Lessons Learned from Prototyping Parallel Computer Architectures for AI Algorithms

Since many years algorithms from the eld of arti cial intelligence (AI) have been targeted for parallelization, i.e., partitioning the search problem and distributing the subproblems among multiple processing nodes. This paper reports on our experience in parallelizing and distributing AI algorithms, i.e., the design and prototype implementation of parallel computer architectures for AI algorithms. While the presented prototypes have been designed for speci c applications, the underlying algorithms are quite general and frequently-used AI search methods, i.e., constraint satisfaction and simulated annealing. Our prototypes achieve speedup factors of several orders of magnitude. keywords: parallel and distributed AI; specialized computer architecture; constraint satisfaction; simulated annealing