In what ways a machine learning model can be trained/improved using video games?
In what ways a machine learning model can be trained/improved using video games?
In what ways a machine learning model can be trained/improved using video games?
There’s a whole bunch of YouTube videos of people training AI to play games.
Here a pretty good example:
he said how can a model be trained by playing/using games, not how to train a model to play games I think
The specific example I linked to uses machine learning to try and master the game by playing it.
Most aim cheats these days use ML to avoid detection
I guess if you had an generalized AI meant to play video games, you would presumably need to feed it a bunch of games so it could get better.
Although I'd imagine it'd be difficult creating the "win conditions" for it to improve.
For example in a game like Super Mario Bros it's fairly simple - there's a point counter on the top right you can just feed into the model. In a racing game, you could just take the time per lap and make that the fitness measurement. This makes it so anybody can just download some machine learning library and an emulator and go to town.
But in a game like Skyrim or GTA- how would you measure fitness?
For that reason, I think you would probably need a very generalized AI like Chatgpt with some sort of interface that allows it to make decisions on the spot with an overarching directive.
It's an interesting question. I think we're very close to something like this, although it's prohibitively expensive right now. At least for an AI to do in real-time. Maybe if you had an emulator you could sort of do something like this if you're able to freeze the game every time you send a prompt to the AI.
Ie you're playing Skyrim. You take a screenshot of the screen and then turn that into pixel data. You take your current health. You take your inventory. You take your level, your mana, your quest log etc. Then you send that all to the AI and say "Hey, your goal is to complete the main quest line. What do you do next?" The emulator freezes the game, waits for a response. AI says "I will move towards the next village. Press W to move forward." then the interface sends the "W" command to the emulator. Plays a frame or two and then freezes again. Sends another prompt to the AI.
and so on and so on. That is technically possible right now, although like I said - very expensive. Every call to an AI like that costs $$$. The more data you send, the more expensive. And sending the screenshot data every frame or two is a lot of data. Not to mention you would presumably need many thousands to get anywhere in the game.
perhaps there are AIs you can run on your PC that won't be as good as chatGPT but maybe good enough to do this? not sure. I've been wanting to look into it but I also don't want to fry my graphics card
Mario AI
One of the things Notch (of Minecraft fame) was known for before MC was a Mario remake that infinitely generated levels and had API hooks for quantifying progress. This was used in AI competitions.
Similarly, one of the old Trackmania games has been modified to output position and track info around the player along with fitness measures, allowing someone to train AIs.
Conventional AI/ML: I... don't think so, unless they are models specially designed for playing games (with extreme examples being the likes of AlphaZero for chess/shogi/go). Actually I guess you could say that whatever DeepMind was doing was technically training/improving their models using video games? Case in point, their innovation on AlphaZero led to AlphaFold (which literally got a Nobel Prize) and the recent weather prediction tool
Generative AI: The field seems to have a major existential crisis due to running out of high-quality training data. So maybe there would be a way to use games to augment training data? I'm not an expert in Generative AI/LLM training so I'm not sure of the details
AGI: I'm certain an AGI model can be improved using video games but I don't think researchers even remotely have a concept of how to build an AGI model yet