Survival games for humans and machines

Abstract

Survival games can be described as video games where the player searches for food and treasures, while avoiding obstacles and hostile attacks. Ms.Pac-Man and Minecraft are two well-known examples. Currently there are AI models that outperform human players at Ms.Pac-Man, while AI models playing Minecraft above the human level has been a long-standing challenge. This paper concerns what we call pure survival games, which take place in previously unseen worlds containing only food, water, and obstacles. The challenge of the player is to navigate and survive in those worlds by continuously finding resources and avoiding obstacles. Arguably, animals need to master physical analogues of pure survival games in order to survive and reproduce. Here we begin to explore human and machine performance on pure survival games. We define two games called the Grid game and the Terrain game and two corresponding AI agents based on deep reinforcement learning: the Grid agent and the Terrain agent. We explore to what extent these agents can match human performance and how their performance is affected by variations in their perception, memory, and reward models. We find that (1) the Terrain agent performs above human level, while the Grid agent performs below human level; (2) the smell, touch, and interoception models contribute significantly to the performance of the Grid agent; (3) the memory model contributes significantly to the performance of the Grid agent; and (4) the performance of the Grid agent is relatively stable under three quite different reward signals, including one that rewards survival and nothing else.