See, I think you're right. It would be very hard to make what most people consider to be an AI. But consider this: we see AIs everywhere in video games. Mostly, we see it in pathfinding. When we see those little testificates detect that it is night, and run inside, that is a pre-programmed response. But put something in their way, they will either jump over it, or move around it. Normally, this can't be done with a basic turtle programming, cause it will "bump" into the wall repeatedly until you stop it and tell it to turn around.
I'm going to throw in my two cents here (and ruin it for everyone :)/>/>) and say that pathfinding is
not artificial intelligence. Finding a method from travelling from one point to another isn't AI, it's just following A*. The AI in that sort of behavior is deciding where to move to- the basis of this however is still usually not AI either, it's just a series of more complicated algorithms with weighted heuristics and some random behaviors added for variation. How a traditional, connectivist AI might tackle this is choose points relatively randomly, creating a connected graph of all points it can move to. The movement points which lead to the highest success rate (doing more damage to the player or killing the player) would increase the weightings and consequently the chance of moving to that given point, until eventually a much smaller set of "optimal" points have been achieved. That's approaching artificial learning.
So that might look like:
local points = { ... } --All available points, populated at startup
function moveToPoint(index)
local path = points[index]
--pathfinding logic here
return moveSuccess --If our move was successful (dead player etc.)
end
function normalizeHeuristics()
local sumH = 0.0
for i=1,#points do
sumH = sumH + points[i].H
end
for i=1,#points do
points[i].H = points[i].H * (1.0/sumH)
end
end
function choosePoint()
normalizeHeuristics()
--This could be more sophistocated, if need be
local chosenH = math.random()
local sH = 0
for i=1,#points do
sH = sH + points[i].H
if chosenH <= sH then
moveToPoint(i)
break
end
end
end
That's a very shoddy implementation- in reality you'd want a large connected graph rather than just a set of points, so the turtle can choose optimal paths with weightings at each intersection, a little like how BigSHinyToys did above.
Another possible avenue is teaching turtles information about the minecraft world and allowing them to make
inferrences as well as decisions about it, that can also be considered learning though this requires guided rather than implicit teaching. An example is an assignment given to third year AI students at Flinders on teaching computers about family relationships (computer responses in
BOLD):
Jack is a boy
UnderstoodJack is Sam's parent
UnderstoodWho is Sam's Father?
Jack is Sam's Father.An ancestor is parent of an ancestor
Understood.Frederick is Jack's parent.
Understood.Who are Sam's ancestors?
Jack,Frederick are Sam's Ancestors. And you can go on to teach him about uncles and godparents and all sorts of other things. I sort of imagine this being used to teach turtles how to craft things- they randomly take ingredients from a chest, then arrange them in patterns to see what causes what. Successful combinations are stored and tested with later heuristics- ingredients that have a high rate of success (wood, sticks, cobble etc) would be weighted as more "interesting" to craft with, while reagents which have fewer possible combinations (redstone, smooth stone) would be "less interesting" and the computer would use them less often.
You could train it with a similar system as shown above:
Craft Sticks 4 with [wood, , , wood, , , , , , ]
Understood.Craft IronPickaxe with [cobble,cobble,cobble, , stick, , , stick, ]
Understood.How do you craft 2 IronPickaxe?
Craft 4 Stick with 2 wood, then craft 2 IronPickaxe with 6 cobble and 4 sticks.I might give that one a try. Of course will all experimentation learning is very slow so it might take awhile before he can do anything interesting but I'll post what I come up with.
AI, AI 
Creativity for the AI
