AN ASSOCIATIVE ARCHITECTURE FOR GENETIC ALGORITHM-BASED MACHINE LEARNING

In the realm of machine-based learning, system architects seek to design machines with ever-greater levels of human-like autonomy and intelligence. Recent advances in massively parallel computing, coupled with massively parallel artificial intelligence, may one day lead to vast computational resources that will offer new solutions to problems in autonomous agents, database mining, speech processing and translation, adaptive vision systems, visualization systems, and animation. In this article, the author presents his investigation of a Learning Classifier System that he developed, through modeling and simulation, based on hybrid AI algorithms computationally matched to a specialized associative architecture he designed. A key characteristic of Learning Classifier Systems is that they have intense computational requirements, often in the form of highly repetitive yet simple search operations, which is why an associative architecture was selected. The proposed architecture couples a 64-bit content-addressable memory with a very basic 1-bit processing element and includes a reconfigurable bus under the autonomous control of each PE. In the programming model for mapping parallel data onto the architecture, groups of physical PEs are organized as single logical PEs to provide a flexible means of meeting varied memory requirements. A simulator for the associative architecture and LCS algorithms was built with a MasPar MP-1 system that consisted of 8,192 4-bit processors. The LCS implementation was tested on a difficult letter prediction problem that challenged the adaptive abilities of the LCS. The article addresses the successful outcome and evaluates the performance of the algorithm simulations.