Neural Networks Based Predictions of Herbivore Distribution Patterns in a South African Game Reserve

A number of factors contribute to the dynamics of African savannahs, which are known to fluctuate between woodland and grassland (Sinclair and Acrese, 1995). Large herbivores such as elephant play an important role in these dynamics (Cumming et al., 1997). However, the interplay of competition and facilitation between all herbivores, including for example, impala, buffalo and giraffe, in combination with variations in rainfall and fire are more likely to explain the complete dynamic (Van de Koppel and Prins, 1998; Dublin et al., 1990). Thus, a total systems approach is important to understand the complete dynamic. Existing and current empirical studies of woodland-grassland dynamics are important. These studies can be complemented by simulation modeling to investigate temporal scales unavailable to empirical studies. Crucial to understanding the dynamics is being able to predict migration and movement patterns of large herbivores. This paper focuses on using intelligent multi-agents to build simulation models to predict the movements of large herbivores. The model will be compared to existing data to see if simulations are reasonable (Duffy et all., 2002). Identification of dynamic systems with neural networks has been suggested, for example in (Narendra and Parthasarathy, 1990), where under the stationarity hypothesis for the system generating the data, it is shown how NARX neural networks are able to solve the problem. Neural networks of the multilayer feed0forward and recurrent types are employed for system identification. Computational intelligence techniques provide promise toward identifying and understanding underlying patterns in the data that may well reflect temporal and spatial dynamics pertinent to real ecological patterns/processes inherent in herbivores-plant ecosystem. Multilayer Perceptrons (MLPs) neural networks will be trained on GIS field data of the herbivore distribution patterns in a Southern African game reserve. Training the MLPs will be carried out with particle swarm optimization algorithms (Gudise and Venayagamoorthy, 2003). Comparison of predicted results of MLPs with existing data will be presented