NASAUs Jet Propulsion Laboratory, Pasadena, California
Boolean neural networks have been proposed to impiement part of an intermediate level (the rule level) of a hierarchical architecture of a control s ystem for the artificially intelligent control of a robot hand A very-large-scale integ rated- circuit prototype of such a network has been built and is undergoing laboratory tests, all as part of a continuing effort to delineate further the hierarchical-c ontrol concept and design effective hierarchical control circuits. The underlying hierarchical-control concept is a multi-level hierarchy, each leve l of which can be described as performing a different kind of control computation that corresponds to a different kind of mental process analogous to the mental processes of a person engaging in a grasping task (see Figure 2). The lowest level would inteface directly with sensors and actuators, the middle level (the r ule level that is the iocus oi attention in this article) would command responses to patterns that it would recognize in feedback signals, and the highest level would be dedicated to planning and the recognition of sequences of patterns. In ascending the hierarchy, one would encounter increasingly symbolic types of logic. The hierarchical architecture (see Figure 2) requires several additions to the standard architectures of previous planner and actuator controllers. First, one h as to generate the set of conditions that constitute the qualitative component of th e specification of the evolution of a task. Second, one has to store a set of alter nate trajectories. Within the rule level of the hierarchy, range selectors would transform the analog outputs of the sensors into Boolean variables (logic-level signals). Two Boolean neural networks would process the Boolean variables according to memorized Rif . thenS rules supplied by the planner (which would reside in the highest level of the hierarchy). This concept relies on two baslc assumptions: that a small set ol actuator trajectories can cover most grasping tasks and that the evolution of a grasping task can be represented by logical conditions. The Boolean networks would be operated in alternation under control of an adaplatlon-and-rule controller. which would also supervise the loading of a new plan and its rules. While one Boolean network was operating, the other one would be receiving the updated plan. Once updating was completed, the newly updated network would be switched into operatlon and the other swltched out of operation to be updated in turn. In this way, updating could proceed wlthout intefering with the correct processing of feedback signals. The prototype Boolean neural network is an adaptive, self-organizing logic network in that it takes an active part in reconfiguring its own loglc gates. In so doing, it strives to optimize its configuration and/or performance with respect t o such criteria as minimality and consistency of the rule base. During processing, the network acts as a programmable logic array. During adaptation, as new rules are added, the network automatically reconfigures itself into a logic clrcuit tha t seeks to maintain a mlnimum and consistent rule base. There is no explicit programmlng of the network, and the internal conflguratlon of the network is not unique but rather depends on the initial state and on the history of the prevlous adaptations. The network accepts new rules that are sequentially presented to it by an external controller. This process allows each node in the network to determlne its relation with the new rule and determine whether it should be involved in the adaptation process. The adaptation may involve addltlon or deletlon of nodes or compaction of subnetworks. A central controller is used for coordinatlon, but the adaptlve process itself is completely distributed In the network, and modifications to the network are performed wlth considerable concurrency.
Point of Contact:
Paolo Fiorini
Mail Stop 198-219
Jet Propulsion Laboratory
4800 Oak Grove Drive
Pasadena, CA 91109
818-354-9061
Paolo.Fiorini@jpl.nasa.gov
Telerobotics Program
page
Please email the site webmaster
with any comments, criticisms or corrections
for this page.
Maintained by: Dave Lavery
Last updated: May 10, 1996