Abstract by Joshua Greaves
Learning Inductive Biases With AutoRL
Deep Reinforcement Learning (RL) has seen rapid progress in recent years. Many researchers have created different building blocks that include inductive biases inspired by various aspects of human cognition in an attempt to create more intelligent machines. Despite all this research, it is unclear how practitioners should compose even the simplest building blocks to achieve strong results. We present AutoRL, which uses Neural Architecture Search (NAS) to search over deep neural network topologies and hyperparameters to find the appropriate inductive biases that lead to fast learning for deep RL agents.