Perspective Taking in Deep Reinforcement Learning Agents

What deep learning can tell us about higher cognitive functions like mindreading?

“…In our work (Labash et al, 2019), we implemented two agents ("subordinate" and "dominant") who were competing for reward. We investigated whether the behavior of the artificial agents reveals some rudimentary skills of perspective taking, similar to the chimpanzee work (Hare et al 2000;.…”

“…In the last section we offered a potential goal in the acquisition of mindreading skills: In the context of multi-agent competition (Tampuu et al, 2017), mindreading could emerge through a process of an agent trying to maximize the probability of reward (e.g. food intake) while avoiding competitive interference by other agents (Labash et al, 2019). However, we are of course not implying that maximizing the probability of food rewards is the sole goal that would lead to mindreading skills to emerge.…”