Why an AI for Outwitters is hard to do

[Kamikaze28]

What you are describing here is a reflex-based agent: "if this, then do that.". You can cobble together a heuristic all you want but it will still remain a predictable and uncreative opponent which can be defeated by knowing its reflexes.
Case in point: you could guide your agent into ambushes all day long, it will always fall for it.
Furthermore you have to evaluate everything after every move or spawn since you could discover previously unknown enemy units which would change the outcome of your "thinking" process.

Here's the next problem in your approach:

(04-12-2012 05:12 AM)Harti Wrote:  - reach+kill opposing Special unit
- heal unit (if damage is high enough)
- reach+attack opposing base
- reach an opposing unoccupied wit flag
- reach a neutral unoccupied wit flag
- reach+kill unit on neutral wit flag (and get it yourself)
- reach one's own unoccupied wit flag (for defending purposes)
- reach+kill Sniper/Medic units
- reach+kill Heavy units
- reach+kill Soldiers
- kill Runners nearby
- etc...

Your heuristic basically has to put all of these possible actions in an order of importance (preferably based on the current game state). You haven't even listed all possible "sub-goals" and with these 11 examples, there are already 55 comparisons to be checked by possibly every unit on your field.

The reflex based agent will not plan ahead, not even for the next turn. It will not anticipate possible enemy movements based on previous observations.

My estimation is that even new players would be able to defeat this AI by the 5th game if not earlier. And after that, they can consistently defeat it by exactly reproducing the same moves as before, since the AI will react exactly the same way. All you would get out of this is a tutorial after approximately months of development. It's just not worth it in my opinion.

Fun fact: the conversation-AI equivalent of your idea would be ELIZA.

Sorry to disappoint you, but I hope you found the mental exercise worthwhile