A modified ARTMAP network, with applications to scheduling of a robot-vision-tracking system

The use of a robot-vision-tracking system to efficiently process different types of objects presented randomly on a moving conveyor belt requires the system to schedule pick and place operations of the the robot to minimize robot processing times and avoid constraint violations. In this paper we present a new approach: a modified ARTMAP neural network is incorporated in the robot-vision-tracking system as an ''intelligent'' component to on-line schedule pick-place operations in order to obtain optimal orders for any group of objects. When the robot-vision-tracking system is working in a changing environment the neural networks used in the optimal scheduling task must be capable of updating their weights aperiodically based on the data collected intermittently in real operations in order to create a continuously effective system. The ARTMAP network developed by Carpenter et al, (1991), which can rapidly learn mappings between binary input and binary output vectors by using a supervised learning law, has good properties to deal with this task. In special situations, however, the ARTMAP must employ a complement coding technique to preprocess incoming patterns to be presented to the network. This doubles the size of input patterns and increases learning rime. The Modified ARTMAP network, proposed herein, copes with these special situations without using complement coding, and has been shown to increase the overall system speed. The basic idea is to insert a matching check mechanism that internally changes the learning order of input vector pairs in respending to an arbitrary sequence of arriving input vector pairs. Simulation results are presented for scheduling a number of different objects, demonstrating a substantial improvement in learning speed and accuracy.