Double-edged sword role of reinforcement learning based decision-makings on vaccination behavior

Abstract

Pre-emptive vaccination has been proven to be the most effective measure to control influenza outbreaks. However, when vaccination behavior is voluntary, individuals may face the vaccination dilemma owing to the two sides of vaccines. In view of this, many researchers began to use evolutionary game theory to model the vaccination decisions of individuals. Many existing models assume that individuals in networks use the Fermi function based strategy to update their vaccination decisions. As we know, human beings have strong learning capability and they may continuously search for the optimal strategy based on the surrounding environments. Hence, it is reasonable to use the reinforcement learning (RL) strategy to reflect the vaccination decisions of individuals. To this end, we here explore a mixed updating strategy for the vaccination decisions, specifically, some individuals called intelligent agents update their vaccination decisions based on the RL strategy, and the other individuals called regular agents update their decisions based on the Fermi function. We then investigate the impact of RL strategy on the vaccination behavior and the epidemic dynamics. Through extensive experiments, we find that the RL strategy plays a double-edged sword role: when the vaccination cost is not so high, more individuals are willing to choose vaccination if more individuals adopt the RL strategy, leading to the significant suppression of epidemics. On the contrary, when the vaccination cost is extremely high, the vaccination coverage is dramatically reduced, inducing the outbreak of the epidemic. We also analyze the underlying reasons for the double-edged sword role of the RL strategy.