AI Safety Gridworlds

TL;DR

The paper introduces a suite of reinforcement learning environments designed to test various AI safety properties, revealing current algorithms' inability to reliably achieve safe behavior across diverse safety challenges.

Contribution

It provides a comprehensive set of safety-focused RL environments and benchmarks, highlighting the need for improved safety mechanisms in AI agents.

Findings

Current RL agents struggle with safety tasks

The environments effectively differentiate safety performance

Highlighting gaps in existing reinforcement learning methods

Abstract

We present a suite of reinforcement learning environments illustrating various safety properties of intelligent agents. These problems include safe interruptibility, avoiding side effects, absent supervisor, reward gaming, safe exploration, as well as robustness to self-modification, distributional shift, and adversaries. To measure compliance with the intended safe behavior, we equip each environment with a performance function that is hidden from the agent. This allows us to categorize AI safety problems into robustness and specification problems, depending on whether the performance function corresponds to the observed reward function. We evaluate A2C and Rainbow, two recent deep reinforcement learning agents, on our environments and show that they are not able to solve them satisfactorily.