: The Threat Detection Group (TOG) at Defense R&D Canada-Suffield has undertaken a research program on the feasibility of remote sensing of minefields. One of the projects is the Remote Minefield Detection (RMD) hierarchical algorithm originally developed for single band active airborne infrared imagery (AAII) and later adapted for vehicle mounted passive infrared imagery (FLIR) The objectives of this contract are: (i) to Implement the RMD hierarchical algorithm on a down-scaled version of transputer architecture for real-time mine detection on FLIR Imagery, (ii) to develop the high and top levels of the algorithm Including expert system and knowledge base, and (iii) to research the possibility to upgrade the current hardware platform to modern advanced computational elements Phase 1 and 2 of the contract focused on the implementation of Low and Middle levels of a real-time RMD system on a transputer network to detect mines in real FLIR images. The work first described the hardware requirement for each module of the algorithm, then outlined the software development, and finally presented some preliminary test results. Although transputers were attractive computing elements 15 years ago when this project started, they have become obsolete quickly today Thus a decision has been made by the Project Authority to implement the RMD hierarchical algorithm on the new hardware platform, namely a network of Intel Pentium-class processors, which was recommended by a study from the University of British Columbia. Phase 3 of the contract Involved building that PC network, and developing and testing the new software version.

An extension to the convergence theorem for single neurons learning under the AR-P algorithm is proved. The extension shows that if the conditions of the single-neuron theorem are satisfied and if the environment satisfies one of two sufficient conditions, the weights in an arbitrarily large one-layer network will converge with probability one to values with which the network correctly classifies the training input set. One condition requires that for all output vectors, the probability of reinforcement being one (success) has one of two values: the output vectors having at least lk correct elements have the higher probability, whereas the output vectors having less than lk correct elements have the lower probability. The alternative condition requires that the reinforcement have a higher probability of being one for output vectors having a higher number of correct elements. The extension and its proof are significant because they further the understanding of the factors affecting the convergence of multilayer networks under reinforcement learning