{"title": "Neuronal Group Selection Theory: A Grounding in Robotics", "book": "Advances in Neural Information Processing Systems", "page_first": 308, "page_last": 315, "abstract": null, "full_text": "308 \n\nDonnett and Smithers \n\nNeuronal  Group  Selection  Theory: \n\nA  Grounding in  Robotics \n\nJim Donnett and Tim Smithers \nDepartment of Artificial Intelligence \n\nUniversity of Edinburgh \n\n5  Forrest Hill \n\nEdinburgh  EH12QL \n\nSCOTLAND \n\nABSTRACT \n\nIn this paper, we  discuss a current attempt at applying the organi(cid:173)\nzational principle  Edelman  calls  Neuronal  Group  Selection  to  the \ncontrol of a real,  two-link robotic  manipulator.  We begin by moti(cid:173)\nvating the need for  an alternative to the position-control paradigm \nof classical robotics,  and suggest  that a  possible avenue  is  to look \nat the primitive animal limb 'neurologically ballistic' control mode. \nWe  have  been  considering  a  selectionist  approach to coordinating \na  simple perception-action task. \n\n1  MOTIVATION \nThe majority of industrial robots in the world are mechanical manipUlators - often \narm-like  devices  consisting of some  number  of rigid  links  with  actuators  mounted \nwhere  the  links  join  that  move  adjacent  links  relative  to  each  other,  rotationally \nor  translation ally.  At  the  joints  there  are  typically  also  sensors  measuring  the \nrelative  position of adjacent links,  and it is  in  terms of position that manipulators \nare generally controlled (a desired motion is specified as a desired position of the end \neffector, from  which can be derived  the necessary  positions of the links comprising \nthe  manipulator).  Position control dominates largely for historical reasons,  rooted \nin  bang-bang  control:  manipulators bumped between mechanical stops placed so as \nto enforce a  desired trajectory for  the end effector. \n\n\fNeuronal Group Selection Theory:  A Grounding in Robotics \n\n309 \n\n1.1  SERVOMECHANISMS \n\nMechanical stops  have  been  superceded  by  position-controlling  servomechanisms, \nnegative feedback systems in which, for a typical manipulator with revolute joints, a \ndesired joint angle is compared with a feedback signal from the joint sensor signalling \nactual measured angle;  the difference controls the motive power output of the joint \nactuator proportionally. \n\nWhere  a  manipulator  is  constructed  of  a  number  of  links,  there  might  be  a  ser(cid:173)\nvomechanism for  each joint.  In  combination,  it  is  well  known  that  joint  motions \ncan  affect  each other adversely, requiring careful design  and analysis to reduce  the \npossibility of unpleasant dynamical instabilities.  This is  especially important when \nthe  manipulator will  be required  to execute fast  movements involving many or  all \nof the joints.  We  are  interested in  such dynamic tasks,  and  acknowledge some suc(cid:173)\ncessful servomechanistic solutions  (see  [Andersson  19881, who describes a ping pong \nplaying robot),  but seek an  alternative that is  not as  computationally expensive. \n\n1.2  ESCAPING POSITION CONTROL \n\nIn  Nature,  fast  reaching and  striking is  a  primitive and fundamental mode of con(cid:173)\ntrol.  In  fast,  time-optimal,  neurologically  ballistic  movements  (such  as  horizontal \nrotations  of the  head  where  subjects  are  instructed  to  turn  it  as  fast  as  possible, \n[Hannaford  and Stark  1985]),  muscle activity patterns seem  to  show  three phases: \na launching phase  (a burst of agonist), a  braking phase (an antagonist burst), and a \nlocking phase  (a second agonist  burst).  Experiments have shown  (see  [Wadman  et \nal.  1979])  that at least the first  100 mS of activity is  the same even if a movement is \nblocked mechanically  (without forewarning the subject), suggesting that the launch \nis specified from  predetermined initial conditions  (and is not  immediately modified \nfrom  proprioceptive information).  With  the  braking  and  locking  phases  acting  as \na  damping  device  at  the  end  of  the  motion,  the  complete  motion  of  the  arm  is \nessentially specified by the initial conditions -\nditional robot positional control.  The overall coordination of movements might even \nseem naive and simple when compared with the intricacies of servomechanisms (see \n[Braitenberg 1989, N ahvi and Hashemi 19841 who discuss the crane driver's strategy \nfor  shifting loads quickly  and time-optimally). \n\na mode radically differing from tra(cid:173)\n\nThe concept of letting insights  (such  as  these)  that can  be gained from  the biolog(cid:173)\nical sciences  shape  the  engineering  principles  used  to  create  artificial  autonomous \nsystems is finding favour with a  growing number of researchers in  robotics.  As  it is \nnot generally trivial to see how life's devices can be mapped onto machines, there is \na  need for  some fundamental experimental work to develop  and test the basic  the(cid:173)\noretical and empirical components of this  approach, and we  have been  considering \nvarious robotics problems from  this perspective. \nHere, we discuss  an experimental two-link manipulator that performs a simple ma(cid:173)\nnipulation  task -\nhitting  a  simple  object  perceived  to  be  within  its  reach.  The \nperception  of the object  specifies  the  initial conditions that determine an  arm  mo-\n\n\f310 \n\nDonnett and Smithers \n\ntion that reaches it.  In relating initial conditions with motor currents, we  have been \nconsidering  a  scheme  based  on  Neuronal  Group Selection  Theory  [Edelman  1987, \nReeke  and  Edelman  1988],  a  theory  of brain  organization.  We  believe  this  to  be \nthe  first  attempt  to  apply  selectionist  ideas  in  a  real  machine,  rather than just  in \nsimulation. \n\n2  NEURONAL GROUP SELECTION THEORY \nEdelman proposes Neuronal Group Selection  (NGS)  [Edelman  1978]  as  an organiz(cid:173)\ning principle for higher brain function - mainly a biological basis for perception -\nprimarily  applicable  to  the  mammalian  (and  specifically,  human)  nervous  system \n[Edelman  1981].  The essential idea is  that groups  of cells,  structurally varied  as  a \nresult  of developmental processes,  comprise  a  population  from  which  are  selected \nthose  groups  whose  function  leads  to  adaptive  behaviour  of  the  system.  Similar \nnotions  appear  in  immunology  and,  of  course,  evolutionary  theory,  although  the \neffects of neuronal group selection  are manifest  in  the lifetime of the organism. \n\nThere  are  two  premises  on  which  the  principle  rests.  The first  is  that  the  unit  of \nselection  is  a  cell  group of perhaps 50  to  10,000 neurons.  Intra-group connections \nbetween cells  are  assumed to vary (greatly)  between groups,  but other connections \nin the brain (particularly inter-group) are quite specific.  The second premise is  that \nthe kinds of nervous systems whose organization the principle addresses are able to \nadapt  to circumstances  not  previously  encountered by  the  organism  or  its  species \n[Edelman  1978]. \n\n2.1  THREE CENTRAL TENETS \n\nThere are  three important ideas  in  the NGS  theory  [Edelman  1987]. \n\n\u2022  A  first  selective  process  (cell  division,  migration,  differentiation,  or  death) \nresults  in  structural  diversity  providing  a  primary  repertoire  of  variant  cell \ngroups. \n\n\u2022  A second selective process occurs as the organism experiences its environment; \ngroup  activity  that  correlates  with  adaptive  behaviour  leads  to  differential \namplification  of intra- and  inter-group  synaptic  strengths  (the  connectivity \npattern remains  unchanged).  From the primary  repertoire  are  thus selected \ngroups whose  adaptive functioning  means  they are more  likely to find  future \nuse  -\n\nthese groups form  the  ,econdary repertoire. \n\n\u2022  Secondary repertoires themselves form  populations,  and the NGS  theory ad(cid:173)\n\nditionally  requires  a  notion  of  reentry,  or  connections  between  repertoires, \nusually  arranged  in  maps,  of which  the  well-known  retinotopic  mapping  of \nthe visual system  is  typical.  These  connections  are critical for  they correlate \nmotor and sensory repertoires,  and lend the world the kind of spatiotemporal \ncontinuity we  all experience. \n\n\fNeuronal Group Selection Theory:  A Grounding in Robotics \n\n311 \n\n2.2  REQUffiEMENTS  OF  SELECTIVE SYSTEMS \n\nTo be selective, a system must satisfy three requirements IReeke and Edelman 1988]. \nGiven a configuration of input signals  (ultimately from the sensory epithelia, but for \n'deeper' repertoires mainly coming from other neuronal groups), if a group responds \nin a specific  way it has  matched the input  IEdelman 1978].  The first  requirement of \na selective system is that it have a sufficiently large repertoire of variant elements to \nensure  that an  adequate match can  be found for  a wide range of inputs.  Secondly, \nenough of the groups in  a  repertoire must 'see' the diverse input signals effectively \nand quickly so that selection can operate on  these  groups.  And finally,  there must \nbe  a  means  for  'amplifying'  the  contribution,  to  the  repertoire,  of groups  whose \noperation when  matching input signals has led  to adaptive behaviour. \n\nIn  determining  the  necessary  number  of groups in  a  repertoire,  one  must consider \nthe  relationship  between repertoire  size  and  the specificity  of member groups.  On \nthe  one  hand,  if groups  are  very specific,  repertoires  will  need  to  be  very large  in \norder to recognize a  wide range of possible inputs.  On the other hand, if groups are \nnot  as  discriminating,  it  will  be possible  to have smaller numbers  of them,  but  in \nthe limit (a single group with virtually no specificity) different signals will no longer \nbe  distinguishable.  A  simple  way  to  quantify  this  might  be  to  assume  that  each \nof N  groups has  a  fixed  probability,  P,  of matching an input configuration;  then  a \ntypical measure  IEdelman  1978]  relating  the  effectiveness  of recognition,  r,  to  the \nnumber of groups is  r  = 1 -\n\n(1 - p)N  (see  Fig.  1). \n\nr \n\nlog N \n\nFigure  1:  Recognition  as  a  Function of Repertoire Size \n\nFrom  the shape  of the  curve  in  Fig.  1,  it  is  clear  that,  for  such a  measure,  below \nsome  lower threshold  for  N,  the  efficacy  of recognition  is  equally  poor.  Similarly, \nabove an upper threshold for N, recognition does not improve substantially as more \ngroups  are  added. \n\n3  SELECTIONISM IN  OUR EXPERIMENT \nOur manipulator is required to touch an object perceived to be within reach.  This is \na  well-defined  but  non-trivial problem  in  motor-sensory coordination.  Churchland \nproposes a  geometrical solution for  his  two-eyed 'crab' IChurchland  1986]'  in which \n\n\f312 \n\nDonnett and Smithers \n\neye  angles  are  mapped  to  those  joint  angles  (the  crab  has  a  two-link  arm)  that \nwould  bring the  end  of the  arm  to  the  point currently foveated  by  the  eyes.  Such \na  novel solution, in which computation is implicit and massively parallel, would be \nwelcome;  however,  the crab is  a simulation, and no heed  is  paid to the question  of \nhow the  appropriate sensory-motor mapping could be  generated for  a real arm. \nReeke  and  Edelman  discuss  an  automaton,  Darwin  III,  similar  to  the  crab,  but \nwhich by selectional processes develops the ability to manipulate objects presented \nto  it  in  its  environment  [Reeke  and  Edelman  19881.  The  Darwin  III  simulation \ndoes  not  account  for  arm dynamics;  however,  Edelman  suggests  that  the  training \nparadigm is  able  to handle dynamic effects  as  well as  the geometry of the problem \n[Edelman  19891.  We are attempting to implement a mechanical analogue of Darwin \nIII,  somewhat simplified,  but which will experience the real dynamics of motion. \n\nS.l  EXPERIMENTAL ARCHITECTURE AND HARDWARE \n\nThe mechanical arrangement of our manipulator is shown in  Fig.  2.  The two links \nhave  agonist/antagonist  tendon-drive  arrangement,  with  an  actuator  per  tendon. \nThere are strain gauges in-line with the tendons.  A manipulator 'reach' is specified \nby six parameters:  burst amplitude and period for each of the three phases, launch, \nbrake, and lock. \n\n'I. \n\n'I. \n\n'I. \n\n'I. \n\n',tendons \n\n'I. , , , \n\nl \n\n'I. \n\n'0 \n\n\" \n\" \\ \nupper-arm  D ri \nU \nforearm/  ~ forearm \n\nleft actuator \n\nupper-arm \nright actuator \n\nleft actuator \n\nright actuator \n\nFigure 2:  Manipulator Mechanical Configuration \n\n\fNeuronal Group Selection Theory:  A Grounding in Robotics \n\n313 \n\nAt  the  end  of the  manipulator is  an  array  of eleven  pyroelectic-effect  infrared  de(cid:173)\ntectors  arranged  in  a  U-shaped  pattern.  The  relative  location  of  a  warm  object \npresented to the  arm is  registered  by  the sensors,  and is  converted to eleven 8-bit \nintegers.  Since  the  sensor output  is  proportional  to detected infrared  energy flux, \nobjects at the same temperature will give  a  more  positive reading if they are  close \nto  the  sensors  than  if they  are  further  away.  Also,  a  near  object  will  register  on \nadjacent  sensors,  not  just on  the  one  oriented  towards it.  Therefore,  for  a  single, \nsmall  object,  a  histogram  of the eleven  values  will  have  a  peak,  and  showing  two \nthings  (Fig.  3):  the  sensor  'seeing'  the  most  flux  indicates  the  relative  direction \nof the object,  and  the  sharpness  of the peak is  proportional to the distance of the \nobject. \n\n(object distant \n\nand to the left) \n\n(object near and \n\nstraight ahead) \n\nFigure 3:  Histograms for  Distant Versus Near Objects \n\nModelled  on  Darwin  III  [Reeke  and  Edelman  1988],  the  architecture  of the  selec(cid:173)\ntional perception-action coordinator is as in Fig.  4.  The boxes represent repertoires \nof appropriately interconnected groups of 'neurons'. \n\nDarwin III responds mainly to contour in a two-dimensional world, analogous to the \nrecognition of histogram shape in our system.  Where Darwin Ill's 'unique response' \nnetwork is  sensitive to line segment lengths and orientations, ours is sensitive to the \nlength of subsequences in the array of sensor output values in which values increase \nor decrease by the same amount, and the amounts by which they change; similarly, \nwhere Darwin Ill's 'generic response' network is sensitive to presence of or  changes \nin orientation of lines, ours responds to the presence of the subsequences mentioned \nabove, and the positions in  the  array where  two subsequences abut. \n\nThe recognition repertoires are reciprocally connected, and both connect to the mo(cid:173)\ntor repertoire which consists of ballistic-movement 6-tuples.  The  system considers \n'touching perceived object'  to be  adaptive,  so  when  recognition  activity correlates \nwith a given 6-tuple,  amplification ensures that the same response will be favoured \nin future. \n\n\f314 \n\nDonnett and Smithers \n\n4  WORK  TO DATE \nAs  the sensing system  is  not  yet functional,  this  aspect  of the system  is  currently \nsimulated in  an  IBM  PC/AT. The rest  of the electrical and  mechanical hardware \nis in place.  The  major difficulty currently faced  is  that the selectional system will \nbecome computationally intensive on a serial machine. \n\nWORLD \n\nFEATURE \nDETECTOR \n\nFEATURE \nCORRELATOR \n\nclassification \n\ncouple \n\nCOMBINATION \nRESPONSES \n(UNIQU~ \n\nCOMBINATION \nRESPONSES \n\ncim\"':r\"~f,~./(GENERIC) \n\n~ motor map \nMOTOR \nACTIONS \n\nFigure 4:  Experimental Architecture \n\nFor  each  possible  ballistic  'reach',  there  must  be  a  representation  for  the  'reach \n6-tuple'.  Therefore, the motor repertoire must become large as  the dexterity of the \nmanipulator is  increased.  Similarly,  as  the  array of sensors is extended  (resolution \nincreased,  or field  of view  widened),  the  classification  repertoires  must  also  grow. \nOn  a  serial  machine,  polling  the  groups  in  the  repertoires  must  be  done  one  at \na  time,  introducing  a  substantial delay between  the registration  of object  and  the \nactual touch, precluding the interception by the manipulator of fast  moving objects. \nWe are exploring possibilities for  parallelizing the selectional process  (and have for \nthis reason constructed a network of processing elements), with the expectation that \nthis  will  lead  us  closer  to  fast,  dynamic  manipulation,  at  minimal computational \nexpense. \n\n\fNeuronal Group Selection Theory:  A Grounding in Robotics \n\n315 \n\nReferences \n\nRussell  L.  Andersson.  A  Robot Ping-Pong  Player:  Experiment in Real- Time  Intel(cid:173)\nligent  Control.  MIT Press,  Cambridge, MA,  1988. \n\nValentino Braitenberg.  \"Some types of movement\" , in C.G. Langton, ed.,  Artificial \nLife, pp.  555-565,  Addison-Wesley,  1989. \n\nPaul M.  Churchland.  \"Some reductive strategies in cognitive neurobiology\".  Mind, \n95:279-309,  1986. \n\nJim  Donnett  and  Tim  Smithers.  \"Behaviour-based  control of a  two-link  ballistic \narm\".  Dept.  of Artificial Intelligence, University of Edinburgh, Research Paper  RP \n.158,  1990. \nGerald M.  Edelman.  \"Group selection and phasic reentrant signalling:  a  theory of \nhigher brain function\",  in  G.M.  Edelman and V.B.  Mountcastle, eds.,  The  Mindful \nBrain,  pp.  51-100,  MIT Press,  Cambridge, MA,  1978. \n\nGerald  M.  Edelman.  \"Group  selection  as  the  basis  for  higher  brain function\",  in \nF.O.  Schmitt et  al.,  eds.,  Organization  of the  Cerebral  Cortex,  pp.  535-563,  MIT \nPress,  Cambridge, MA,  1981. \n\nGerald M.  Edelman.  Neural  Darwinism:  The  Theory  of Neuronal  Group  Selection. \nBasic  Books,  New  York,  1987. \nGerald M.  Edelman.  Personal correspondence,  1989. \n\nBlake Hannaford and Lawrence Stark.  \"Roles of the elements of the triphasic control \nsignal\".  Experimental Neurology,  90:619-634,  1985. \nM.J.  Nahvi and M.R. Hashemi.  \"A synthetic motor control system;  possible paral(cid:173)\nlels with transformations in cerebellar cortex\", in J .R. Bloedel et al., eds.,  Cerebellar \nFunctions,  pp.  67-69, Springer-Verlag,  1984. \n\nGeorge  N.  Reeke  Jr.  and  Gerald  M.  Edelman.  \"Real  brains  and  artificial  in(cid:173)\ntelligence\",  in  Stephen  R.  Graubard,  ed.,  The  Artificial  Intelligence  Debate,  pp. \n143-173,  The MIT Press,  Cambridge, MA,  1988. \nW.J.  Wadman,  J.J.  Denier van  der Gon,  R.H.  Geuse,  and  C.R.  Mol.  \"Control of \nfast  goal-directed arm movements\".  Journal of Human  Movement Studies,  5:3-17, \n1979. \n\n\f", "award": [], "sourceid": 204, "authors": [{"given_name": "Jim", "family_name": "Donnett", "institution": null}, {"given_name": "Tim", "family_name": "Smithers", "institution": null}]}