Operator adaptation in evolutionary computation and its application to structure optimization of neural networks

Choosing the appropriate forecasting model for predictive parameter control

All commonly used stochastic optimisation algorithms have to be parameterised to perform effectively. Adaptive parameter control (APC) is an effective method used for this purpose. APC repeatedly adjusts parameter values during the optimisation process for optimal algorithm performance. The assignment of parameter values for a given iteration is based on previously measured performance. In recent research, time series prediction has been proposed as a method of projecting the probabilities to use for parameter value selection. In this work, we examine the suitability of a variety of prediction methods for the projection of future parameter performance based on previous data. All considered prediction methods have assumptions the time series data has to conform to for the prediction method to provide accurate projections. Looking specifically at parameters of evolutionary algorithms (EAs), we find that all standard EA parameters with the exception of population size conform largely to the assumptions made by the considered prediction methods. Evaluating the performance of these prediction methods, we find that linear regression provides the best results by a very small and statistically insignificant margin. Regardless of the prediction method, predictive parameter control outperforms state of the art parameter control methods when the performance data adheres to the assumptions made by the prediction method. When a parameter's performance data does not adhere to the assumptions made by the forecasting method, the use of prediction does not have a notable adverse impact on the algorithm's performance.

EVOLUTIONARY MULTI-OBJECTIVE OPTIMISATION OF NEURAL NETWORKS FOR FACE DETECTION

For face recognition from video streams speed and accuracy are vital aspects. The first decision whether a preprocessed image region represents a human face or not is often made by a feed-forward neural network (NN), e.g. in the Viisage-FaceFINDER® video surveillance system. We describe the optimisation of such a NN by a hybrid algorithm combining evolutionary multi-objective optimisation (EMO) and gradient-based learning. The evolved solutions perform considerably faster than an expert-designed architecture without loss of accuracy. We compare an EMO and a single objective approach, both with online search strategy adaptation. It turns out that EMO is preferable to the single objective approach in several respects.

The dual role of fragments in fragment-assembly methods for de novo protein structure prediction

In fragment-assembly techniques for protein structure prediction, models of protein structure are assembled from fragments of known protein structures. This process is typically guided by a knowledge-based energy function and uses a heuristic optimization method. The fragments play two important roles in this process: they define the set of structural parameters available, and they also assume the role of the main variation operators that are used by the optimiser. Previous analysis has typically focused on the first of these roles. In particular, the relationship between local amino acid sequence and local protein structure has been studied by a range of authors. The correlation between the two has been shown to vary with the window length considered, and the results of these analyses have informed directly the choice of fragment length in state-of-the-art prediction techniques. Here, we focus on the second role of fragments and aim to determine the effect of fragment length from an optimization perspective. We use theoretical analyses to reveal how the size and structure of the search space changes as a function of insertion length. Furthermore, empirical analyses are used to explore additional ways in which the size of the fragment insertion influences the search both in a simulation model and for the fragment-assembly technique, Rosetta.

Mutation-Based Genetic Neural Network

Evolving gradient-learning artificial neural networks (ANNs) using an evolutionary algorithm (EA) is a popular approach to address the local optima and design problems of ANN. The typical approach is to combine the strength of backpropagation (BP) in weight learning and EA's capability of searching the architecture space. However, the BP's "gradient descent" approach requires a highly computer-intensive operation that relatively restricts the search coverage of EA by compelling it to use a small population size. To address this problem, we utilized mutation-based genetic neural network (MGNN) to replace BP by using the mutation strategy of local adaptation of evolutionary programming (EP) to effect weight learning. The MGNN's mutation enables the network to dynamically evolve its structure and adapt its weights at the same time. Moreover, MGNN's EP-based encoding scheme allows for a flexible and less restricted formulation of the fitness function and makes fitness computation fast and efficient. This makes it feasible to use larger population sizes and allows MGNN to have a relatively wide search coverage of the architecture space. MGNN implements a stopping criterion where overfitness occurrences are monitored through "sliding-windows" to avoid premature learning and overlearning. Statistical analysis of its performance to some well-known classification problems demonstrate its good generalization capability. It also reveals that locally adapting or scheduling the strategy parameters embedded in each individual network may provide a proper balance between the local and global searching capabilities of MGNN.