Machine Learning PvP Bot Plays LMS in Old School RuneScape

Here's some footage of a project I've been messing around with. It's a PvP bot that focuses on NH fights, trained via deep reinforcement learning on a RSPS, and then runs as a plugin on a third party OSRS client. It supports general NH fights, but is specialized for LMS/PvP Arena.

(copy of video description)
This video contains footage of an AI-based Old School RuneScape (OSRS) PvP bot playing Last Man Standing. The version of the ML model shown was trained using reinforcement learning on a simulated version of the game for 5 years of game time. The model learned through self-play, and was trained using PPO.

 

This is super impressive. How did you create the training data?

 

Thanks! It learned by training against itself on a modified RSPS. I plan to make a video going over how I did it in more detail, as well as open sourcing the relevant code

 

That is pretty insane, wow.

 

Very cool.

 

bots these days are insane lol

how much $ would u get if u sold it?

 

I'm not sure, but I don't have plans to sell it. I plan to open source most of the training code such as the machine learning stuff, and the RSPS used for training, but I have no plans to release the plugin (that actually lets you use the trained models on the live game) to not hurt PvP.

 

could this method technically be used in a pvm scenario? or an account building scenario? like, could you machine learn an ai bot with hundreds(or thousands) of hours of TOA or chambers?
im AI dumb, so apologize for my stupid question(if it is)

 

Theoretically, yes, but you'd have to code the entirety of those raids (or whatever you want the AI to learn) into a RSPS for the AI to be able to play them which would be a massive effort. If you tried to train on RS itself without a RSPS, it'd require far too much computational resources + membership costs for bonds for it to be worth it, and would be really slow since you couldn't speed up the game like you could with an RSPS.

 

this is sick dude...

 

Looks like it fails at anyone who recognizes it’s a bot and knows how to one tick flick. You take a lot of damage from the guy one tick ballista’ing you.

Please add weights to the overhead prayers and let it predict which next attack is most likely instead of being lazy and having it choose overheads based on what weapon the opponent is wearing.

 

I suggest training a ML model on millions of YouTube videos. You could make bank selling botted maxed accounts or rev caves pking bot farms.

 

I can assure you this is not how it works.

 

pretty sure I both know and was in vc with you while working on this lmao