What should an AI assessor optimise for?

TL;DR

This paper investigates whether training an AI assessor directly on a target metric is always optimal or if alternative metrics can lead to better predictions, revealing surprising results about metric choices.

Contribution

It experimentally explores the impact of training assessors on different metrics and shows that optimizing for more informative metrics is not always advantageous.

Findings

Optimizing for the target metric is not always best.

Some monotonic transformations improve assessor performance.

Logarithmic scores benefit classification score maximization.

Abstract

An AI assessor is an external, ideally indepen-dent system that predicts an indicator, e.g., a loss value, of another AI system. Assessors can lever-age information from the test results of many other AI systems and have the flexibility of be-ing trained on any loss function or scoring rule: from squared error to toxicity metrics. Here we address the question: is it always optimal to train the assessor for the target metric? Or could it be better to train for a different metric and then map predictions back to the target metric? Us-ing twenty regression and classification problems with tabular data, we experimentally explore this question for, respectively, regression losses and classification scores with monotonic and non-monotonic mappings and find that, contrary to intuition, optimising for more informative met-rics is not generally better. Surprisingly, some monotonic transformations are promising. For example, the logistic loss is useful for minimis-ing absolute or quadratic errors in regression, and the logarithmic score helps maximise quadratic or spherical scores in classification.