Statistical discrimination in learning agents

TL;DR

This paper investigates how information processing biases lead to statistical discrimination in learning agents, using a theoretical model and multi-agent reinforcement learning experiments, revealing architecture and environment influences on bias.

Contribution

It introduces a theoretical framework and empirical analysis demonstrating how agent architecture and training data bias affect statistical discrimination in social dilemmas.

Findings

Statistical discrimination arises from both population bias and agent architecture.

Recurrent neural network agents show less discrimination than other architectures.

Bias persists in agents trained on biased populations despite learning.

Abstract

Undesired bias afflicts both human and algorithmic decision making, and may be especially prevalent when information processing trade-offs incentivize the use of heuristics. One primary example is \textit{statistical discrimination} -- selecting social partners based not on their underlying attributes, but on readily perceptible characteristics that covary with their suitability for the task at hand. We present a theoretical model to examine how information processing influences statistical discrimination and test its predictions using multi-agent reinforcement learning with various agent architectures in a partner choice-based social dilemma. As predicted, statistical discrimination emerges in agent policies as a function of both the bias in the training population and of agent architecture. All agents showed substantial statistical discrimination, defaulting to using the readily available correlates instead of the outcome relevant features. We show that less discrimination emerges with agents that use recurrent neural networks, and when their training environment has less bias. However, all agent algorithms we tried still exhibited substantial bias after learning in biased training populations.