AI for Retail

CHAPTER 10Simulators and AI

Simulators and AI go hand in hand. Since AI can be scaled infinitely, why would you limit its learning to only examples from the real world? Why not let it learn in a simulated environment, where it can play billions of games getting billions of “years” of experience in just a few days?

Whether you are developing a ballistic missile defense system (BMD), a self‐driving car, a robot arm, or a self‐driving store, you should focus a meaningful amount of your time on building a simulator that is an accurate representation of what would happen in the real world. If you do, you can grow your training data infinitely and AI accuracy with very little cost.

When I first graduated from college, my first job was at Lockheed Martin, where I worked on ballistic missile defense (see Figure 10.1). The system is designed to detect the launch of an intercontinental ballistic missile and quickly react by launching its own missile to hit it out of the sky. This system is currently deployed on all Navy cruisers and destroyers. To test such a system in the real world requires millions of dollars and a year of preparation. But we may want to test this system way more often than that. Maybe we want to test it one hundred times a day on new threat types, with varying weather conditions, in positions all over the Pacific Theater. The right answer was to develop a massive simulator that can do exactly that. This is used extensively in the Navy for us to fully test our system ...