Scalable architecture for parallel distributed implementation of genetic programming on network of workstations

We present an approach for developing a scalable architecture for parallel distributed implementation of genetic programming (PDIGP). The approach is based on exploitation of the inherent parallelism among semi-isolated subpo pulations in genetic programming (GP). Proposed implementation runs on cost -efficient configurations of networks on workstations in LAN and Internet e nvironment. Developed architecture features single global migration broker and centralized manager of the semi-isolated subpopulations, which contribu te to achieving quick propagation of the globally fittest individuals among the subpopulations, reducing the performance demands to the communication network, and achieving flexibility in system configurations by introducing dynamically scaling up opportunities. PDIGP exploits distributed component object model (DCOM) as a communication paradigm, which as a true system mod el offers generic support for the issues of naming, locating and protecting the distributed entities in proposed architecture of PDIGP. Experimentally obtained results of computational effort of proposed PDIGP are discussed. The results show that computational effort of PDIGP marginally differs from the computational effort in canonical panmictic GP evolving single large p opulation. For PDIGP running on systems configurations with 16 workstations the computational effort is less than panmictic GP, while for smaller conf igurations it is insignificantly more. Analytically obtained and empiricall y proved results of the speedup of computational performance indicate that PDIGP features linear, close to ideal characteristics. Experimentally obtai ned results of PDIGP running on configurations with eight workstations show close to 8-fold overall speedup. These results are consistent with the ant icipated cumulative effect of the insignificant increase of computational e ffort for the considered configuration and the close to linear speedup of c omputational performance. (C) 2001 Elsevier Science B.V. All rights reserve d.