I’m not sure I buy the story about applying this to cooperation in any sort of complicated environment, among agents anywhere near human level. It seems like you need the “police” to actually know what the right thing to do is, and you need the agents to not be able to get around the policing.
Maybe the idea is that you could use this kind of policing in cases where you can’t just alter the reward function of the agent?
Interesting paper!
I’m not sure I buy the story about applying this to cooperation in any sort of complicated environment, among agents anywhere near human level. It seems like you need the “police” to actually know what the right thing to do is, and you need the agents to not be able to get around the policing.
Maybe the idea is that you could use this kind of policing in cases where you can’t just alter the reward function of the agent?