{"title": "Primitive Manipulation Learning with Connectionism", "book": "Advances in Neural Information Processing Systems", "page_first": 889, "page_last": 895, "abstract": null, "full_text": "Primitive Manipulation Learning with \n\nConnectionism \n\nYoky  Matsuoka \n\nThe Artificial Intelligence  Laboratory \n\nMassachusetts  Institute of Techonology \n\nNE43-819 \n\nCambridge,  MA  02139 \n\nAbstract \n\nInfants' manipulative exploratory behavior within the environment \nis a vehicle of cognitive stimulation[McCall 1974].  During this time, \ninfants practice and perfect sensorimotor patterns that become be(cid:173)\nhavioral modules which will be seriated and imbedded in more com(cid:173)\nplex actions.  This paper explores the development of such primitive \nlearning systems  using  an  embodied  light-weight  hand  which  will \nbe  used for  a  humanoid being  developed  at the  MIT Artificial  In(cid:173)\ntelligence  Laboratory[Brooks and Stein  1993].  Primitive grasping \nprocedures  are  learned  from  sensory  inputs  using  a  connectionist \nreinforcement  algorithm while two submodules preprocess  sensory \ndata to  recognize  the  hardness  of objects  and  detect  shear  using \ncompetitive  learning  and  back-propagation  algorithm  strategies, \nrespectively.  This  system  is  not  only  consistent  and  quick  dur(cid:173)\ning the initial learning stage,  but also adaptable to new  situations \nafter  training is completed. \n\n1 \n\nINTRODUCTION \n\nLearning manipulation in an unpredictable, changing environment is a complex task. \nIt requires  a  nonlinear  controller  to respond  in  a  nonlinear system  that contains  a \nsignificant amount of sensory  inputs and noise [Miller , et  al  1990].  Investigating the \nhuman manipulation learning system and implementing it in  a physical system has \nnot  been  done  due  to its complexity and too many unknown parameters.  Conven(cid:173)\ntional  adaptive  control  theory  assumes  too  many  parameters  that  are  constantly \nchanging in a  real  environment [Sutton,  et al  1991,  Williams 1988].  For an embod(cid:173)\nied  hand,  even  the  simplest  form  of learning  process  requires  a  more  intelligent \ncontrol  network.  Wiener  [Wiener  1948]  has proposed  the idea of \"Connectionism\" , \nwhich  suggests  that  a  muscle  is  controlled  by  affecting  the  gain  of the  \"efferent-\n\n\f", "award": [], "sourceid": 1150, "authors": [{"given_name": "Yoky", "family_name": "Matsuoka", "institution": null}]}