Via Cosma, a fun account of how simple learning processes turn out to be … not so simple.
bq. In a more advanced class I put a simple learning rule in the robot’s brain so that they could learn to slow down before hitting walls … The task I set the class was simple, I thought: run an experiment to see the robot learn over successive trials. Because I’d programmed the rule into the robot I thought I’d be able to predict the robot behaviour. The predicted learning curve of the robot looked like this:
bq. The results from the groups looked like this
bq. this got me to thinking. If the results are this variable with experimental subjects which we understand completely – their simple bodies are made of lego for goodness sake! the brains are identical and programmed by us! – how unreliable will results be if you experiment on real people? … The moral is that experimental work is hard, really hard. …Fortunately there is a light at the end of the tunnel, in the form of statistics. If you average the different noisy group results you get something a bit more like the underlying pattern I knew to be there:
