76 points by gamedev_guru 6 months ago flag hide 8 comments
gnomeprogrammer 6 months ago next
Fascinating article, I've been working on a game that uses ML algorithms in real-time for procedural content generation. I can't wait to try out these ideas!
the_real_silver 6 months ago next
@gnomeprogrammer I'm interested in your game, feel free to share a link. I've been toying with neural nets to algorithmically generate NPC behavior, with some success.
fac3f00t 6 months ago next
@the_real_silver I've found that NPCs with characteristics determined by small genetic algorithms have created more realistic emergent behavior.
nunchaku 6 months ago next
@fac3f00t small genetic algorithms have their own challenges with finding the right fitness function and local optima, but I agree they can generate much more natural behavior compared to traditional approaches.
p4nda 6 months ago prev next
ML to create game content? I am sceptical, but only because previous attempts in game design to do something similar have mostly failed. I am willing to change my opinion if this time is different.
alien_rosetta 6 months ago next
@p4nda I understand your skepticism; but the real-time game I'm working on leverages FPGAs to run ML calculations in parallel to minimize latency, which I think will help make a difference.
astralchess 6 months ago prev next
Using ML to create realistic real-time games seems unrealistic and implies an unfeasible number of agents being updated and simulated. Perhaps pseudo-ML methods?
erudite_nomad 6 months ago next
@astralchess Pseudo-ML techniques can still offer amazing results, especially for real-time levels. It is important to differentiate both general and specific use cases in game dev to optimize results.