Neuromimetic Algorithms Processing: Tools for Design of Dedicated Architectures

The finality of the research subject is to derive a new parallel architecture of computer dedicated to connectionist algorithms and especially neuromimetic algorithms. The study of this problem induces us to carry out a simulation model. This model allows to highlight difficulties of parallel implementation on multiprocessor architecture of connectionist algorithms, i.e. projection of application graphs on the target machine graph in order to achieve optimal performances. Implementation is difficult because data transfers between processors induce mapping and scheduling problems. Then, some mapping strategies and optimization heuristics are used, as Node Swapping, Simulated Annealing or Genetic Algorithms. According to the results the model simulation allows to vary several parameters of the studied machine: number of processors, architecture, particular parallel mode (SIMD, SPMD,...), interconnection network (static or dynamic) and communication mode. It also allows to research a trade off between processing loads and communication loads. Such a way, some basic specifications of dedicated architecture will be chosen from the comparison of a set of alternatives.