I’m still a bit mad about the trains in grand theft auto games. The early ones had trains you could actually properly interact with rather than them just sort of mindlessly going around the map being indestructible juggernauts.
I want to be able to derail and steal it.
There are indie games that support derailing trains, so it ain’t hard
Oh that’s really interesting; I hadn’t considered racing games as a genre to benefit from this type of machine learning. I guess I figured there’s not so much to AI there that it’s necessary, at least when we already know the “ideal lap line” for cars to follow, but yeah it gets a lot harder when considering other drivers on the track and a huge array of unique car models with their own handling and performance characteristics.
I played Forza Horizon 4 and the Drivatars are pretty convincing. They make exactly the kind of mistakes on the track that I make and they can be challenging but beatable in a way that’s much more fun than any other racing game I had played before.
For more mountainous or thick forest areas this is understandably. It required combustion engine trains, simply because of steep mountains, where it’s difficult to put down power lines or forests with a lot of trees who can easily destroy power lines. USA however is mostly flat. Looking at some like Austria or Swiss, if I see this correctly, they also are on a good way. Here we have a lot of hybrid but in general our train transport is a mess of mixed.
I don’t know what it’s using specifically under the hood, but in Street Fighter 6 Capcom recently added a new AI opponent you can fight that they say is trained on actual player ranked matches and fights more like a human opponent. You can even have it try to mimic your own playstyle if you’ve played enough.
It can do some odd things and its mimicry isn’t perfect. But it definitely doesn’t feel like the typical high difficulty CPU opponent which uses things like input reading to react faster than a real player ever could.
You can train it in mirror matches, but the V Rivals that you can fight other than your own mirror are an amalgamation of a particular rank. There’s a whole lot of skill variance in Master rank alone, so it might be good for training me against Dhalsim, because hardly anyone plays Dhalsim, so no one knows the matchup, but it won’t help me learn how to beat Punk, specifically.
Yeah, there are some disappointing limitations for sure, but it definitely is interesting, and does at least feel more like a human player than the normal CPU opponents.
The challenge is that AI for a video game (even one fixed game) is very problem specific and there’s no generalized approach/kit for developing AI for games. So while there’s research showing AI can play games, it’s involved lots of iteration and AI expertise. Thats obviously a large barrier for any video game and that doesn’t even touch the compute requirements.
There’s also the problem of making AI players fun. Too easy and they’re boring, too hard and they’re frustrating. Expert level AI can perform at expert level, which wouldn’t be fun for the average player. Striking the right difficulty balance isn’t easy or obvious.
I wouldn’t mind an AI using unorthodox strategies, but yeah that’s a good point that fine tuning it to be fun is a big challenge. Speaking of “non-player-like behavior”, I wonder if AI could be used to find multiplayer exploits sooner, though the problem there is you don’t really have much training data besides QA and playtesters before a full release.
Historically, AI has found and used exploits. Before OpenAI was known for chatgpt, they did a lot of work in reinforcement learning (often deployed in game-like scenarios). One of the more mainstream training strategies (pioneered at OpenAI) played sonic and would exploit bugs in the game, for example.
The compute used for these strategies are pretty high though. Even crafting a diamond in Minecraft can require playing for hundreds of millions of steps, and even then, AI might not constantly reach their goal. Theres still interesting work in the space, but sadly LLMs have sucked up a lot of the R&D resources.
Dunno how much you played of the franchise, but if you got stuck early on (e.g. the dreaded Marble Zone in the punishing first game), maybe you could abuse save states? The franchise got several emulated releases, and I imagine it’s not uncommon for them to allow such a function natively. And at least to me, Sonic 2 plays much better and I remember kid me finding Sonic 3 even sharper.
Iirc there are also cheat codes. I definitely remember reading about them in a magazine back then and having the best time flying around this zone as super sonic.
I remember planting so, so many rings in a single place with debug mode enabled in Sonic 2 just so I could play through each stage as Super Sonic without effort- aside from the super slippery controls.
ECHO, the 3rd person action\puzzle game was a fun concept to script in your machine dopplegangers to learn on you (and repeat after you one of the set actions you can do) and reset every cycle.
I don’t think it would work by itself without such limiting.
I always got the impression it wasn’t a learning AI but rather a very limited “Has the player pressed the run button? if YES: AI can use run next cycle”
Yes it is, it’s 100% scripted. And yes, in the environment where you can do like 10 different actions, they start to do their routine adding ones that you used in that cycle before they get reset. In a sense, they act no more natural than monsters from a tabletop game.
But these do make me think that if we talk gamedesign with a LLM as an actor, it should too have a very tight set of options around it to effectively learn. The ideal situation is something simplistic, like Google’s dino jumper where the target is getting as far as it can by recognising a barrier and jumping at the right time.
But when things get not that trivial, like when in CS 1.6 we have a choice to plant a bomb or kill all CTs, it needs a lot of learning to decide what of these two options is statistically right at any moment. And it needs to do this while having a choice of guns, a neverending branching tree of routes to take, tactics to use, and how to coexist with it’s teammates. And with growing complexity it’s hard to make sure that it’s guided right.
Imagine you have thousands of parameters from it playing one year straight to lose and to win. And you need to add weight to parameters that do affect it’s chance to win while it keeps learning. It’s more of a task than writing a believable bot, that is already dificult.
And the way ECHO fakes it… makes it less of a headache. Because if you limit possible options to the point close to Google’s dino, you can establish a firm grasp on teaching the LLM how to behave in a bunch of pre-defined situations.
And if you won’t, it’s probably easier to ‘fake it’ like ECHO or F.E.A.R. does giving a player an impression of AI when it’s just a complicated scri orchestrating the spectacle.
For most games, it's not difficult to make AI that can absolutely destroy humans. But it turns out to be very difficult to make AI that feels like a fun and engaging challenge to a human. Hardest of all is making AI that realistically plays like a human does.
I’m playing the PC version of SMCP, and the only difference I can notice, maybe due to the better hardware, is that the game seems to be a bit faster on PC than on PS2. And have yet to test any of the other collections Sega made for/with the Sonic games.
The Rain World Animation Process.
While the title suggests only animation, the AI is tied directly into the animations so you gat a 2 for 1 deal in this video.
bin.pol.social
Najstarsze