Literature/1979/Albus


 * http://www.isd.mel.nist.gov/documents/albus/Loc_5.pdf

Key points

 * Cerebellar Model Articulation Controller, CMAC

Abstract

 * [...] A neurophysiological model is described which demonstrates the capacity to learn, to generalize, to compute multivariate mathematical functions, and to decompose input commands into sequences of output commands in a context-sensitive manner. Evidence is presented that clusters of neurons with such properties are arranged in hierarchical structures in the brain so as to produce AND/OR task compositions. At the lowest levels in the motor system these clusters transform coordinates and compute servo functions. At the middle levels they decompose input commands into sequences of output commands which give rise to behavior patterns. Mechanisms by which feedback can alter these decomposition sequences to compensate for perturbations and uncertainties in the environment are described. [...]

Extracts

 * The reader should understand ... that any model of neuronal mechanisms of the higher cognitive processes must of necessity involve speculation and metaphorical language. It should also be understood that a multidisciplinary approach to such a large and complex subject cannot avoid a mixing of jargon and an oversimplification of many difficult issues.
 * One of the fundamental techniques used in problem solving is the decomposition of problems into subproblems which are simpler to solve than the original problem. This procedure can be repeatedly applied to subproblems, and then to sub-subproblems, until finally the end product is a set of primitive problems for which there are known one-step solutions.
 * The concept of solving difficult problems, and the closely related concept of controlling large systems, by problem or task decomposition are old ones. It is implicit in many forms of control hierarchies such as exist in military command structures, business management organizations, and industrial manufacturing procedures, and has for more than a century been assumed to be a mechanism used by the brain for generating and controlling behavior.
 * Most artificial -intelligence research in goal or task decomposition has been done in a planning context as opposed to a control context. In planning the emphasis naturally tends toward mechanisms for searching AND/OR graphs to find, and hopefully optimize, solutions. This implies a detached evaluation of alternatives where the constraints of real-time interaction with a dynamic and unpredictable environment can be deemphasized or ignored altogether. [...] The effect is to place enormous demands on the planning programs for attending to details, and the resulting behavior appears most unintelligent because of lengthy interplan periods of open-loop activity in which there is little or no interaction with sensory feedback.
 * An obvious but seldom recognized fact is that planning is not characteristic of the behavior of most biological organisms. In most creatures the central nervous system is primarily a control mechanism for goal-seeking, not planning. Only in the most advanced species does the brain demonstrate any significant capacity for foresight, imagination, and systematic evaluation of potential future scenarios in the formulation of plans.
 * The distinguishing feature of goal-seeking, as opposed to planning, is that it is a real-time control process resulting in physical activity. Goal-seeking produces a sequence of overt actions which may be represented by a single uninterrupted string of primitive actions[...].
 * Planning ... is not a real-time control process, nor is it best characterized by a single uninterrupted trajectory. Planning involves the postulation of hypothetical situations, the evaluation of predicted or imagined results, and the optimization of solution paths prior to the initiation of overt behavioral activity.
 * The overwhelming weight of evidence from the evolutionary record ... as well as from contemporary behavioral science ... indicates that the original and still primary function of the brain is not to think and plan, but to act and react. When one observes the behavior of creatures in the lower to middle ranges of intelligence such as ants, bees, fish, birds, and mice, there is ample evidence for goal-seeking ... but virtually none for planning.
 * The ability to learn, or modify behavioral algorithms through experience, also has nothing to do with the ability to plan. Virtually all species can learn, or modify their responses to sensory stimuli, but only the most advanced use learned associates to any significant degree for generating internal representations of future or hypothetical situations which can then be evaluated in the formulation of plans.
 * The rarity and late arrival of the ability to plan suggests that a highly developed precursor, or substrate, was required from which planning capabilities could evolve. Both the similarities and the differences between planning and goal-seeking suggest that the mechanism for sensory -interactive goal-directed behavior may have been this precursor.

Chronology

 * James S. Albus (1975). "A new approach to manipulator control: the cerebellar model articulation controller (CMAC)," Journal of Dynamic Systems, Measurement and Control, Sept. 1975, pp. 220-227. (Online PDF)
 * James S. Albus (1975). "Data storage in the cerebellar model articulation controller (CMAC),"  Journal of Dynamic Systems, Measurement and Control, Sept. 1975, pp. 228-233.
 * James S. Albus (1975). "Data storage in the cerebellar model articulation controller (CMAC),"  Journal of Dynamic Systems, Measurement and Control, Sept. 1975, pp. 228-233.