Metacognition in the Prisoner’s Dilemma

In this paper, we show ACT-R agents capable of metacognitive reasoning about opponents in the repeated prisoner’s dilemma. Two types of metacognitive agent were developed and compared to a non-metacognitive agent and two fixed-strategy agents. The first type of metacognitive agent (opponent-perspective) takes the perspective of the opponent to anticipate the opponent’s future actions and respond accordingly. The other metacognitive agent (modeler) predicts the opponent’s next move based on the previous moves of the agent and the opponent. The modeler agent achieves better individual outcomes than a non-metacognitive agent and is more successful at encouraging cooperation. The opponent perspective agent, by contrast, fails to achieve these outcomes because it lacks important information about the opponent. These simple agents provide insights regarding modeling of metacognition in more complex tasks.