Rule-based Improvisation Instruction systems?

[omphalopsychos]

In response to the Miles comment, I want to clarify that machine learning doesn't just work for imitation. Miles would be an interesting case study, but the approach there would be very different from a Parker bot. Essentially, all you're doing with ML is you're trying to optimize a set of parameters to a function by minimizing cost on some output from the function. That cost can be defined as error against some "true" data but it can also be a penalty on undesired outcomes. So, you say a computer can't feel jealous or competitive; I respond: kinda. You can assign penalties to different outcomes, again without have to explicitly program the internal state of the bot. Think for example of how Youtube recommends videos to watch or Spotify recommends music. They are applying mostly techniques using interest similarity among users, but they also optimize performance of their recommendations based on user feedback. That's part of the reason why the "like" button is so valuable in social media.

Here's an example of a non-mimetic algorithm getting created. The programmers made a bot that had a goal: "walk". They didn't teach it by showing it examples of humans walking. Instead they let it move at random, bumping into things, crawling, rolling around, and generally spazzing out, like a child. They assign negative penalties for bad outcomes in order to reinforce good behavior. Over sufficient time iterating and training on performance, the bot learned not just how to walk, but also how to run, jump, avoid walls, etc.

[christianm77]

Optimise by a reward system based on badly paid gigs.

[zirenius]

Wow, this thread has grown a lot since I took a little time off and went on vacation.

Thank you very much to all contributors for a lot of interesting reading and discussion to get back into. Back aboard the ship I go!