Evolving Agents that Learn More Like Animals🔗
Abstract🔗
Deep neuroevolution, a combination of deep neural networks and evolutionary algorithms, has recently shown to be a competitive alternative to other deep reinforcement learning approaches. In this talk, I will present some of our work on creating agents that evolve to learn complex tasks, such as playing games or controlling robots. I will show that these algorithms do not only allow agents to perform in simple environments but also enable them to (1) learn 3D tasks directly from pixels, (2) learn models of the world for rapid planning, and (3) adapt quickly to task changes through a biologically-inspired form of meta-learning.