Cognitive Risk-Sensitivity in Multi-Agent Coordination

Liu and Fang (2023) propose ARSP, where agents use risk-seeking bonuses to coordinate in general-sum games. But this assumes rational risk assessment. Psychology research (e.g., Merrick & Shafi, 2013) shows humans often exhibit cognitive biases like loss aversion or overconfidence. This idea asks: Can we model these biases as risk-sensitivity parameters in learning algorithms? For instance, an agent with loss aversion might overweight downside risk, leading to safe actions (defection in Prisoner’s Dilemma). We’d extend ARSP to include bias parameters learned from data (e.g., via inverse reinforcement learning) and test whether they better predict human behavior in experiments (Blume et al., 2002). This challenges the rationality assumption in game-theoretic learning and could improve human-AI collaboration by aligning AI risk-sensitivity with human biases.

References:

1. Learning and communication in sender-receiver games: an econometric investigation. A. Blume, D. DeJong, George R. Neumann, N. Savin (2002).
2. A game theoretic framework for incentive-based models of intrinsic motivation in artificial systems. K. Merrick, Kamran Shafi (2013). Frontiers in Psychology.
3. Learning Adaptable Risk-Sensitive Policies to Coordinate in Multi-Agent General-Sum Games. Ziyi Liu, Yongchun Fang (2023). International Conference on Neural Information Processing.