A test-bed for the Attainable Utility Preservation (AUP) method for quantifying and penalizing the change an agent has on the world around it. This repository further augments this expansion to DeepMind's AI safety gridworlds. For discussion of AUP's potential contributions to long-term AI safety, see here.

With this work, we improve AUP and introduce vAUP, a modular extension to AUP with different variations, which are applicable to environments with a no-op action and action-driven environments alike. This method allows to choose between variants based on the environments to solve the safety property of avoiding side effects and to optimize an agent for a correct reward function implicitly. We evaluated all introduced variants on the same AI safety griworlds and show that this approach induces safe, conservative and effective behavior.

Please see the corresponding slides for my talk during YTIC 2023 and the bachelor's thesis with all details and results.

Work under review at International IFIP Cross Domain (CD) Conference for Machine Learning & Knowledge Extraction (MAKE) CD-MAKE 2023