Those Who (2019) is a virtual world simulation, consisting of artificial life forms that evolve over time, a set of self-propelled resource particles, and the environment within which both are embedded. Agents are trained to collect resources by a neural network using reinforcement learning, and their performance in doing so varies depending on their genotype. The simulation is implemented in Unity, using the ML-Agents toolkit to train the agents first to navigate the environment with intelligence, while a genetic algorithm works in real time to evolve new forms and behaviors as the simulation progresses. The quantities and values of resources in the virtual world can be linked to real-world inputs, such as the market for rare-earth minerals. Focusing on the conceptual relationship between reinforcement learning and natural selection, the project poses fundamental questions about the similarities between processes of learning and evolution, as discussed by Leslie Valiant in the book Probably Approximately Correct. The underlying framework is intended as a playground or tool for the free exploration of behavior and the emergence of intelligence, of whatever kind.
https://garage.digital/en/those-who-2019
