Replicating human bias through synthetic data generation using deep learning

Abstract

The growth of Artificial Intelligence (AI) emphasizes the need for greater understanding of human and AI interactions. Human models formed during interactions with AI are subjected to cognitive biases, warranting further study to improve joint performance. Mitigating bias involves real-time bias detection; however, the exigency of extensive training data poses challenges, particularly in research contexts where such data sets may be unavailable. We propose an algorithm that amalgamates reward shaping, reinforcement learning, and imitation learning to emulate human game play data with sparse training data. This algorithm is evaluated in two experiments: the first successfully replicates human game play data utilizing diminutive training data, and the second extends the initial experiment by accurately replicating the anchoring effect cognitive bias from human game play data. Conclusions and suggestions for future research are discussed.