CoinRun: Solving Goal Misgeneralisation

Goal generalisation is shown to be a solvable by way of the ACE concept extrapolation algorithm on CoinRun

ArXiv

ArXiv preprint

Patrick Leask

Latent mechanistic interpretability

Goal generalisation Misgeneralisation ACE CoinRun Alignment

Abstract

Goal misgeneralisation is a key challenge in AI alignment – the task of getting powerful Artificial Intelligences to align their goals with human intentions and human morality. In this paper, we show how the ACE (Algorithm for Concept Extrapolation) agent can solve one of the key standard challenges in goal misgeneralisation: the CoinRun challenge. It uses no new reward information in the new environment. This points to how autonomous agents could be trusted to act in human interests, even in novel and critical situations.