Intrinsically motivated reinforcement learning in socio-economic systems: The dynamical analysis

Abstract

We conduct a theoretical analysis of the effects of intrinsic motivation on learning dynamics. We study a simple example of a single agent adapting to unknown environment; the agent is biased by the desire to take those actions she has little information about. We show that the intrinsic motivation may induce the instability (namely, periodic oscillations) of the learning process that is stable in case of rational agent. Most interestingly, we discover that the opposite effect may arise as well: the cyclic learning dynamics is stabilized by high levels of agent intrinsic motivation. Based on the presented results we argue that the effects of human intrinsic motivation in particular and bounded rationality in general may appear dominant in complex socio-economic systems and therefore deserve much attention in the formal models of such systems.