3 ms·
I think that’s only true if it’s a predictable reward function; if you add a random variable, in the words of Mr. Miyagi, “no can defense.”
by siavosh 8y ago
I think that’s only true if it’s a predictable reward function; if you add a random variable, in the words of Mr. Miyagi, “no can defense.”
- oblio 8y agohttps://www.joelonsoftware.com/2018/01/12/birdcage-liners/ https://www.joelonsoftware.com/2018/01/12/birdcage-liners/ > How do we prevent extinction? By only giving treats some of the time. So the dog learns something more subtle. When my master says come and I obey, I might get a treat. Sometimes I do, sometimes I don’t. That way, if I obey and don’t get the treat, I shouldn’t panic. I should still always come when he says come because that’s still the best way to get the most treats. Intermittent reinforcement works better. > This sounds like what Facebook was doing to me.