Evolutionarily Stable Strategies and Artificial Intelligence
https://www.quantamagazine.org/artificial-intelligence-discovers-tool-use-in-hide-and-seek-games-20191118/
Recently, researchers at OpenAI trained bots to play hide and seek in an environment which they were able to manipulate, and came across some surprising results. As initially predicted, the hiders found ways to barricade themselves into “forts” built from blocks in the environment in order to evade capture, and the seekers managed to use ramps in the environment to scale those forts and catch the hiders. However, when researchers locked those ramps in place away from the forts, an unforseen outcome arose: the seekers brought unlocked blocks to the locked ramps, climbed onto the blocks with the ramps, and “surfed” around the environment on the blocks to search for and capture the hiders.
Many people consider these findings to represent a model of evolution of intelligence, due to the bots evolving their strategies based on what they learn from their own actions and those of other bots. I also noticed from the video that they appear to be using evolutionarily stable strategies similar to what we learned about in lecture. The bots took turns being passive and aggressive depending on which stage of hiding or seeking they were at, and the most successful strategies appeared to be when a larger fraction of the seekers banded together in being aggressive, leaving a smaller fraction passive.