An autonomous system with inner modeled associative memory

具有内部建模联想记忆的自治系统

• 批准号: 13450176
• 负责人: OBAYASHI Masanao
• 金额: $ 4.16万
• 依托单位: Yamaguchi University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (B)
• 财政年份: 2001
• 资助国家: 日本
• 起止时间: 2001 至 2003
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/en/grant/KAKENHI-PROJECT-13450176/
• 关键词: Learning; Chaotic Neural network; Reinforcement Learning; Nonlinear system control; Chaos; Edge of chaos; Associative memory; Robot; 階層型連想記憶; 非線形制御; 免疫システム; ニューラルネットワーク; 海馬連合野モデル; 時系列データ; 1対多; ロボット制御; カオスの辺縁; 関数型記憶行列; 情報生成; 自己組織化

To realize the autonomous system with inner modeled associative memory, it is very useful to research the function of the human brain and learning method of the living things. Under these considerations, we mainly developedi) for reinforcement learning, a faster learning method with asymmetric probability density function to realize trial and error effectivelyii) for reinforcement learning, a learning method introducing the idea of time-varying parameters in order to adapt to dynamical environments which change rapidlyiii) chaotic neural networks with function typed synaptic weights which enable easier retrieval of stored patterns than conventional in the case that each stored patterns have strong correlation each otheriv) a retrieval method for time-series data, particularly for plural time-series data which have same first data but different after thatv) robust chaotic control method introducing the idea of edge of chaos that the system is in the state between chaotic state and non-chaotic statevi) high precision of chaotic time series prediction by introducing the stochastic ascent gradient reinforcement learning method as prediction methodThese validity has been clarified by simulation studies

为了实现具有内模联想记忆的自主系统，研究人类大脑的功能和生物的学习方法是非常有益的。在此考虑下，我们主要针对强化学习开发了一种基于非对称概率密度函数的快速学习方法，以有效实现强化学习的试错。一种引入时变参数思想的学习方法，以适应快速变化的动态环境iii)具有函数型突触权重的混沌神经网络，在每个存储模式彼此具有强相关性的情况下，比传统的存储模式更容易检索iv)时间序列数据的检索方法；特别是对于首数据相同后数据不同的多重时间序列数据，采用鲁棒混沌控制方法，引入混沌边缘的思想，使系统处于混沌状态与非混沌状态之间；引入随机上升梯度强化学习方法作为预测方法，混沌时间序列预测精度高；仿真研究表明，这些方法的有效性得到了验证

项目成果

M.Obayashi: "Associative Memory and Mutual Information In a Chaotic Neural Network Introducing Function Typed Synaptic Weights"Transactions of Institute of Electrical Engineers of Japan. Vol.123, No.9. 1631-1637 (2003)

M.Obayashi：“引入函数类型突触权重的混沌神经网络中的联想记忆和互信息”日本电气工程师学会汇刊。

K.Kobayashi: "A chaotic memory search model based on associative dynamics using features in stored patterns"Proc.of 41^<st> Society of Instrument and Control Engineering Annual Conference. 2919-2924 (2002)

K.Kobayashi：“使用存储模式中的特征，基于关联动力学的混沌记忆搜索模型”Proc.of 41^<st> 仪器与控制工程学会年会。

大林正直, 夕田憲治, 大宮理恵, 小林邦和: "開放型記憶行列を持つカオスニューラルネット連想記憶システムと相互情報量"電気学会論文誌C. 123・9. 1631-1637 (2003)

Masanao Obayashi、Kenji Yuta、Rie Omiya、Kunikazu Kobayashi：“具有开放记忆矩阵和互信息的混沌神经网络联想记忆系统”日本电气工程师学会会刊 C. 123・9 (2003)。

K.Umesako, M.Obayashi, K.Kobayashi: "Evolutionary and time-varying reinforcement learning system based on overlap of rules"Proc.of 6th Japan-France Congress on Mechatronics and 4^<th> Asia-Europe Congress on Mechatronics (JFM2003). 202-207 (2003)

K.Umesako，M.Obayashi，K.Kobayashi：“基于规则重叠的进化和时变强化学习系统”第6届日法机电一体化大会和第4届亚欧机电一体化大会的Proc.（

梅迫公輔, 大林正直, 小林邦和: "時変パラメータを持つ進化的強化学習システム"電気学会論文誌C. 124・7. (2004)

Kosuke Umesako、Masanao Obayashi、Kunikazu Kobayashi：“具有时变参数的进化强化学习系统”，日本电气工程师学会汇刊 C. 124, 7. (2004)