They very quickly figured out that they could win at this strategy by just embedding themselves in the ground so you couldn't reach them. I adjusted their code to try and prevent them from embedding themselves in the ground.
-
-
এই থ্রেডটি দেখান
-
They found a god damn glitch (specific set of parameters) that still let them do it despite the restrictions that I put in.
এই থ্রেডটি দেখান -
That's pretty much the reason that prototype got shelved, that kind of AI is terrible overall if you're trying to focus on something as abstract as 'fun' or 'challenge' that isn't clearly definable.
এই থ্রেডটি দেখান -
and yet thats the kind of algorithm that all the major content platforms on the internet use right now to curate their content and part of the reason the internet kinda sucks right now
এই থ্রেডটি দেখান -
if you optimize games for money you get mobile lootbox garbage. if you optimize videos for watch time you get vlogs that are 95% filler. if you optimize social media posts for view counts you get clickbait, lies, and freebooting
এই থ্রেডটি দেখান -
AI is really good at finding local minimums and maximums. Like "oh, promoting 45 minute videos gets more watch time than 5 minute videos". Its terrible at recognizing that promoting that kind of garbage will eventually cause people to leave for better platforms
এই থ্রেডটি দেখান -
this kind of AI curated garbage world is way more likely a scenario for an AI dystopia than a kurzweilian singularity. Step outside and The Algorithm decided that today's hot shit is clothespins. Buy all the clothespins. buy stock in clothespin co. watch videos about clothespins
এই থ্রেডটি দেখান -
all those thumbtacks you bought last week? yeah worthless now.
এই থ্রেডটি দেখান
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
did you let the AI play thousands of games against itself, or how did the genetic algorithm figure this stuff out so quickly? (I realize that's not the point of your thread, was just wondering because that sounds really interesting)
-
its not they were dirt simple. like 10 params to play with. each child would inherit its parents params then pertub them randomly. the winning child would exaggerate its pertubations before breeding the next round to try and make it hone in quicker
-
each degenerate case was the result of one or two params going extreme
-
ah so this happened over the course of just a few games that you played manualy
-
yeah the point was they should evolve as you play, so if you focus on say, big strong shots they would evolve to be smaller and faster. if you do lots of small fast shots, they would get larger and defend-ier
-
ah cool idea, thanks for explaining!
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
I did something like that using neural nets. AI learned their best strategy was to run away. They’d figure out where the player was, then all run and hide in the far corner of the map, occasionally sending a scout to check the player wasn’t getting too close.
-
Full marks as a cool story to tell at GDC, but I had to entirely remove the learning features from the AI because it simply wasn’t fun to play against enemies who had figured out that they actually couldn’t win in a fight.
-
Should have been proud of them for trying to stay alive. :P
-
Oh, I absolutely was! I mean, once I finally figured out why all the game’s enemies were vanishing after about 20 minutes of gameplay.
-
But admittedly, if they can learn by chance or design to cooperate, your player becomes the one taco in a world of deadpools. Shades of Walking Dead, but more brutal.
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
-
-
Looks like Mewgenics just got enemies
-
this has nothing to do with mewgenics
-
I know Isa joke
কথা-বার্তা শেষ
নতুন কথা-বার্তা -
লোড হতে বেশ কিছুক্ষণ সময় নিচ্ছে।
টুইটার তার ক্ষমতার বাইরে চলে গেছে বা কোনো সাময়িক সমস্যার সম্মুখীন হয়েছে আবার চেষ্টা করুন বা আরও তথ্যের জন্য টুইটারের স্থিতি দেখুন।