Article,
Neurocontrol I: Self-organizing Speed-field Tracking
Neural Network World, (1996)
Abstract
The problems of controlling a plant while avoiding obstacles and experiencing perturbations in the plants dynamics are considered. It is assumed that the plant's dynamics is not known in advance. To solve this problem a self-organizing artificial neural network (ANN) solution is advanced here. The ANN consists of various parts. The first part discretizes the state space of the plant and also learns the geometry of the state space. The learnt geometrical relations are represented by lateral connections. These connections are utilized for planning a speed field, allowing collision free motion. The speed field is defined over the neural represention of the state space and is transformed into control signals with the help of interneurons associated with the lateral connections: connections between interneurons and control neurons encode the inverse dynamics of the plant. These connections are learnt during a direct system inverse identification process by Hebbian learning. Theoretical results and computer experiments show the robustness of approach.
