生物網路逆向工程之演化式最佳化計算

Abstract

S-system 的數學建模是從連續時間序列的數據來進行，它提供我們一個相互作用的網路模式。 然而系統識別是一個棘手的工作，因為S-system被形容成一個高維度的非線性微分方程組，很多研究採取各種進化計算的技術來識別系統的參數，有些更進一步完成結構識別的網路架構。然而所刪除之多餘動力階數(redundant kinetic orders)相對於應該保存下來的而言，並非足夠的小到可以明顯的區分開來。另一方面，如何在準確性和計算時間上取得平衡也是一個重要的議題。對於結構識別，我們發展了一個新的基因演算法來提高收斂性並維持多樣性，所提出之開拓和開採的基因演算法(EEGA)能改善最佳的個體和確保全域的搜尋，並且為了確保得到一個合理的基因調控網路我們引進模糊合成(fuzzy composition)來推導重建指標，這性能指標讓EEGA擁有自我交談的多目標學習，提出的基於模糊重建的多目標基因演算法確保修剪行為是安全(修剪的門檻設定在10-15)。 對於參數估計，我們發展兩個最佳化演算法以減少計算時間並確保正確度，以取代並改善基於人口的局部收尋能力之計算方法，我們提出一個相反的觀念;我們整合了隨機的操作進到梯度的最佳化，而非將梯度法融入隨機演化法中，這個技術擁有梯度法和進化演算法的優點。為了顯示這個技術在解決問題的品質和計算時間上的效能，學習的搜尋空間定在一個寬廣的範圍，且所有的參數初始化到一個非常不好的點。另外，我們進一步探討並分析生物系統的動態行為。

S-system modeling from time series datasets can provide us an interactive network. However, system identification is a tough work since S-system is described as highly nonlinear differential equations. Much research adopts various evolution computation technologies to identify system parameters, and some further to achieve skeletal-network structure identification. However, the truncated redundant kinetic orders are not small enough as compared to the preserved terms. On the other hand, how to make a trade-off between the accuracy (reliability) and computation time is important. For structure identification, we develop a new genetic algorithm for achieving convergence enhancement and diversity preservation. The proposed exploration and exploitation genetic algorithm (EEGA) can improve the best-so-far individual and ensure global optimal search at the same time. Further, to ensure that a reasonable gene regulation network is inferred, fuzzy composition is introduced to derive a reconstruction index. This performance index let EEGA possess self-interactive multi-objective learning. The proposed fuzzy-reconstruction-based multi-objective genetic algorithm (FRMOGA) show that a safety pruning action is guaranteed (The truncation threshold is set to be 10-15). For parameter estimation, we develop two optimization algorithms to reduce the computation time and keep the accuracy. Instead of improving the local-search ability of the population-based computational methods, an inverse aspect was proposed. We integrate stochastic operations into the gradient-based optimization, instead of incorporation of the latter into the former. This technology has the advantages in gradient-based methods and solves the problems in evolution algorithms. To show the performance in the solution quality and the computation time, the learning was implemented in a wide search space ([0, 100] for rate constants and [-100, 100] for kinetic orders) and all parameters were initialized at an undesirable point (the neighborhood of 80). Furthermore, we discuss and analyze the dynamic behavior of S-type biological systems.