#66 – Michael Cohen on Input Tampering in Advanced RL Agents
Hear This Idea
00:00
The Importance of Random Mutation in Reward Learning
For some difficult and economically valuable tasks in the world, it seems critical that you need to learn as you go about what sorts of things you need to do in order to get reward. Yeah. Maybe this is kind of aside from the actual interesting parts here of the thought experiment, but I was just going to say in terms of the real world and stuff, that also kind of makes sense where you can't imagine just a reinforcement learning thing being allowed to do anything in the World. You could imagine there being some collateral damage if you released that into the real world. But maybe that doesn't matter if you train it in a simulated environment and that goes away.
Play episode from 30:07
Transcript
