Artificial Intelligence Safety And Security Independent Study Week 9

This week I won't be doing any exploratory reading. Instead, I will be studying curiosity rewards in open ended problems, such as Game of Life or Minecraft. I will start by getting a reference implementation running, and plugging in a game of life simulator with some RL-friendly interface to take actions in the game. I think the setup be like this: There is a 400x400 board of active GOL pixels that the AI can observe. In the middle of it, there is some limited fence where the AI can do various things. This way, the AI can take actions within this limited fence. However, a 400 bit output channel is a bit much for most RL learners, so instead I would provide a few actions. Here is where I diverge into a few different schemes:

  1. Clipboard/Registers: The AI can draw boxes on the visible pixels, and copy them to a clipboard. It could even have multiple clipboards. This way, if it likes a glider it can copy it and paste several of them or rotate them and experiment with this
  1. Random Starting: A way to find patterns in GOL is to put down noise and watch for interesting patterns to evolve. There should be a button to replace the AI's control area with random noise, for when it wants to sample interesting stuff.

  2. Drawing from scratch: It should have a way to draw arbitrary shapes, with some sort of pen or stamps.

  3. Play/Pause: It might want to pause the simulation, and draw or copy and paste inside its control area. A healthy AI would only pause for short times, because the paused game is not only boring but also unlikely to give a reward. Which leads to...
  4. Extrinsic rewards: No-reward RL is good, but this setup could also be used in a weakly supervised way, or at least with some secret goal states that I expect. Some goals I think are meaningful:

As much as I want to do this, I'm having trouble working on it. Maybe if I could make it into a real AI gym and convince myself that other people would use it, I could do it.