Abstract

AI Safety Gridworlds are a suite of 2D reinforcement learning environments that test for desirable safety properties of an agent, such as correct objective specification and robustness.

We will first discuss the paper’s approach to formalising safety properties in environments. Next, we will demo some of the environments and discuss whether they are reasonable tests of desirable properties. Finally, we will discuss why certain algorithms (among variations of RAINBOW and A2C ) seem to have better safety performance than others.

This week’s talk is a good opportunity to get a big-picture view of AI safety from a practical perspective. No prior knowledge of AI safety is needed.

You can try the environments for yourself by cloning this git repo: https://github.com/deepmind/ai-safety-gridworlds/tree/master/ai_safety_gridworlds

Paper: AI Safety Gridworlds (Leike et. al., 2017) https://arxiv.org/abs/1711.09883