Evolution of stochastic bio-networks using summed rank strategies

Stochastic models defined in the stochastic pi calculus are evolved using genetic programming. The interpretation of a stochastic model results in a set of time series behaviors. Each time series denotes changing quantities of components within the modeled system. The time series are described by their statistical features. This paper uses genetic programming to reverse engineer stochastic pi-calculus models. Given the statistical characteristics of the intended model behavior, genetic programming attempts to construct a model whose statistical features closely match those of the target process. The feature objectives comprising model behavior are evaluated using a multi-objective strategy. A contribution of this research is that, rather than use conventional Pareto ranking, a summed rank scoring strategy is used instead. Summed rank scoring was originally derived for high-dimensional search spaces. This paper shows that it is likewise effective for evaluating stochastic models with low- to moderate-sized search spaces. Two models with oscillating behaviors were successfully evolved, and these results are superior to those obtained from earlier research attempts. Experiments on a larger-sized model were not successful. Reasons for its poor performance likely include inappropriate choices in feature selection, and too many selected features and channels contributing to an overly difficult search space.