景记忆 神经元激励 行为控制 

分 类 号：TP242[自动化与计算机技术—检测技术与自动化装置]