Just like in nature, game AIs can be surprisingly clever when circumstance meets potential.
Take for example a simple AI that moves randomly in one of four directions: up, down, left, or right. This simplest of intelligence can be made much more interesting when generalized. Rather than binding the random number generator to movements, why not generate a number from the space of all possible actions?
So, let’s assume we add three new options: fight, pickup item, and drop item. Suddenly this random agent can interact meaningfully with the environment. It can defend itself or establish territory. It can also modify it’s surrounding or even use tools. All of this we get from a change of dice roll from 1 of 4 to 1 of 7.
Now think of what we could do with logic and incentive! Now we get into the messy part of classifying AIs. There are many ways, including standard theory, that we can approach this problem. Here I’ll just throw muddy ideas against the wall, and interested readers can check back later to learn what ideas stuck.
1: Does the AI have “free will” in selection of actions or plans?
2: Does the AI have a personal memory?
3: Does the AI have an inherited or shared memory?
4: Does the AI have a generically challenging environment?
5: Does the AI have a personally challenging environment?
6: How large is the effect space of the environment?
7: Are there other AIs present in the environment?
8: Is the AI turing complete?
9: Are there other turing complete AIs in the environment?
These are all I can think of right now. I am sure there are others that are equally or more important, but I guess those will have to wait.