DDraceNetwork - showroom

Swezzy 2024-07-25 12:04

amk

2

8

wrong bind xDDD

Xio

wrong bind xDDD

6

https://clips.twitch.tv/FriendlySlickPanWutFace-oPxDWic1NfFmnSR_

Stepfunn - I CAN'T

Watch Stepfunn's clip titled "I CAN'T"

1

2

omg thats me owo

1

2

1

5

Lil Motive Vert

omg thats me owo

Gay

2

4

4

18:27

A mobile player carried another mobile player through kobra3

18:28

He uses the new client

Neural network learning to beat a simple map

10

2

CovER

Neural network learning to beat a simple map

this so satisfying to watch, damn

1

CovER

Neural network learning to beat a simple map

LOL let it play some super hard map

2

CovER

Neural network learning to beat a simple map

teach each of them new tricks as generations progress, making them better the older they are, same as real players nod

Steinchen

LOL let it play some super hard map

Multeasymap kek

Just fly1 (they will never learn)

Steinchen

Just fly1 (they will never learn)

Idk if they can learn how to fly, but i'll try to tune the model to beat the real maps

CovER

Idk if they can learn how to fly, but i'll try to tune the model to beat the real maps

Weapons would also be really interesting :D

1

CovER

Neural network learning to beat a simple map

Is this just a visualisation? I hope so. You can compute physics hundreds of times faster than realtime.

Teero

Is this just a visualisation? I hope so. You can compute physics hundreds of times faster than realtime.

Yes, it's just a visualisation

Nice

is the src available or closed?

Teero

Nice

is the src available or closed?

Closed, i think i will open it after some time

1

CovER

Neural network learning to beat a simple map

omg, how is that works

Teero

Is this just a visualisation? I hope so. You can compute physics hundreds of times faster than realtime.

It learns at speed of 5760 ticks per second

Hmm that's kind of slow no? I can get up to 3-4mil

If i remove learning and bots would run randomly speed would be 170000 TPS (edited)

kk

Teero

Hmm that's kind of slow no? I can get up to 3-4mil

On which hardware you get 3-4 mil? Or how you would do that?

I'm using an optimized version with multithreading that's prob why.

1

19:27

dw

CovER

Neural network learning to beat a simple map

@Mr.Gh0s7

CovER

Neural network learning to beat a simple map

are you using python gym enviorment?

Tater

are you using python gym enviorment?

No, i'm using libtorch and the environment is DDNet) (edited)

btw it will train faster if you replace all the air in the map with spawn tiles

Tater

btw it will train faster if you replace all the air in the map with spawn tiles

Good idea, thanks

CovER

No, i'm using libtorch and the environment is DDNet) (edited)

so you wrote all the code in c++?

Tater

so you wrote all the code in c++?

Yes

woah nice

help

CovER

Good idea, thanks

you may also want to apply random velocity when they spawn, but it's optional

Tater

you may also want to apply random velocity when they spawn, but it's optional

I spawn them a little bit randomly

19:51

I change the x coordinate a bit

yeah but if you just have the spawn tiles at the start then they can't learn the ending of the map until they reach it

Tater

yeah but if you just have the spawn tiles at the start then they can't learn the ending of the map until they reach it

Yeah, you're right

19:52

I will try to spawn them not only at the beginning

19:52

That's a really good idea i think

how long did you train it for that video?

6 hours

hmm

19:53

I think it should do better after 6 hours

But i dont use max of my CPU and GPU

19:53

I change hyperparameters to find the best to learn

which learning algorithm are you using?

I run 4 instances of learning at the moment

Tater

which learning algorithm are you using?

PPO

CovER

I run 4 instances of learning at the moment

4*64 tees?

Tater

4*64 tees?

Yes with different hyperparameters. They run independently

hyper paramaters for what, PPO?

19:54

NN size?

Tater

hyper paramaters for what, PPO?

Yes like epochs, mini batch size, learning rate and etc.

ah ok

19:55

I think you should still get better results

19:56

what do they receive for world data?

Tater

NN size?

4363 inputs 1024 units 512 units 1613 outputs

lol what?

CovER

4363 inputs 1024 units 512 units 1613 outputs

Layers

1613 outputs?

Angles

that's too many

Yeah i know

also you don't need a seperate output for each angle

19:57

you can let it control it's angle with 2 outputs

Tater

you can let it control it's angle with 2 outputs

Oh

direction = vec2(sin(x1),cos(x2))

Omg i'm stupid

19:58

I'll try that too (edited)

4363 inputs is also too many lol

I give him the blocks 33*33 grid around him

hmm

19:59

I guess it's an option

19:59

I never tried that

and multiply it by 4 because there is air, hookable, unhookable, and dead tile (edited)

we were using raycasts from 16 angles

20:00

times 2 for hookable and freeze

20:00

so 32 inputs

CovER

and multiply it by 4 because there is air, hookable, unhookable, and dead tile (edited)

isn't air and empty the same?

Yeah, i forgot something

20:01

Dead tile at the end

we had a pathfinding direction input to tell it where to go

1

Maybe relocate this convo to #somewhere else

2

lol