NNetView: A Real-World Neural Programming Environment

For studying and developing control structures based on neural networks for systems that behave in a real world niche (i.e. autonomous robots), a proper monitoring, visualization and control environment is a necessity. In this paper, a new type of robot control/neural network simulation environment, “NNetView”, is presented. The innovative aspect is a real-world graphical interface which is entirely based on a connectionists approach. The interface allows user-friendly construction of neural network architectures in a visual programming style. The environment enables robot programming while displaying the robot's internal states directly on the computer screen. Unit (neuronal) activity from 2-dimensional arrays, connection strength and statistical data can be made visible at demand. The easy accessible external world and the visual "neural programming" approach enables a large variety of researchers, coming from many different fields, to conduct their simulations or real-world experiments. With NNetView we want to contribute to a new standard for neural network based robot control methodologies. Conducted real-world experiments so far include bilateral robot control, classical conditioning based on color stimuli, optical flow and edge detection.