Do Artificial Agents Reproduce Human Strategies in the Advisers’ Game?

Game theory has been recently used to study optimal advice-giving strategies in settings where multiple advisers compete for a single client’s attention. In the advisers’ game, a client chooses between two well informed advisers to place bets under uncertainty. Experiments have shown that human advisers can learn to play strategically instead of honestly to exploit client behavior. Here, we analyze under which conditions agents trained with Q-learning can adopt similar strategies. To this end, the agent is trained against different heuristics and itself.