(371c) Implementation of Distributed Optimisation Algorithm On Computer Grid | AIChE

(371c) Implementation of Distributed Optimisation Algorithm On Computer Grid

Authors 

Du, D. - Presenter, University of Surrey
Cecelja, F. - Presenter, University of Surrey
Kokossis, A. - Presenter, University of Surrey


High-throughput synthesis of complex process applications based on stochastic optimization algorithms such as Tabu Search and Simulated Annealing, which normally involves a large amount of components and parameters and is usually processed sequentially, often experiences slow convergence or trapping in local optimum. This paper presents a novel distributed optimization algorithm, the cascade algorithm, which has been implemented for solving large scale process synthesis problems by splitting long sequential Markov chain into smaller and parallel sections running on computer grid. Along with the cascading, the self-supervised intelligent systems based on knowledge model in the form of production rules is employed to benefit from knowledge generated at each stage of optimization and reuse it in consequent stages. Results have been obtained under different computing environments and show that the execution time to convergence reduces with the increase of number CPUs with faster CPUs contributing more than slower ones. As for the self-supervised system, search directions that determine which move can be made in a stochastic optimization process, are biased by putting different weights to the patterns of intrinsic parameters for each solution according to on-time analytical results of these parameters. Results show that the optimization search converges more quickly by applying more parameters in the production rules in the model. A conclusion can be made that appropriate use of grids and production rules in a self-supervised knowledge-based optimization system is possible to improve the optimization performance, such as convergence speed, of high-throughput systems, i.e. reactor network design synthesis. Future work will focus on the integration of knowledge-based optimization system with ontology, in which more underlying relations among the intrinsic parameters may be explored and put into production rules to further guide the optimization search.