Optimizing the Neural Architecture of Reinforcement Learning Agents

Reinforcement learning (RL) enjoyed significant progress over the last years. One of most important steps forward was wide application neural networks. However, architectures these networks are quite simple and typically constructed manually. In this work, we study recently proposed architecture search (NAS) methods for optimizing RL agents. We create two spaces test NAS methods: Efficient Neural Architecture Search (ENAS) Single-Path One-Shot (SPOS). Next, carry out experiments on Atari benchmark conclude that modern find agents outperforming a manually selected one.