Evolving brains in evolving environments

Abstract

The spectacular and headline-grabbing recent successes in deep learning have made little or no use of evolutionary algorithms. Has this left Darwin turning in his grave, like one of the Earthworms he was so fond of studying? Or is he flying high, like a Galapagos Finch, safe in the knowledge that the inferior land-borne creatures below him will soon reach the edge of their harsh island and realize that, to go any further without drowning, they will need evolution?We will discuss and illustrate two ways in which evolution may be on the critical path to Artificial General Intelligence, AGI. On the environment side, highly connected ecosystems of evolving agents --- such as the environments within which our own intelligence evolved --- may provide a 'natural curriculum' that naturally optimizes the nature and magnitude of the learning challenge presented to each of the agents in those ecosystems. If so, simulations of virtual ecosystems could become an efficient engine of innovation in AI. On the brain side, evolutionary algorithms and deep learning interact in various ways, e.g. searching over hyperparameters, copying weights in neural networks can contribute to continual learning, evolution can control which parts of giant neural networks learn, neural networks can learn mutation operators, reward functions can be evolved, and diversity maintenance mechanisms can be applied to reinforcement learning.