Refining of Learning in Human Decision Making Models: A Simple Step towards Machine Unlearning

In this paper the challenge of human decision-making (HDM) modeling by gambling task is targeted, specifically focusing on the Iowa Gambling Task (IGT). Traditional models train on data generated during gameplay, where decisions are made by the subject, although at times these decisions are random, leading to the model's inadequate understanding of natural human behavior. Here, we proposed a novel approach by utilizing realtime data from the IGT to model the HDM process involving uncertainty, risk, and ambiguity. Methodology in this study incorporates decay reinforcement learning for expectancy value updates, a prospect utility function for calculating utility values, and a time-dependent choice rule to determine the deck selection in the next trial. Additionally, mean square deviation (MSD) is employed to measure the disparity between actual and simulated data. By imposing restrictions on the continuous selection of decks, particularly limiting the number of consecutive selections, is a way towards concept of unlearning through the faulty pattern in data. The significance of this research lies in its ability to mitigate the impact of random deck selections during model training. By imposing constraints and aligning the model with more realistic human decision patterns, our approach offers a refined and accurate representation of human decision-making in gambling tasks.