This is very impressive. Although, I can't help but get distracted by how cute 1:07 is.
That was pretty cool to watch! It's interesting to see how much stuff like this has come along.
Do they have some form of approximation of visual sensors. Or are they just reading memory and have God like knowledge of the map and entity position?
The agent is trained merely using visual input and auxiliary information you as a player would be aware of
When they show the first first person part it looks like the bot knows another bot is going to come around the corner and aims along the wall.
That could just be bots being fucking dumb though.
It's a result from emulating audio with a visual short range radar as the model does not support audio input.
So are we talking actually reading data from the screen, like machine vision? Or an approximation from reading memory, like "Agent X should see agent Y under these circumstances."
Messing around with machine learning is really fun once you get going. Personally I hope to see some ML Agents in racing games some day. I managed to get some agents to drive around a track but they were hilariously stupid and not really going fast enough to be considered racing. Still with all the buzz surrounding ML you would be surprised as to how easy the basics are. It takes more effort to create a traditional bot that needs predefined behaviors versus a neural network, you set up the parameters and let the network figure it out on its own. (Massively simplified but at its core that is what it is)
Yes - the convolutional stack in the model extracts salient information in raw pixel data. Here is an example which shows some of the principles behind the model in a less complex environment. The article can be found in the related blogpost
https://www.youtube.com/watch?v=LW20UbquVBU
In most cases they let the pixels of the screen serve as an input and it interprets what is going on at each frame. In my driving agents case I instead set up three sensors at the front of each car which allowed them to "see" in front of them. Making them all omnipotent would be nearly as interesting if you ask me.
Well that makes this about 100 times cooler.
That's the whole point Kan, they're "learning" through trial and error from the visual feedback.
Giving the AI insider information and telling them to aim at heads is so 20 years ago :P
I've seen people make "learning" bots in the past, for a variety of of games. Most notable was the Mario one, but that if I remeber correctly relied on reading memory in order to figure out where things were. It still had to figure out how to complete a level and not die. From what I've seen alot of game playing bots people have use rarely have a vision component, so I was a little reserved in thinking this was something new. But yea this is the bees knees. I wonder if cameflouge would work, or maybe the neural net would learn how to see through cameflouge very easily.
i really hope this means botmatches are going to return to battlefield
A player with good game sense would hear the movement and check the minimap, then once they realize it is an enemy they will react accordingly, completely valid behavior. Player movement in battlefield is loud af and easy to pinpoint
Yeah but I don't know what the AI are able to react to. MAybe they can hear, maybe they know the location of every other AI character, maybe they only know what is on screen right now.
Though according to wranglor they can hear.
And Battlefront.
Could it be? The return of bots?
So the bots are taught to play the objective? I would already pick them over most of my teammates.
Maybe that's the real goal of this. Teach bots to play the objective to better learn how to teach pubs to play.
Really hope they retroactively add them to battlefield 1, it would really help with being able to experiment with vehicles, weapons and tactics. As well as extend the game's lifespan. I still go back and play the older battlefields thanks to bots.
I'm thinking more along the lines of padding. Bot matches are nice, but imagine padding player numbers with bots. 32 players per team plus a mess of semi confident non cheating bots. Kinda liketitan falls grunts, but can actually contribute and isn't just fodder to shoot.
And have them dynamically prioritize players who buy less lootboxes. This is EA after all.
I wonder what would happen if somebody put these bots in a private server and forgot for 4 years.
it's pretty amazing, wondering if they're gonna implement it in BF
Are there any solid practical examples and tutorials for making ML bots? I've always wanted to give it a shot but its always seemed really complex compared to regular programming
This guy explains the theory quite well.
https://www.youtube.com/watch?v=aircAruvnKk
There are plenty of tutorials on youtube but the problem I found is that they always end up using a python library that contains a lot of ML functionality. Even the tutorials that claimed to be "from scratch" always used a library that does a lot of the heavy lifting for you. In my case I got some help from a friend and a lot of trial and error since I wanted to implement a neural network from scratch without any libraries inside Unreal Engine 4. It takes significantly less work than one might think to make something basic, like I did. Granted my driving agents are super limited and stupid but they do improve over time.
http://pahlavan.se/dump/MLCars.webm
The footage is low res since I did a poor job of recording it but the red car is the best car of the previous generation and after each run the two best cars are "bred" and a new generation of cars is spawned based on the two "parents".
Looking forward to have bots in games again but how are they going to have the code be used for games?
Would the best AI end up with hundreds of lines of codes ready to be packed and used?
Sorry, you need to Log In to post a reply to this thread.
