Autonomous, Harmonious and Purposive Acquisition of Various Functions of Robots by Reinforcement Learning and the Relation to the Intelligence Formation

强化学习自主、协调、有目的地获取机器人各种功能及其与智能形成的关系

• 批准号: 15300064
• 负责人: SHIBATA Katsunari
• 金额: $ 4.16万
• 依托单位: Oita University
• 依托单位国家: 日本
• 项目类别: Grant-in-Aid for Scientific Research (B)
• 财政年份: 2003
• 资助国家: 日本
• 起止时间: 2003 至 2006
• 项目状态: 已结题
• 来源: https://kaken.nii.ac.jp/grant/KAKENHI-PROJECT-15300064/
• 关键词: Reinforcement Learning; Recurrent Neural Network; Symbol; Robot; Spatial Abstraction; Intelligent Exploration; Temporal Abstraction; Emergence of Intelligence; 乗算ニューロン; 抽象化; 予測; 文脈; 決定論的探索; 一様探索; ゲートニューロン; 報酬期待ニューロン; 実用的リカレント学習(PRL); ニューラ

This research was aimed to show that by the learning using the training signals that are derived by reinforcement learning, various functions emerge according to the necessity in a neural network to which sensor signals are directly entered and whose outputs are motor commands. The main fruits are as follows.1. It is said that neural networks are not good at symbol processing. However, it was shown that the output representation of a neural network became binary only by reinforcement learning.2. It was shown that a real robot could learn box-pushing behavior using neural network without giving any informatio a about image processing, image recognition, or the given task.3. It was shown that a real robot could learn to reach an object in some degree even in a quasi-real world where various objects and colorful leaflets exist.4. It was shown that a recurrent neural network trained by reinforcement learning could learn some tasks that are thought to be relevant to the spatial or temporal abstraction.

本研究旨在表明，通过使用强化学习衍生的训练信号进行学习，在直接输入传感器信号并输出运动命令的神经网络中，根据需要出现各种功能。主要的水果如下。据说神经网络不擅长符号处理。然而，只有通过强化学习，神经网络的输出表示才会变成二值。结果表明，机器人可以在不提供任何图像处理、图像识别或给定任务信息的情况下，利用神经网络学习推盒行为。结果表明，即使在存在各种物体和彩色传单的准真实世界中，真正的机器人也能在一定程度上学会够到物体。研究表明，通过强化学习训练的递归神经网络可以学习一些被认为与空间或时间抽象相关的任务。

项目成果

階層型ニューラルネットにおける中間層ての適応的空間再構成と中間層レベルの汎化に基づく知識の継承

分层神经网络中基于中间层自适应空间重构和中间层泛化的知识继承

• 发表时间: 2007
• 期刊: 計測自動制御学会論文集 Vol. 43, No.1
• 作者: 柴田克成;伊藤宏司
• 通讯作者: 伊藤宏司

Learning of Reaching a Colored Object Based on Direct-Vision-Based Reinforcement Learning and Acquired Internal Representation

基于直视强化学习和获得的内部表征的到达彩色物体的学习

• 发表时间: 2004
• 期刊: Proc. of The 9th AROB (Int'l Sympo. on Artificial Life and Robotics) Vol. 2
• 作者: K.Yuki;M.Sugisaka;K.Shibata
• 通讯作者: K.Shibata

An Explanation of Emergence of Reward Expectancy Neurons Usine Reinforcement Learning and Neural Net

使用强化学习和神经网络解释奖励期望神经元的出现

• 发表时间: 2005
• 期刊: Abstract Book of Fourteenth Annual Computational Neuroscience Meeting
• 作者: Shinya Ishii;Munetaka Shidara;Katsunari Shibata
• 通讯作者: Katsunari Shibata

強化学習による探索行動の学習

使用强化学习学习探索行为

• 发表时间: 2005
• 期刊: 計測自動制御学会システム・情報部門学術講演会2005講演論文集
• 作者: Shinya Ishii;Munetaka Shidara;Katsunari Shibata;柴田克成
• 通讯作者: 柴田克成

Discretization of Analog Communication Signals by Noise Addition in Communication Learning

通信学习中通过加噪实现模拟通信信号的离散化

• 发表时间: 2004
• 期刊: Proc. of The 9th AROB (Int'l Sympo. on Artificial Life and Robotics) Vol. 2
• 作者: K.Shibata;M.Nakanishi
• 通讯作者: M.Nakanishi