SafeLife: AI Safety Environments Based on Conway's Game of Life
partnershiponai.orgWith some graphics this looks like it could be a rather fun game to play manually. I wonder if there is such a thing? Maybe multiplayer support as well?
Right in the readme it describes how to launch interactively via `safelife play puzzles`
Yes but it’s missing decent graphics, sounds, animations, imagination... etc
May be useful, but it seems to me that the reward function still is relatively easy to specify? Much of the difficulty in AI safety is due to specify what humans really want.
Perhaps the AI can observe a human playing the game and learn a reward function?
A subarea of AI research focused on learning the reward function is Inverse Reinforcement Learning. Here’s an article on it:
Learning from humans: what is inverse reinforcement learning? https://thegradient.pub/learning-from-humans-what-is-inverse...
The problem is very easy to solve if the reward function (avoid altering the green life patterns) is specified. The aim in SafeLife version 1.0 (future versions will add more safety problems) is to find an agent/architecture that naturally has conservatism with respect to side effects, without being told which particular side effects in particular are bad.
I see, thanks!
> Much of the difficulty in AI safety is due to specify what humans really want.
Much of the difficulty of programming (for someone else) is due to the same thing.
I am confused by how this is supposed to be useful. It seems like the researchers are defining side-effects as things that "disrupt the world" (of this life game) and training an AI to avoid this.
But this seems like at best one of a whole host unexpected effects one might consider. AI that discriminates in a way that society frowns on might not "disrupt the world" in such a visible fashion.
I don't see how one can get away with an entity doing stuff for you with that entity understanding your model of the world.
Yes, this is one specific safety problem -- there are many other RL safety problems that deserve high quality benchmarks too. See eg https://arxiv.org/pdf/1606.06565.pdf or https://medium.com/@deepmindsafetyresearch/building-safe-art... for discussions of the problem space.
Hasn't Conway himself been known to say that he don't like the Game of Life since it doesn't give rise to that many interesting mathematical conclusions and how everybody is focusing on it above some of his other math axhievments? Maybe finally he can find the use case he wanted for it. And maybe it also gives reason to all the hours people have spent researching and finding new structures in Game of Life.
What if the AI decides to ignore penalties due to human thinking being inefficient? Well, that shouldn’t be possible I think.