Toribash
Original Post
DeepTR V2
Hello everyone,


After a little while of absence I'm back on the AI of toribash, this time i'm setting up a sequential neural network with a supervised simulation.
This allows the learning of any mods.


The project is under construction. But it works if you modify it for you.

Everybody can participate, as long as you have the knowledge.


By the way nobody will be able to beat this AI.


Project : https://github.com/Stolym/DeepTR_V2
Clan: Tokugawa
Programming: Github
Exciting, hope to see something of this soon. I recently created a deep learning Toribash AI with my good friend Tom, and so far it's been going quite well. We have a different approach to you, so I'm excited to see what yours does.

Once it's done, we should make them fight!

Good luck!
I think it's just a question of supervision.
Basically we have only one network, which is designed to anticipate and attack because it uses the reward of the current environment so win lost.


Let's imagine now that we do this "like a player":


Action "Move to Point (X Y Z)" 1 model
Action "Dodge" 1 model
Action "Jump Attack" 1 model
Action "...."


You see the principles and on top of that we put a new network that will use this action.


It's the same thing as motion matching with a neural network.
Because the 3D model has a skeleton.


Hope, Join me !


Also I forgot to mention a BIG detail actually this AI is not made for a mod that has structures.


the model scales itself on the floor ("normally" " 99% sure x) ").


PS again: For the data, if anyone has a huge replay file of mod lenshu3ng normal rules (> 5.4).


I also just had an idea we could already see a model that adjusts the balance of the body
Last edited by Chips91; 2 Weeks Ago at 02:40 PM.
Clan: Tokugawa
Programming: Github
New updates !


- More stable environment.
- Generation before learning.
- Save and Load model weights.


--------------------------------------------------------------------------------


After 3 hours of training I copy a little his opponent dodge but do not attack.


I think that this is due to the quality of the games and the learning process. In any case 63% ("Accuracy") of the moves are good !
Clan: Tokugawa
Programming: Github
Hello everyone,


Update !


i think i found the right model, for the lens.
But i noticed that it took a lot of time to learn on my PC.


That's why i call on the community to know if someone has a good PC :


- VRAM 12 go (RTX XXXX, PXXXX, Telsa VXXX)
- 4 GHz >= 8 Cores
- >= 32 go RAM

With that we can make learn AI in 3, 4 days.
Instead of a few weeks.

Also a last factor comes into play if we have enough neurons but i think YES ahah
Clan: Tokugawa
Programming: Github
Voilà


I've been training her for 2-3 hours and I've found that I have good replay of her fighting against her.


So in fact the red uke it is doing well ^^


now i have started a new training to do 4-5 hours. purple line


but i will have to go through paperspace because i have a slow gpu and i have to use my pc for other things.
Attached Thumbnails
Screenshot_10.jpg  
Attached Files
replay_file_1.rpl (27.9 KB, 0 views)
replay_file_4.rpl (44.7 KB, 0 views)
replay_file_20.rpl (22.3 KB, 0 views)
replay_file_27.rpl (39.0 KB, 0 views)
Clan: Tokugawa
Programming: Github