Neuronal Group Selection Theory: A Grounding in Robotics

Part of Advances in Neural Information Processing Systems 2 (NIPS 1989)

Bibtex Metadata Paper

Authors

Jim Donnett, Tim Smithers

Abstract

In this paper, we discuss a current attempt at applying the organi(cid:173) zational principle Edelman calls Neuronal Group Selection to the control of a real, two-link robotic manipulator. We begin by moti(cid:173) vating the need for an alternative to the position-control paradigm of classical robotics, and suggest that a possible avenue is to look at the primitive animal limb 'neurologically ballistic' control mode. We have been considering a selectionist approach to coordinating a simple perception-action task.

1 MOTIVATION The majority of industrial robots in the world are mechanical manipUlators - often arm-like devices consisting of some number of rigid links with actuators mounted where the links join that move adjacent links relative to each other, rotationally or translation ally. At the joints there are typically also sensors measuring the relative position of adjacent links, and it is in terms of position that manipulators are generally controlled (a desired motion is specified as a desired position of the end effector, from which can be derived the necessary positions of the links comprising the manipulator). Position control dominates largely for historical reasons, rooted in bang-bang control: manipulators bumped between mechanical stops placed so as to enforce a desired trajectory for the end effector.

Neuronal Group Selection Theory: A Grounding in Robotics

309

1.1 SERVOMECHANISMS

Mechanical stops have been superceded by position-controlling servomechanisms, negative feedback systems in which, for a typical manipulator with revolute joints, a desired joint angle is compared with a feedback signal from the joint sensor signalling actual measured angle; the difference controls the motive power output of the joint actuator proportionally.

Where a manipulator is constructed of a number of links, there might be a ser(cid:173) vomechanism for each joint. In combination, it is well known that joint motions can affect each other adversely, requiring careful design and analysis to reduce the possibility of unpleasant dynamical instabilities. This is especially important when the manipulator will be required to execute fast movements involving many or all of the joints. We are interested in such dynamic tasks, and acknowledge some suc(cid:173) cessful servomechanistic solutions (see [Andersson 19881, who describes a ping pong playing robot), but seek an alternative that is not as computationally expensive.

1.2 ESCAPING POSITION CONTROL

In Nature, fast reaching and striking is a primitive and fundamental mode of con(cid:173) trol. In fast, time-optimal, neurologically ballistic movements (such as horizontal rotations of the head where subjects are instructed to turn it as fast as possible, [Hannaford and Stark 1985]), muscle activity patterns seem to show three phases: a launching phase (a burst of agonist), a braking phase (an antagonist burst), and a locking phase (a second agonist burst). Experiments have shown (see [Wadman et al. 1979]) that at least the first 100 mS of activity is the same even if a movement is blocked mechanically (without forewarning the subject), suggesting that the launch is specified from predetermined initial conditions (and is not immediately modified from proprioceptive information). With the braking and locking phases acting as a damping device at the end of the motion, the complete motion of the arm is essentially specified by the initial conditions - ditional robot positional control. The overall coordination of movements might even seem naive and simple when compared with the intricacies of servomechanisms (see [Braitenberg 1989, N ahvi and Hashemi 19841 who discuss the crane driver's strategy for shifting loads quickly and time-optimally).

a mode radically differing from tra(cid:173)

The concept of letting insights (such as these) that can be gained from the biolog(cid:173) ical sciences shape the engineering principles used to create artificial autonomous systems is finding favour with a growing number of researchers in robotics. As it is not generally trivial to see how life's devices can be mapped onto machines, there is a need for some fundamental experimental work to develop and test the basic the(cid:173) oretical and empirical components of this approach, and we have been considering various robotics problems from this perspective. Here, we discuss an experimental two-link manipulator that performs a simple ma(cid:173) nipulation task - hitting a simple object perceived to be within its reach. The perception of the object specifies the initial conditions that determine an arm mo-

310

Donnett and Smithers

tion that reaches it. In relating initial conditions with motor currents, we have been considering a scheme based on Neuronal Group Selection Theory [Edelman 1987, Reeke and Edelman 1988], a theory of brain organization. We believe this to be the first attempt to apply selectionist ideas in a real machine, rather than just in simulation.

2 NEURONAL GROUP SELECTION THEORY Edelman proposes Neuronal Group Selection (NGS) [Edelman 1978] as an organiz(cid:173) ing principle for higher brain function - mainly a biological basis for perception - primarily applicable to the mammalian (and specifically, human) nervous system [Edelman 1981]. The essential idea is that groups of cells, structurally varied as a result of developmental processes, comprise a population from which are selected those groups whose function leads to adaptive behaviour of the system. Similar notions appear in immunology and, of course, evolutionary theory, although the effects of neuronal group selection are manifest in the lifetime of the organism.

There are two premises on which the principle rests. The first is that the unit of selection is a cell group of perhaps 50 to 10,000 neurons. Intra-group connections between cells are assumed to vary (greatly) between groups, but other connections in the brain (particularly inter-group) are quite specific. The second premise is that the kinds of nervous systems whose organization the principle addresses are able to adapt to circumstances not previously encountered by the organism or its species [Edelman 1978].

2.1 THREE CENTRAL TENETS

There are three important ideas in the NGS theory [Edelman 1987].

• A first selective process (cell division, migration, differentiation, or death) results in structural diversity providing a primary repertoire of variant cell groups.

• A second selective process occurs as the organism experiences its environment; group activity that correlates with adaptive behaviour leads to differential amplification of intra- and inter-group synaptic strengths (the connectivity pattern remains unchanged). From the primary repertoire are thus selected groups whose adaptive functioning means they are more likely to find future use -

these groups form the ,econdary repertoire.

• Secondary repertoires themselves form populations, and the NGS theory ad(cid:173)

ditionally requires a notion of reentry, or connections between repertoires, usually arranged in maps, of which the well-known retinotopic mapping of the visual system is typical. These connections are critical for they correlate motor and sensory repertoires, and lend the world the kind of spatiotemporal continuity we all experience.

Neuronal Group Selection Theory: A Grounding in Robotics

311

2.2 REQUffiEMENTS OF SELECTIVE SYSTEMS

To be selective, a system must satisfy three requirements IReeke and Edelman 1988]. Given a configuration of input signals (ultimately from the sensory epithelia, but for 'deeper' repertoires mainly coming from other neuronal groups), if a group responds in a specific way it has matched the input IEdelman 1978]. The first requirement of a selective system is that it have a sufficiently large repertoire of variant elements to ensure that an adequate match can be found for a wide range of inputs. Secondly, enough of the groups in a repertoire must 'see' the diverse input signals effectively and quickly so that selection can operate on these groups. And finally, there must be a means for 'amplifying' the contribution, to the repertoire, of groups whose operation when matching input signals has led to adaptive behaviour.

In determining the necessary number of groups in a repertoire, one must consider the relationship between repertoire size and the specificity of member groups. On the one hand, if groups are very specific, repertoires will need to be very large in order to recognize a wide range of possible inputs. On the other hand, if groups are not as discriminating, it will be possible to have smaller numbers of them, but in the limit (a single group with virtually no specificity) different signals will no longer be distinguishable. A simple way to quantify this might be to assume that each of N groups has a fixed probability, P, of matching an input configuration; then a typical measure IEdelman 1978] relating the effectiveness of recognition, r, to the number of groups is r = 1 -

(1 - p)N (see Fig. 1).