Evaluating advanced architectures for AI production systems

The author evaluates the suitability of advanced parallel processing architectures for efficiently executing forward-chaining AI (artificial intelligence) production system (PS) algorithms. A PS overview is presented and sequential PS algorithms discussed. Fundamental PS algorithm processes are described and reported metrics reviewed. Architectures constructed or proposed for executing parallel PS are surveyed, and parallel algorithms, such as the parallelized Rete, TREAT, DADO, Shaw, and Loral bit-vector algorithms, are reviewed. Algorithm state-saving and execution modality characteristics are summarized. The suitability of the following types of parallel processing architectures for PS is then evaluated on the basis of metrics, benchmarks, simulations and formal studies: shared memory, bus-based multiprocessors: associative memory processors; data-flow machines hypercube-structured, message-passing architectures; and tree-structured multiprocessor architectures