The presented parallel GA computes densely packed placements for the modules of a circuit with short wirings for the signal nets. The main reasons for enabling the optimization of this real-world combinatorial problem by a GA are
Strategy adaptation returns to the GA some of its robustness, which may have been lost due to the inclusion of problem specific knowledge in the operators. It has been shown that especially the development of the crossover/mutation rate differs for the presented problems. Without adapting this parameter during the optimization, search would be less efficient.
After computing the global routes in the current implementation, the width chosen for a channel is too high in most cases, because one track is added for each net inside a routing region. This leads to layouts with an about 10% larger area than the best results published in literature. Including a sophisticated detailed routing procedure before fixing the widths of the routing regions would avoid this drawback. Nevertheless, the main strength of GA described in this paper is the integration of the routing into the placement, which are usually separated tasks in other (genetic and non-genetic) approaches. Here the exact positions for the modules are fixed not until the computation of all global routes.
Further work in this project will be the implementation of a different recombination operator. Generating an offspring out of the trees of two parent individuals is not very efficient. On the one hand two disjunct subtrees have to be chosen, which decreases the number of potential pairings, and on the other hand a third part has to be added to produce a complete layout. Multi-parent crossover like it is used by Eiben et al. [5] or even implementing a gene-pool recombination operator like it is analysed in the work of Mühlenbein and Voigt [12] might improve the performance of the GA.