A Comparison of Parallel Global Optimisation Algorithms for Reverse Engineering Gene Networks

The approach of reverse engineering gene networks by fitting models has been highly successful, but increasing model complexity means that more powerful global optimisation techniques are required for model fitting; for this, we require faster parallel algorithms. I examine the modeling of the gap gene network in Drosophila, for which a gene network model, in the form of a set of differential equations, is fitted to high-resolution spatio-temporal expression data. Previously model fitting has been performed with Parallel Lam Simulated Annealing, but it has been shown that an island Evolutionary Strategy would be more efficient. Until now a parallel Evolutionary Strategy has not been applied to this problem. I study in detail the performance of the island Evolutionary Strategy when applied to the gap gene problem, including the effect of the distribution of individuals across islands, and demonstrate that the per-island speed of the algorithm increases with number of islands. By splitting up the islands throughout a number of processors, I apply a new coarse-grained parallel Evolutionary Strategy to problem, and study how its performance varies depending on number of nodes. It is found that both the reliability and the speed of the algorithm increase with increased number of nodes, though the efficiency drops off beyond 20 nodes. Finally, I compare the performance of the parallel ES to Parallel Lam Simulated Annealing and find that the Evolutionary Strategy is both faster and more reliable than Simulated Annealing. I discuss ways in which model fitting by optimisation can progress in the future, including the potential for specific improvements that could be made to both the Evolutionary Strategy and Simulated Annealing. I conclude that as well as its superior performance to Simulated Annealing, the Evolutionary Strategy also has more possibilities for future developments, and I discuss such improvements, including asynchronous algorithms, hierarchical algorithms and