Neural Processors, ETL Japan

At the Japanese Electrotechnical Laboratory (ETL) in Tsukaba, at team of scientists lead by Tetsuya Higuchi are developing microprocessors which implement neural networks. These neural processors are able to reconfigure themselves by using genetic algorithms. The hardware is able to change its function in real time so as to adapt to the task at hand. The chips have been used to control the behaviour of a small autonomous robot (image right), and to control the movements of a prosthetic hand.

The GRD Chip
In April 1998 the researchers manufactured what they call the "GRD Chip". GRD stands for "Genetic Reconfiguration of DSPs". The chip consists of a 100MHz 32-bit RISC processor and fifteen 33MHz DSP's. Together these components implement an evolvable neural net. The RISC processor is responsible for the genetic algorithm (GA). This GA produces chromosomes which configure the functions of the DSP's and the connections between them. Note that the GRD chip does not need a host machine for this reconfiguration. The chip is completely self-contained and can reconfigure itself in real time during operation. A number of these chips can be combined to build ever larger neural nets.

The Learning Robot
The GRD chip was tested on a small autonomous robot. The robot is cylindrical in shape with a diameter of about 25cm. It is fitted with 10 infra-red proximity detectors, 4 bumping sensors and two cameras. The infra-red detectors can detect obstacles which are less than 30cm away. The robot also has two motorised wheels. These are pre-programmed with 16 possible motions including forward movement and rotation.

The task of the robot was to track a brightly coloured ball and to move towards it. The robot learns how to avoid obstacles by on-line evolution of a model of its environment. The model is continually updated throughout its lifetime which allows it to adapt to changes in obstacle positions. The robot can also learn to cope when one or more of its sensors is disabled.

Links:
ETL, Japan: www.etl.go.jp/~ehw