AI-Driven Replays

OmgItsGenoma·Feb 11, 2017·1 votes·505 views

Recently had a teammate claiming that if I had played differently in a certain trade the outcome would've been different. I disagreed and argued that there was nothing (at most, very little) significance I could've done at that situation. It almost got us into a sour state of affairs for the rest of the game.

Was thinking if maybe there would be a way to add a function during game replays where you can take the champion of your choosing and "take over" the replay where the AI could faithfully correspond with the patterns of movement and action from the rest of the players. With it you can try a different play-through and give an idea at how a certain situation would have turned out.

Thanks

5 Comments

Mysticman89Feb 11, 20171 votes

I think thats giving a huge amount of credit to the state of AI and machine learning that from a single game that an AI that'd play the same as a given player would be producable.

Heck, I feel like making an AI that plays in the same style and manner as a given player given all of riots data on the player over their entire gameplay history might be close to the pinnacle of what might be remotely feasible, and even then I wouldn't expect it to be done by riot, but rather experts in AI.

Keep in mind that they can't even make bots that remotely come close to being able to jungle properly (e.g. including ganks/counter jungling), and what they're capable of with lane bots is still a far cry from humans. Even with absurd tunings (i.e. doom bots), humans still eventually learn how to stomp them.

OmgItsGenomaFeb 11, 20171 votes

Yes I totally agree at the complexity of programming AI for League to be close to humans is really a huge effort. To be fair, the kind of AI that I'm talking about for the replays is rather a simple AI enough to react to the differences to the players "take over" input. My hope was that it can at least attempt to have similar objectives as the original replay data (an enemy player winning/killing, using an ultimate, or attempting to run away).

There is evidence of attempts, however, to combine state of the art AI with modern gaming. Blizzard and DeepMind partnered up to develop the most sophisticated AI bots for StarCraft 2; it's the most mechanically complex of games on the market. Here below is a work in progress video of how the AI would think:

(https://www.youtube.com/watch?v=5iZlrBqDYPM)

Studies have shown that providing DeepMind with a classic game like Space Invaders, it was able to learn through its mistakes, and eventually master it. It only took one night. See 7:25:

(https://www.youtube.com/watch?v=TnUYcTuZJpM&t=734s)

On the topic of StarCraft 2, it's hoped that one day DeepMind would be able to compete in tournaments. Such prospects to simulate human behavior in the present day are encouraging even for a game like League of Legends.

The real kicker of all this is the fact that an AI called AlphaGo in fact beat a grandmaster player in Othello. It's been considered the holy grail of AI to play in a way that it's unpredictable yet show signs of an understanding of intuition. It's commonly said that Othello is much more complex than Chess. This actually happened. Here's a nice little doc about it:

(https://www.youtube.com/watch?v=g-dKXOlsf98)

Anyway, I hope that was informative. It would be nice to see one day AI being accessible enough to add functionality like that.

Mysticman89Feb 11, 20171 votes

While relatively simplistic AI that might maybe be within riots reach might possibly be able to take over in a given situation, I feel like it would very rapidly diverge from the 'what if...' scenario you were trying to test in the first place, with lots of 'well yeah it worked out like that with bots, but X would have worked with people...' type counter arguments and then very little has been accomplished.

Custom matches are a thing, and combined with replays one can go back and double check the situation (health/items etc) in an order to recreate the scene if one really cared that much, so we as players do have the tools within reason if we have associates willing to step in to play the other roles.

It doesn't seem remotely practical for riot to really put any amount of resources into a replay->AI subtitution type thing though, as they wouldn't be able to do it well enough to achieve the intended goals.

It might be worthwhile more of an academic endeavor if a research group/company involved in state of the art AI wanted to team up with riot and see what they could accomplish, but I suspect AI research focus is going to primarily be focused on 1v1 style games (be it SC2 or more classical games like othello/Chess/Go) for a long time to come, as dealing with human teamwork is tricky to say the least.

OmgItsGenomaSep 7, 20171 votes

This is a bump to this thread regarding new developments. There's a new AI Bot specifically for DOTA 2. I wouldn't be surprised there would be experiments in LoL anytime soon.

https://www.youtube.com/watch?v=jAu1ZsTCA64