Learning to Use Money through Reinforcement Learning

Abstract

Money, in its various forms, has played a pivotal role in shaping civilisations throughout human history. By facilitating cooperation among strangers, currencies have enabled monumental advancements in trade, settlement, and migration that surpassed the limitations of barter systems and other early exchange mechanisms. But despite its benefits and ubiquitous nature, it remains a mystery how humans learned to use these mediums of exchange in the first place. This project explored the origins of money through reinforcement learning, chosen for its resemblance to human learning processes. Two multi-agent Q-learning models were designed, developed, and experimented on, drawing from recent research into safe swapping behaviours. The first model, or the “swapping model”, examines whether a population of agents can naturally learn the behaviours necessary to perform safe swaps with strangers in discrete meetings. The second model, or the “token model”, builds on this by investigating learned agent behaviours when inherently worthless yet persistent tokens are introduced in a continuous stream of meetings. With the versatility to represent a diverse range of swapping scenarios and interactions, these models provide valuable insights into the motivations behind money’s use and the fundamental requirements for a population to adopt such behaviours.