About Space Invaders and automated scaling

Description

Adobe Cloud Platform is the core component of Adobe’s cloud offerings, serving billions of requests per day. Fast reaction time to load spikes and cost-efficient scaling are essential. Therefore, Adobe is constantly evaluating autoscaling algorithms to incrementally improve efficiency.

Deep reinforcement learning (RL) success stories raise hopes for building a scaling algorithm that can learn from historical load data while still being able to react to unusual load spikes in an appropriate way.

Michael Friedrich and Stefanie Grunwald took an algorithm that successfully learns Space Invaders and other Atari Games and applied it to service autoscaling. If a RL algorithm can learn how to play Space Invaders, why shouldn’t it learn how to automatically scale IT services?

Join in to learn how they built an OpenAI Gym using TensorFlow to train a neural network based on historical load data. You’ll deep dive into the technical details of the changes required to adapt the algorithm to the new domain of scaling and see how the next generation of flight controllers for racing drones helped them to improve their results—and why flight controllers solve a similar problem as autoscaling mechanisms.