Preferential training of neurodynamical model based on predictability of target dynamics

Shun Nishide, Harumitsu Nobuta, Hiroshi G. Okuno, Tetsuya Ogata

    研究成果: Article

    1 引用 (Scopus)

    抄録

    Intrinsic motivation is one of the keys in implementing the mechanism of interest to robots. In this paper, we present a method to apply intrinsic motivation in dynamics learning with predictable and unpredictable targets in view. The robots arm is used for the predictable target and the humans arm is used for the unpredictable target in the experiment. The learning algorithm based on intrinsic motivation will automatically set a larger weight to targets that would contribute to decreasing the training error, while setting a smaller weight to others. A neurodynamical model, namely multiple timescales recurrent neural network (MTRNN), is utilized for studying the robots arm/external object dynamics. Training of MTRNN is done using the back propagation through time (BPTT) algorithm. We modify the BPTT algorithm by the following two steps. (1) Evaluate predictability of robots arm/objects using training error of MTRNN. (2) Assign a preference ratio, which represents the weight of the training, to each object based on predictability. The proposed training method would focus more on reducing training error of predictable objects compared to normal BPTT, where training error is equally treated for every object. Experiments were conducted using an actual robot platform, moving the robots arm while a human moves his arm in the robots camera view. The results of the experiment showed that the proposed training method could achieve smaller training error of the robots arm visuomotor dynamics, which is predictable from the robots motor command, compared to general training with BPTT. Evaluation of MTRNN as a forward model to predict untrained data showed that the proposed model is capable of predicting the robots hand motion, specifically with larger number of nodes.

    元の言語English
    ページ(範囲)587-596
    ページ数10
    ジャーナルAdvanced Robotics
    29
    発行部数9
    DOI
    出版物ステータスPublished - 2015 5 3

    Fingerprint

    Robots
    Recurrent neural networks
    Backpropagation
    Experiments
    End effectors
    Learning algorithms
    Cameras

    ASJC Scopus subject areas

    • Control and Systems Engineering
    • Human-Computer Interaction
    • Computer Science Applications
    • Hardware and Architecture
    • Software

    これを引用

    Preferential training of neurodynamical model based on predictability of target dynamics. / Nishide, Shun; Nobuta, Harumitsu; Okuno, Hiroshi G.; Ogata, Tetsuya.

    :: Advanced Robotics, 巻 29, 番号 9, 03.05.2015, p. 587-596.

    研究成果: Article

    @article{dd60fed54bb84fd4944c545cb87cd878,
    title = "Preferential training of neurodynamical model based on predictability of target dynamics",
    abstract = "Intrinsic motivation is one of the keys in implementing the mechanism of interest to robots. In this paper, we present a method to apply intrinsic motivation in dynamics learning with predictable and unpredictable targets in view. The robots arm is used for the predictable target and the humans arm is used for the unpredictable target in the experiment. The learning algorithm based on intrinsic motivation will automatically set a larger weight to targets that would contribute to decreasing the training error, while setting a smaller weight to others. A neurodynamical model, namely multiple timescales recurrent neural network (MTRNN), is utilized for studying the robots arm/external object dynamics. Training of MTRNN is done using the back propagation through time (BPTT) algorithm. We modify the BPTT algorithm by the following two steps. (1) Evaluate predictability of robots arm/objects using training error of MTRNN. (2) Assign a preference ratio, which represents the weight of the training, to each object based on predictability. The proposed training method would focus more on reducing training error of predictable objects compared to normal BPTT, where training error is equally treated for every object. Experiments were conducted using an actual robot platform, moving the robots arm while a human moves his arm in the robots camera view. The results of the experiment showed that the proposed training method could achieve smaller training error of the robots arm visuomotor dynamics, which is predictable from the robots motor command, compared to general training with BPTT. Evaluation of MTRNN as a forward model to predict untrained data showed that the proposed model is capable of predicting the robots hand motion, specifically with larger number of nodes.",
    keywords = "cognitive developmental robotics, intrinsic motivation, recurrent neural network",
    author = "Shun Nishide and Harumitsu Nobuta and Okuno, {Hiroshi G.} and Tetsuya Ogata",
    year = "2015",
    month = "5",
    day = "3",
    doi = "10.1080/01691864.2015.1031279",
    language = "English",
    volume = "29",
    pages = "587--596",
    journal = "Advanced Robotics",
    issn = "0169-1864",
    publisher = "Taylor and Francis Ltd.",
    number = "9",

    }

    TY - JOUR

    T1 - Preferential training of neurodynamical model based on predictability of target dynamics

    AU - Nishide, Shun

    AU - Nobuta, Harumitsu

    AU - Okuno, Hiroshi G.

    AU - Ogata, Tetsuya

    PY - 2015/5/3

    Y1 - 2015/5/3

    N2 - Intrinsic motivation is one of the keys in implementing the mechanism of interest to robots. In this paper, we present a method to apply intrinsic motivation in dynamics learning with predictable and unpredictable targets in view. The robots arm is used for the predictable target and the humans arm is used for the unpredictable target in the experiment. The learning algorithm based on intrinsic motivation will automatically set a larger weight to targets that would contribute to decreasing the training error, while setting a smaller weight to others. A neurodynamical model, namely multiple timescales recurrent neural network (MTRNN), is utilized for studying the robots arm/external object dynamics. Training of MTRNN is done using the back propagation through time (BPTT) algorithm. We modify the BPTT algorithm by the following two steps. (1) Evaluate predictability of robots arm/objects using training error of MTRNN. (2) Assign a preference ratio, which represents the weight of the training, to each object based on predictability. The proposed training method would focus more on reducing training error of predictable objects compared to normal BPTT, where training error is equally treated for every object. Experiments were conducted using an actual robot platform, moving the robots arm while a human moves his arm in the robots camera view. The results of the experiment showed that the proposed training method could achieve smaller training error of the robots arm visuomotor dynamics, which is predictable from the robots motor command, compared to general training with BPTT. Evaluation of MTRNN as a forward model to predict untrained data showed that the proposed model is capable of predicting the robots hand motion, specifically with larger number of nodes.

    AB - Intrinsic motivation is one of the keys in implementing the mechanism of interest to robots. In this paper, we present a method to apply intrinsic motivation in dynamics learning with predictable and unpredictable targets in view. The robots arm is used for the predictable target and the humans arm is used for the unpredictable target in the experiment. The learning algorithm based on intrinsic motivation will automatically set a larger weight to targets that would contribute to decreasing the training error, while setting a smaller weight to others. A neurodynamical model, namely multiple timescales recurrent neural network (MTRNN), is utilized for studying the robots arm/external object dynamics. Training of MTRNN is done using the back propagation through time (BPTT) algorithm. We modify the BPTT algorithm by the following two steps. (1) Evaluate predictability of robots arm/objects using training error of MTRNN. (2) Assign a preference ratio, which represents the weight of the training, to each object based on predictability. The proposed training method would focus more on reducing training error of predictable objects compared to normal BPTT, where training error is equally treated for every object. Experiments were conducted using an actual robot platform, moving the robots arm while a human moves his arm in the robots camera view. The results of the experiment showed that the proposed training method could achieve smaller training error of the robots arm visuomotor dynamics, which is predictable from the robots motor command, compared to general training with BPTT. Evaluation of MTRNN as a forward model to predict untrained data showed that the proposed model is capable of predicting the robots hand motion, specifically with larger number of nodes.

    KW - cognitive developmental robotics

    KW - intrinsic motivation

    KW - recurrent neural network

    UR - http://www.scopus.com/inward/record.url?scp=84930013670&partnerID=8YFLogxK

    UR - http://www.scopus.com/inward/citedby.url?scp=84930013670&partnerID=8YFLogxK

    U2 - 10.1080/01691864.2015.1031279

    DO - 10.1080/01691864.2015.1031279

    M3 - Article

    VL - 29

    SP - 587

    EP - 596

    JO - Advanced Robotics

    JF - Advanced Robotics

    SN - 0169-1864

    IS - 9

    ER -