Statistical discrimination in learning agents

Abstract

Independent researcherUndesired 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 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.

Developing, evaluating and scaling learning agents in multi-agent environments

“…Statistical discrimination refers to the behaviour of estimating the quality of a social partner based purely on the statistical association between some readily perceptible characteristics and quality, without taking into account the specific quality of the individual in question [96]. In recent work [97], we studied statistical discrimination in a temporally and spatially extended context where individuals have the chance to progressively refine their estimate of the quality of partners, and where the limitations of learning agents are explicitly modelled. We discovered that an agent's information processing ability mediates whether they will learn to individuate their partners and choose them based on their intrinsic quality, or fail to do so and engage in statistical discrimination.…”