Imitation Game-Playing Competition

[Maciej Świechowski]

Dear fellow scholars,

I decided to share a small and rather less technical paper I have just published, which is basically two things:

- a mini survey of game-AI-related competitions 

- a motivation for a new type of competition: Imitation Game-Playing Competition

This is a preprint, but the conference has just ended and it will be indexed later in scientific databases:

https://annals-csis.org/proceedings/2020/pliks/126.pdf

I used to work in academia but now I am working mostly in the game industry (games and tools) and I'm guessing it can be surprising to read that such a competition would be beneficial from the industry point of view as well.

Best wishes,

Maciek

[Julian Togelius]

Interesting. The survey is nice but very focused on what's running
right now - there's a whole bunch of other game-playing competitions
from the recent past that are not mentioned. (See competition lists
from past CIGs/FDGs/AIIDEs for a partial list.)

One that is particularly related is the 2k BotPrize, which was
eventually won by two different bots that were largely based on
imitating human players they encountered. So it was a competition
about believability where imitation turned out to be the best road to
believability. (There were other interesting approaches too.)

In this "imitation game" (not Turing's use of the term :) ), would the
agent imitate behaviour in *the same* levels/problems, or different
levels/problems than it was trained on? This changes the character of
the competition radically.

Julian

> --
>
> ---
> You received this message because you are subscribed to the Google Groups "Procedural Content Generation" group.
> To unsubscribe from this group and stop receiving emails from it, send an email to proceduralcont...@googlegroups.com.
> To view this discussion on the web visit https://groups.google.com/d/msgid/proceduralcontent/CA%2B6fXQquEdkxUsV5beU8MUkwgi_50rK7b8Dy9T5Vwp4A1EQbeg%40mail.gmail.com.



--
Julian Togelius
Associate Professor, New York University
Department of Computer Science and Engineering
Editor-in-Chief, IEEE Transactions on Games
jul...@togelius.com | http://julian.togelius.com

[Maciej Świechowski]

pt., 11 wrz 2020 o 23:25 Julian Togelius <julian....@gmail.com> napisał(a):

Interesting. The survey is nice but very focused on what's running
right now - there's a whole bunch of other game-playing competitions
from the recent past that are not mentioned. (See competition lists
from past CIGs/FDGs/AIIDEs for a partial list.)

Yes, indeed, I have chosen the recent ones in this paper and this is a rather short paper, overall.

I remember the competition that was played using the Unreal Tournament engine but forgot it was called 2k BotPrize.

 So it was a competition about believability where imitation turned out to be the best road to
believability. (There were other interesting approaches too.)

That is nice to hear.

 

In this "imitation game" (not Turing's use of the term :) ), would the
agent imitate behaviour in *the same* levels/problems, or different
levels/problems than it was trained on? This changes the character of
the competition radically.

Transfer learning could be the next step.

My idea was just to train them on the same problem but with various "human role models", so this way they can display variety.

Thanks for comments :)