Parallel neural net path-planner on hypercube and transputer

The task of finding a good'' path is of prime importance to any automated vehicle navigation controller. This task has been formulated as problems in computational geometry, which is commonly reduced to a graph search problem. We report here the preliminary results of a new approach that finds a good,'' if not optimal, path dynamically and in real-time by casting the task as a classification problem. The path-planner implements a back-propagation model for classification. The multi-scale representational strategy used for mapping the problem domain onto the input space of the back-propagation model ensures the applicability of the trained network to instances of the problem, which differ in the distribution of obstacles and the size of the problem domain. Software simulations based on the parallel computer EXPRESS operating system run, without modification of code, in Transputer-based systems as well as hypercube concurrent processor, with satisfactory efficiency. 10 refs., 7 figs.