"Decap" Deep Learning - Advanced Competitive Toribash AI
Decap Project banner, because I thought it looked cool. So now it's here.
As a lot of you know, Tom and I have taken the challenge of creating a competitive, deep learning AI for Toribash.
We have now, finally, decided to post something about it to the forums.
We've been collecting a lot of replays over this year, starting around March, 2021. We deployed our first Toribash bot, ReplayMaster, to join random ABD rooms and save the replays for us. This only cost us a permanent ban and a lot of convincing, but it's worked out and has helped us tremendously in expanding our data set. We have also received replays from members of the community which has also helped quite a lot.
The reason we've been collecting all of these replays was to try and create a competitive AI in Toribash, and we have done so using deep learning. Some members of the community have fought against it, but for those of you that haven't,
here are some examples of how it performs:
[ First ] [ Second ] [ Third ] [ Fourth ] [ Fifth ]
It generally performs pretty well, but it's far from our best. We're currently working on expanding inputs to take into account more variables, refining the input data set to be as good of replays as we can possibly filter without manual work, and a few various other alterations that will make this better. The AI in those example videos was only trained for about 53 hours, on roughly 17,000 "refined" ABD replays. The results are very promising and have inspired us to take this as far as we can.
If you're interested in being more involved with updates on the project, you can PM me on Discord [ ChrisUMB#2908 ] if you're interested in joining our private project server where we post updates and host rooms for people to fight against it.