A Comparison of Non-human Primate and Deep Reinforcement Learning Agent Performance in a Virtual Pursuit-Avoidance Task

We compare the performance of non-human primates and deep reinforcement learning agents in a virtual pursuit-avoidance task, as part of an effort to understand the role that cognitive control plays in the deeply evolved skill of chase and escape behavior. Here we train two agents, a deep Q network and an actor-critic model, on a video game in which the player must capture a prey while avoiding a predator. A previously trained rhesus macaque performed well on this task, and in a manner that obeyed basic principles of Newtonian physics. We sought to compare the principles learned by artificial agents with those followed by the animal, as determined by the ability of one to predict the other. Our findings suggest that the agents learn primarily 1st order physics of motion, while the animal exhibited abilities consistent with the 2nd order physics of motion. We identify scenarios in which the actions taken by the animal and agents were consistent as well as ones in which they differed, including some surprising strategies exhibited by the agents. Finally, we remark on how the differences between how the agents and the macaque learn the task may affect their peak performance as well as their ability to generalize to other tasks.