AI Safety Gridworlds

Jan Leike, Miljan Martic, Victoria Krakovna, Pedro A. Ortega, Tom Everitt +3 more
11/27/2017
cs.LGcs.AI

Official code: google-deepmind/ai-safety-gridworlds (639 stars, Apache-2.0, last commit 2022-05-18).

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.

View on arXivView PDF

Code Implementations(1)

google-deepmind/ai-safety-gridworldsOfficial100%
Apache-2.0

This is a suite of reinforcement learning environments illustrating various safety properties of intelligent agents.

639127PythonNov 3, 20174 years agoApache-2.0

Similar Papers

A Dynamical Systems Framework for Reinforcement Learning Safety and Robustness Verification

89% match2025

Safety-Gymnasium: A Unified Safe Reinforcement Learning Benchmark

88% match1 repo5772023

Safe CoR: A Dual-Expert Approach to Integrating Imitation Learning and Safe Reinforcement Learning Using Constraint Rewards

88% match2024

DeepSafeMPC: Deep Learning-Based Model Predictive Control for Safe Multi-Agent Reinforcement Learning

88% match2024

Multi-Agent Constrained Policy Optimisation

88% match1 repo2382021

Safeguarded Progress in Reinforcement Learning: Safe Bayesian Exploration for Control Policy Synthesis

87% match2023

Cite this paper

@article{leike2017safety,
  title  = {AI Safety Gridworlds},
  author = {Jan Leike and Miljan Martic and Victoria Krakovna and Pedro A. Ortega and Tom Everitt and Andrew Lefrancq and Laurent Orseau and Shane Legg},
  year   = {2017},
  eprint = {1711.09883},
  archivePrefix = {arXiv},
  url    = {https://arxiv.org/abs/1711.09883}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/1711.09883.svg)](https://codeofpaper.com/paper/1711.09883)

Discussion