ABSTRACT
For many, even in the game development community, the goal of AI development is to make the "winningest" AI that one can from the perspective of the developer, which sounds like a reasonable goal. A group of players played a turn-based strategy game against two dozen AI opponents; ranked them on difficulty, realism, and fun; and explained what they believed each AI's strategy was. Complex techniques often achieved higher scores but seemed no more difficult, realistic, or fun to players than significantly simpler AIs. No factor correlated highly with how much players enjoyed playing a particular opponent. The AI considered the most realistic was a Monte-Carlo-based algorithm that was arguably the least human-like AI in the test. Use of available information, use of direct versus summary information, technique complexity, predictability, strategy obviousness, persistency, revenge, and hustling are eight properties of an AI opponent.
