|
      
|
|
|
|
Biological creatures are controlled by large networks of simple computational units which process sensory input data, and drive muscle actuators. Despite decades of research we have no theory to guide us in the design of such networks.
In the last decade researchers have successfully used an approach based on genetic reproduction to "evolve" neural controllers for a variety of simple control tasks. These controllers have always used a large amount of state and bandwidth - typically, every aspect of network state is represented as a 64-bit floating point number, which approximates the continuity of the real world. This is expensive, especially when implemented in hardware.
In my research I am looking at the effects of quantisation of values on evolved networks, and also the evolution of discrete processing through boolean networks, cellular automata, and random asynchronous circuits. I believe these discrete systems are capable of acting as cyclic attractors and hence can produce patterns of activity which can control a robot.
