[1]王东署,杨凯.基于状态转移学习的机器人行为决策认知模型[J].郑州大学学报(工学版),2021,42(6):8-14.[doi:10.13705/j.issn.1671-6833.2021.04.012]
　Wang Dongshu,Yang Kai,Behavior Decision-making Cognitive Model of Mobile Robot Based on State Transfer Learning[J].Journal of Zhengzhou University (Engineering Science),2021,42(6):8-14.[doi:10.13705/j.issn.1671-6833.2021.04.012]

点击复制

基于状态转移学习的机器人行为决策认知模型()

分享到：

《郑州大学学报(工学版)》[ISSN:1671-6833/CN:41-1339/T]

卷:

42

期数:

2021年6期

页码:

8-14

栏目:

出版日期:

2021-11-10

文章信息/Info

Title:

Behavior Decision-making Cognitive Model of Mobile Robot Based on State Transfer Learning

作者:

王东署,杨凯

郑州大学 电气工程学院，河南 郑州 450001

Author(s):

Wang Dongshu; Yang Kai;

School of Electrical Engineering, Zhengzhou University, Zhengzhou 450001, China

关键词:

行为决策; 状态转移; 发育神经网络; 连续学习; 适应性

Keywords:

behavior decision-making;  state transfer;  developmental neural network;  continuous learning;  adaptability

分类号:

TP242. 6

DOI:

10.13705/j.issn.1671-6833.2021.04.012

文献标志码:

A

摘要:

当传统神经网络训练样本不足时，网络对场景的识别错误率较高，且在执行任务的过程中无法连续学习，从而导致传统神经网络对陌生环境的适应性较差。针对这些问题，提出一种仿生的机器人行为决策认知计算模型。该计算模型采用半监督方法和状态转移学习方法，先使用少量训练样本对发育神经网络进行训练，使其具备基本的行为决策能力; 机器人在实际环境中探索时，可以不断学习新的场景数据; 当机器人完成任务时，计算模型会按照某种概率回忆所经历的特定场景，即回放在线执行任务时新学习到的经验数据，并结合状态转移机制，不断调整自身决策效果。这种方法可以使网络模型快速收敛到稳定状态，在未知环境中具有很强的适应性。为了验证模型的可行性，设计了真实的机器人运行环境，使用 ＲIKIＲOBOT 移动机器人来进行导航测试。实验结果表明: 所提方法在未知环境中经过3～ 5次的决策调整即可收敛到稳定状态，且决策效果不断改善。通过不断积累知识，机器人可以应对各种复杂环境，在未知环境中具有很强的适应性。

Abstract:

Due to the small sample size of traditional neural networks, the error rate of the recognition of the scene was very high, and it could not continuously learn during the execution of the task. This would lead to poor adaptability of traditional neural networks to unfamiliar environments. In response to these problems, a bionic robot behavior decision-making cognitive computing model was proposed. The algorithm used semi-supervised and state transition learning methods. Firstly, a small number of training samples were used to train the developmental neural network, so that it could have some basic behavioral decision-making capabilities. When the robot was exploring in the actual environment, it could continuously learn new unlearned scene data. When the robot completed the task, the network model would recall the specific scene it had experienced according to a certain probability, and combined the state transfer mechanism to continuously adjusted its own decision-making effect. This method could make the network model quickly converge to a stable state, and had strong adaptability in unknown environments. In order to verify the feasibility of the model, a real robot operating environment was designed, and the RIKIROBOT was used for navigation testing. Experimental results showed that this developmental model could converge to a stable state after 3 to 5 decision-making adjustments in an unknown environment, and the decision-making effect was continuously improved. Robots could deal with various complex environments by continuously accumulating knowledge, and had strong adaptability in unknown environments.

常用功能