Hey

For some time now, I have been studying artificial intelligence and it's aspects. In this semester we have a class about expert systems. Our professor mentioned that they are used in games, and offered me to do a paper on this subject. The problem is that beyond some tutorial or dialog management, I could not identify where these systems are used, so that is my question - what can be done with expert systems in games. Of course, I'm only asking for guidelines, I am willing to do the necessary research.

In a broad sense, almost all game AI is an expert system at some level. In it's generic form, an expert system is a rule-based system designed to make sense of a very small, specific set of inputs. These rules can be as simple as a binary check on something (a la an if/then statement), a series of branches, or a weight-based system where you may say the amount of X is proportional to Y. All of these rules, however, are hand-coded at some point -- whether it be all out scripting of sequences or hand-tuned weights for coefficients.

The word "expert" is somewhat misleading in this arena, however. Game AI rules aren't so much designed to mimic what a subject-matter expert would use (although it is a good idea to use it as a guide) as they are crafted by what a designer wants things to be like (whether it be realistic, stylized, etc.). Regardless, the system is taking a playbook of human knowledge and deciding which "play" should be run at that moment based on matching it with the inputs specified. (In fact, as my language starts to overlap here, "playbooks" in tactical or sports games are very obvious examples of expert systems.)

So, of the common techniques in game AI, the following could easily be considered as (or similar to) expert systems:

FSMs and their transition rules
Behavior Trees
Utility-based methods
Scripted AI
Rule-based AI

The above list easily accounts for 90+% of game AI.

The following start to diverge from the true definition of expert systems:
Goal-based architectures
Planners
NNs
GAs

The main reason for that is those systems tend to be more deliberative than reactive. They are assembling the information and the output themselves. As such, they are also the ones that can produce interesting, unexpected behaviors (although when you simply consider chaos theory and the explosion of state spaces, some of the initial list -- notably utility-based methods -- can do a pretty good job of that as well).

Unfortunately, the same reason that allows for open-endedness in things such as NNs and GAs is exactly the reason that they are wrong for most types of games... that the designer (i.e. "expert") no longer has control of the experience.