 |
    |
Design of Multithreaded Estimation of Distribution Algorithms
Jiri Ocenasek1, Josef Schwarz2, and Martin Pelikan1
1Computational Laboratory (CoLab)
Swiss Federal Institute of Technology ETH
Hirschengraben 84
8092 Zürich, Switzerland
{jirio,pelikanm}@inf.ethz.ch
2Faculty of Information Technology
Brno University of Technology
Bozetechova 2
612 66 Brno, Czech Republic
schwarz@fit.vutbr.cz
Abstract. Estimation of Distribution Algorithms (EDAs) use a probabilistic model of promising solutions found so far to obtain new candidate solutions of an optimization problem. This paper focuses on the design of parallel EDAs. More specifically, the paper describes a method for parallel construction of Bayesian networks with local structures in form of decision trees in the Mixed Bayesian Optimization Algorithm. The proposed Multithreaded Mixed Bayesian Optimization Algorithm (MMBOA) is intended for implementation on a cluster of workstations that communicate by Message Passing Interface (MPI). Communication latencies between workstations are eliminated by multithreaded processing, so in each workstation the high-priority model-building thread, which is communication demanding, can be overlapped by low-priority model sampling thread when necessary. High performance of MMBOA is verified via simulation in TRANSIM tool.
LNCS 2724, p. 1247 ff.
lncs@springer.de
© Springer-Verlag Berlin Heidelberg 2003