Automation is sometimes seen as replacing humans. However, in most cases only some elements of tasks performed by humans is replaced.

For example, an autopilot does not replace human pilots in an airplane. It merely makes the task of flying easier for human pilots, still requiring vigilant supervision of potential failure modes of automation. For many more years, this state of partial automation will remain in safety critical tasks.

Situation is not very different in deremining and adminstering medication to a patient. Whether it is determining insulin dose, or chemotherapy dose, humans are still involved in the critical decisions, often aided by computers and software.

Therefore, human decision making process under partial automation needs to be examined closely. In addition, understanding the process of human decision making is important even for human behavior not involving automation per se.

Here we will explore this topic using Jupyter notebooks and Python for live interaction with the user. I want to systematically explore surprises that occur along the way. In doing so, I want to point out various method pitfalls relating to human decisions. Sometimes, these pitfalls are present in computer automation methods.

In these examples, I will first describe a problem and the methods used commonly. Then I will describe assumptions underlying the methods and show that it works as intended when used on a simulation that matches the assumptions. Next I will show what happens when used on a more realistic situation simulated using a more complex model. Granted that reality is always more complex than even the complex simulation, it can still show problems encountered in practice.

Topics

1. Prediction for Automation
2. Poor Control Inspite of Perfect Prediction
3. Errors and Uncertainty in Prediction
4. Interconnted Systems

It seems logical that if you want to control something, you need to measure it and predict it.

For example, if you want to achieve a particular health outcome such as blood glucose level or blood pressure, measuring it, predicting its response to nutrition, exercise and medication seems sufficient. Once you can predict it well, then achieving your goal of controlling it appears a matter of diligent execution by changing the inputs.

We all know from personal experience, that it is easier said than done.

Prediction for automation is different from generic statistical prediction because errors in prediction can have drastic consequences. Moreover, it is common that new modalities of the system (the environment in Reinforcement Learning) are discovered as a direct result of the use of the predictions to change inputs.