Celia Alexis Wagar's CritPoints

Thoughts on all the people saying they should put OpenAI and Deepmind into single player games to make the enemies smarter? https://www.theverge.com/2017/8/11/16137388/dota-2-dendi-open-ai-elon-musk

This entire line of discussion of is stupid. You can’t compare AI beating humans at multiplayer games with AI used in singleplayer games. The two are trying to accomplish fundamentally different purposes. AI in single player games is a Game Design issue, not a Technology issue. Developing AI that will beat players every single time at a multiplayer game usually isn’t a process of making the AI smarter the same way a human opponent can improve, it’s usually a matter of the AI having better timing, reaction speed, and knowing the best decision to make for each situation. AI doesn’t play the game the same way humans do, with our slow reaction time, our ability to only keep like 9 things in our head at maximum, our limited ability to simulate and predict what the game state will be. If you want to build a perfect AI that always wins at say, street fighter, you could just make it uppercut when you attack, throw when you block, and block when you uppercut, and the game is over. There’s already a number of bots like this at the top of the SFV leaderboards and the only reason they don’t win every match is because of netplay lag.

People keep saying they want smarter AI, like it’s a technology problem. As if shoving a neural net into an FPS is going to lead to games with amazingly deep strategy and tactics. An aimbot is gonna win 100% of the time versus you, and a sufficiently advanced neural net is going to end up being a slightly better aimbot for all intents and purposes.

AI in single player games is a game design problem. Shoving an AI that optimizes winning into the enemies of a single player game isn’t going to result in smart opponents that are fun to fight, it’s going to result in enemies that exploit the limited strategies that win 100% of the time. Making enemies that adapt to the tools you’re currently using doesn’t take crazy advanced learning technology, it takes basic if > then logic, or if you want to be fancy, counting how many times a player has done a thing recently, then performing the action that beats that thing the player is doing. It takes them building specific, recognizable behaviors into the enemy, because enemies in single player games don’t have AI that makes them behave more complexly, they have human designers who decide to animate additional actions the AI can take, and record additional lines of dialogue stating that they are doing those actions.

If you want AI that does different things on different difficulty levels, then play God Hand, Halo, DMC, MGR, or Bayonetta.

And of course, I posted this article to twitter already that sums up my thoughts on something like OpenAI for a multiplayer game as well.
http://www.wildml.com/2017/08/hype-or-not-some-perspective-on-openais-dota-2-bot/