Hybrid evolutionary search methods for training an artificial neural network

In recent decades, hydrologic models have become increasingly complex in order to better simulate the physical processes that occur in watersheds. Highly detailed modeling, however, comes only at the expense of increased computational requirements. Such computational demand is typically not problematic while simulating hydrologic responses of a watershed for predefined land management policies; however, it is a major drawback when repeated simulations are required as part of an iterative, decision management approach. This paper presents a hybrid evolutionary search method for training an Artificial Neural Network (ANN) that will simulate hydrologic responses (e.g., flows, sediment and nutrient yield) and economic profits that can be generated as a result of particular watershed landscapes. The ANN is trained to mimic outputs of the comprehensive, but computationally intensive, hydrologic model known as Soil and Water Assessment tool (SWAT). The hybrid search method is derived by combining a Particle Swarm Optimizer (PSO) and the Back Propagation algorithm (BP). Test results indicate that the developed data-driven models are capable of simulating SWAT outputs with greatly reduced computational demands. The ultimate goal of this study will be to integrate this SWAT-based ANN for use in a watershed management decision model. Copyright ASCE 2005.