Application of tabu search to optimizations in chemical engineering.

摘要

This dissertation presents efforts to develop an optimization approach based on Tabu Search (TS) for chemical engineering optimization problems. Modifications were made to the basic TS algorithm to enhance its performance when applied to continuous optimization problems. Two intensification strategies were developed. By modifying the distribution of the neighbors and reducing the search space dynamically, TS is able to consistently find precise global solutions. Three different neighbor generations strategies were compared with extensive experiments on global optimal benchmarks. Our random subset strategy, which generates neighbors by varying a randomly selected set of variables, is shown to be better than the commonly used approaches, which vary either only one variable or all variables to construct a new neighbor. Constraint handling techniques were integrated with TS to extend its applicability to chemical engineering problems. We developed an approach, which guarantees feasible solutions by solving the set of equations consisting of equality constraints. The approach can be used to solve optimization problems involving both linear and nonlinear equality constraints. Termination criteria were incorporated in TS to make more effective use of computation time. A termination-on-convergence criterion stops the process by identifying unproductive search. This approach is a good complement to termination at the maximum number of iterations. Several parallel TS schemes were implemented with OpenMP. Impressive speedup is achieved with functional decomposition strategy, which is suitable for problems involving a computationally intensive process to find the best neighbor. The communication among threads increases the probability of locating the global solution, which enables multi-search threads strategy to be useful for problems with many local optima. Step-by-step parameter settings are provided for the modified TS. Starting points and ranges are recommended for each parameter. Based on the initial runs, further adjustment following the guidelines will provide better performance for specific problems.