Reinforcement Learning – Behaviouralism for the Machine?

I was at a day long seminar/conference on campus yesterday showcasing a variety of technology projects that are running on campus. One of the sessions that I went to that was somewhat interesting (if poorly presented) was on reinforcement learning (RL). Basically they have the machine/program hunt for a maximal reward signal and from there devise an optimal means to achieve that state repeatedly. It’s certainly interesting stuff and it seems much like classical conditioning to me. This was mentioned in the presentation as it seems now that this theory grew out of animal behavior and is now being applied there as well.

Another theory that was examined during the session was Temporal Difference Learning – as a means of helping to solve/understand RL.  Generally, this feels to me to be much like scaffolding or perhaps even (if I read into it a bit more) constructivism where you base your learning on what you bring to the problem.

One of the things that people in Education are always having to deal with is the idea that there is no way to systematize what they do. And even though some people understand that the impact of a kindergarden teacher really isn’t seen until after a student is at least several years away from kindergarten, there are no numbers to point to. These theories may help some people cook up some numbers.  But even if one day we have robots that are human like, I think it’s still going to be next to impossible to use numbers to define how we teach our children, as there are far too many variables involved and too few of them can be controlled. That is unless we start teaching our children for specific tasks, rather than teaching them skills to survive a changing environment – which in itself negates ever being able to recreate a teaching experiment.