Examples of Memorable AI

Just thought i would list some issues beyond the black box issue that neural nets have. It is worth mentioning as you would need to know about these issues before designing an AI which uses them effectively

Beyond the issue of Neural networks being a black box it is also worth mentioning that they are statistical models. So they
always get things wrong a certain percentage of the time (no matter how accurate). there is no Neural network that provides 100% reliable predictive qualities. Your training model minimizes that but there will always be an error margin.

Also please don't forget to mention the performance impact of neural nets. They can be beasts on CPU/GPU

IMHO if you do use neural nets I would keep them isolated to simple tasks purely because of the overall performance impact. That way the actual networks stay small and your training models can stay pretty agile. Don't forget that the more inputs you feed into the network the larger it has to become and if I am not mistaken the growth rate is almost exponential. additionally the larger those networks become and the more training data they need to solve the signal to noise ratio increases drastically.

Most games use some sort of decision tree (just like Jimym shows) to determine AI behavior I would still use decision trees and they provide a nice way to generalize the data for use in a statistical model such as a neural net.

When building a solution using neural nets just keep in mind that they should be used to solve a specific set of behaviors and not a generalized set (they tend to be as bad as any other AI at those types of problems)