GTC 2020: Learning to Route Using Multi-Agent Reinforcement Learning
After clicking “Watch Now” you will be prompted to login or join.
Click “Watch Now” to login or join the NVIDIA Developer Program.
Learning to Route Using Multi-Agent Reinforcement Learning
Amine Kerkeni, Instadeep | Alexandre Laterre, Instadeep
Learn about deep multi-agent reinforcement learning on GPUs. We'll schedule trains in the Flatland routing environment, although such scheduling is relevant to a wide range of industries. We'll present different reinforcement learning (RL) solutions to address this specific challenge. Our step-by-step approach will help you discover some of the main difficulties in designing multi-agent deep RL systems and help you solve these challenges using GPU compute. We'll present best practices for overcoming these hurdles. We'll also compare centralized and decentralized multi-agent deep RL solutions. Specifically, we'll look at the Deep Deterministic Policy Gradient as Independent Learner and Multi-Agent algorithms as a centralized learner. We'll provide working code, and you'll undertake a series of tasks evaluating how different solutions handle problem complexity and efficiency in GPU usage.