Open environments and competition among species are two driving forces of Darwinian evolution that are largely absent in recent works on evolutionary approaches toward AI models. Within a given generation, faster impalas and faster cheetahs are more likely to survive (and reproduce) than their slower counterparts—leading to the evolution of faster and faster impalas and cheaters over time. Can these and other principles of genetics and natural selection guide us towards major advances in AI?
Due in part to co-evolution, wolves have evolved a social intelligence that allows them to hunt in packs—thereby killing animals much larger than themselves and earning a food reward that benefits the entire pack, not just individual wolves.
Indeed, a very recent study by OpenAI demonstrates the emergence of complex interactive behaviors among agents trained by reinforcement learning, despite a seemingly impoverished reward system. Notably, rewards were given for team (akin to species) performance, not individual performance—much like all the wolves in a pack are rewarded when a large kill is made.
Given the simple task of playing hide-and-seek, these agents learn to collaborate to perform the complex tasks that at first blush, seem to have little to do with the primary hide-and-seek objective. Competition between teams is a driving factor in the emergence of these collaborative within-team behaviors. Note that these agents were trained by reinforcement learning rather than evolutionary computation. However, the relevant point is that competition between teams of agents (or species) can promote novel and complex behaviors.
Baby wildebeests gain agile muscle control and navigation abilities with minutes of birth. Clearly evolution has given them an innate sense of gravity, physics, objectness, and an advanced sensorimotor control system. They do no need to learn these concepts and capabilities through their reinforcement learning system.
Peng Liu