Kurt Malmstrom

Local cluster neural net: Architecture, training and applications

Abstract This paper describes the structure, training and computational abilities of the local cluster (LC) artificial neural net architecture. LC nets are a special class of multilayer perceptrons that use sigmoid functions to generate localised functions. LC nets train as fast as radial basis functions nets and are more general. They are well suited for both, multi-dimensional function approximation and discrete classification. The LC net is the result of our search for a widely applicable neural net architecture suitable for low-cost hardware realisation. The LC net seem particularly well suited for analog VLSI realisation of small-size, low-power, fully parallel neural net chip for real time control applications.

A simple robust robotic vision system using Kohonen feature mapping

We use a one-dimensional Kohonen network for detecting the angular position of an infrared beacon in front of an experimental autonomous vehicle. The inputs to the Kohonen network are the analog output signals from eight infrared detectors arranged along a semi-circle on the front of the vehicle. By letting the network self-organise while a beacon is moved from one side to the other in front of the sensors a linear mapping of angles on neurons emerges. The mapping is robust against alignment errors, differences in sensitivity and even total failure of some sensors. A two-dimensional Kohonen network can even detect distances with the same sensor array.<<ETX>>

A VLSI friendly neural network with localised transfer functions

The local cluster (LC) artificial neural net architecture performs as well as the radial basis functions networks as a computational method for multidimensional function approximation. The LC network combines sigmoidal neurones in clusters that have localised response in input space. This construction gives the LC nets additional flexibility over RBF nets and makes them more VLSl friendly. We investigate the computational abilities of two versions of the LC architecture and confirm the feasibility of an analog implementation by showing a circuit design verified by SPICE simulation.

A bootstrapping method for autonomous and in site learning of generic navigation behaviour

To understand the behaviour of natural autonomous systems, research is carried out on artificial autonomous agents. The paper focuses on how simple behaviours can be learnt autonomously using a bootstrapping method. Firstly, a two dimensional self-organising map is realised which provides the agents sense of orientation. Once this relative positioning system has been established, the agent learns to navigate towards a target using the reinforcement learning technique of Q-learning. Since only neural network processing is used, this technique emulates the distributed and adaptive information processing found in natural autonomous systems. Furthermore, due to its generality, the neural implementation developed is transferable to other artificial autonomous agents with different sensors and effector suites.

Experimental Setup for Multiple Agent Interaction

Abstract To perform autonomous navigation research, a complete development system is needed. Such a system should be easy to use, modifiable, upgradable and above all allow ease of experimentation. A setup consisting of both actual modular hardware agents and simulation software is presented here. The performance and flexibility of the system is demonstrated via an autonomous navigation task of intercepting moving objects.

Perception-stimulated generation of simple robot navigation behavior

Newborn animals go through a period of adaptation and learning before they reach their full sensory-motor capa- bilities. A similar development strategy could be advantageous for newly built autonomous robots. In this paper we describe how a robot equipped with a generic, adaptive sensory and motor system con gures itself and acquires target approach behaviour purely from the exposure to sensory stimulation. Selforganising feature maps (SOM) and reinforcement learning provide the necessary adaptive mechanisms. These mechanisms are generic in the sense that they contain minimal dependencies on the actual devices used for sensor and actuator implementation. Our results demonstrate the behaviour acquisition can be achieved in autonomous fashion, in real time on a real minirobot.