#66 – Michael Cohen on Input Tampering in Advanced RL Agents
Hear This Idea
00:00
The Importance of Quantization in Reinforcement Learning
In inverse reinforcement learning, there is just real ambiguity about what we feel is going on. I think that if you just did reinforcement learning to try to make sure a person is satisfied, and they're just reporting their own satisfaction, that maximizing their satisfaction while not quite what we want would at least require a world that isn't pretty good. In the imitation case, well, it seems less potential for catastrophic outcomes,. There's less potential for being extremely useful or even just surpassing human capabilities.
Play episode from 02:13:01
Transcript
