Learning

As we have seen, feedback can be used to control systems, by measuring the output, feeding it back, and producing suitable corrective action if needed. It all sounds simple, but much effort is needed to achieve good control - we have a three year degree Cybernetics & Control Engineering concentrating on this.

If a system is complicated, particularly if (as often happens) it changes as it operates, a more sophisticated control mechanism is needed, one which adapts to changes in the system - the system must be able to learn.

Learning is a feedback process - as typified by the statement 'you learn by your mistakes' - you do something, assess how well it was done, and on that basis you refine and next time you do it differently (hopefully better). Again the usual block diagram can be adapted here.

As a test bed for research into how systems learn, we have used our simple mobile robots. These can move around and perceive their environment through simple sensors. Students program the 'rules' to allow the robots to move around either avoiding obstacles, or follow objects - an interesting problem solving experiment.

Then we considered how such a device could learn those same rules - this is done by the robot using trial and error - they try an action, see if it was successful - if so, then that action is more likely to be used in that situation - if not the action is less likely to be used.
 

<< Temperature control | Neural networks >>