[<< | Prev | Index | Next | >>]

Tuesday, September 25, 2001

Now and Then

I just walked through the door from the garage to the kitchen, and when I opened it I felt a twinge of caution that I not slam it into the fridge. In fact it's not actually a concern right now because that only happens when the fridge is pulled out (e.g., to clean behind it) which it's not. But I remember the first time I had it pulled out for a while, and I hit it with the door a number of times -- kicking myself every time after the first for being so stupid as to do it again...

It relates to a challenging AI problem -- how to learn temporal correlations. Generally speaking, the approach is to hold old state (but how much and for how long?), or some residual of it, and to send training signals back in time (mathematically speaking). I.e., when I bang the fridge and kick myself, that negative reinforcement should affect the residual prior state of opening the door, so that I'll be more alert next time I'm opening the door.

But it doesn't seem to work that way, and the approach has never been particularly practical or successful in AI either -- largely because of the assignment problem (knowing which part of which prior state is to "blame").

Of course, I did eventually learn not to hit the fridge. And today I think I figured out how I learned it... and how to get an AI to learn such things. And it gets around most or all of the problems with the other approach. And it's really just an extension of other mechanisms already in place.

And no, I'm not going to tell what it is. :P

[<< | Prev | Index | Next | >>]

Simon Funk / simonfunk@gmail.com