Designing Autonomous AI

Chapter 8. Giving Your AI Brain the Information It Needs to Learn and Decide

Reality is complex and nuanced.

Most of the systems and processes for which you will design autonomous AI are like complex three-dimensional objects. Imagine you’re trying to explore the surface of a Picasso sculpture. The surface is asymmetrical and irregular, so to truly understand it, you’re going to have to look at it from different angles.

The same is true for decision-making. Humans and autonomous AI need to look at each process they optimize or system they control from a variety of perspectives. They try different scenarios and consider different variables during their practice. When designing autonomous AI, consider what information the AI will need to make decisions well under lots of different conditions. I like to start by considering what information I would need to make the same decision. If I don’t know, I ask experts.

Most AI that you design will learn by practicing in a virtual environment. This allows the AI to learn without experiencing the financial and safety consequences of trial and error on the real system. Virtual environments also usually significantly accelerate (sometimes by orders of magnitude) the training by allowing much more practice in the same amount of time.

Simulations are the virtual environments that provide the training gym where your AI will practice. They are the digital version of sensors that measure what’s happening in the real system. As a machine teacher, ...