The graphic on the screen basically explains that a special interface is created for the A.I. to directly interact with the game (top picture).
The text in the middle of both picture explains that the operations from the A.I. and the probability evaluations and more create a feedback system.
The bottom picture explains that it is also possible to use a console simulator to run Duels at high speed with a text-only interface.
The two squares on the right basically explain that information from the Duel and board is organized and distributed to the A.I., which then responds with its inputs and requests. The A.I. agent (written in Python) is an A.I. framework / trained model.
Q: This looks very different from the usual screen that we see in Master Duel, could you give us a rundown of what we’re looking at?
Mr. Hattori: Yes. The first thing I would like to point out is at the bottom-left, we see a percentage. This percentage is the winning odds from the perspective of the moves of the A.I. In this bottom-left box, we can see the total estimated winning odds, but in the bottom right, we can see the top 5 moves. So we’ll be looking at the A.I. choosing the #1 move. Right now, the winning odds are low, but soon enough, there will be a move that heavily increases the winning odds, so let’s pay close attention to that.
(The A.I. uses The Melody of Awakening Dragon to search for Blue-Eyes Chaos MAX Dragon and is now able to Ritual Summon it, increasing the winning odds to about 80%)
Q: From my perspective, this almost looked like two CPUs Dueling each other, but it’s not just that, is it?
Mr. Hattori: Correct, these are not CPUs, but rather external programs operating Master Duel. In other words, machine learning is being used to teach the A.I.. The data of the users is not being utilized at all, and the A.I. is learning from scratch how to Duel by Dueling over and over, without even teaching it the rules.
Q: By “over and over”, around how many times would you say?
Mr. Hattori: It has trained Dueling the Solo Mode CPU over 100000 times.
Q: So even without teaching it the rules, you just make it Duel over 100000 times, and the result is this current A.I.?
Mr. Hattori: Yes.
“It has probably played more Duels than humans!”, “That’s some though training!”
Q: There is so much information on the screen, that looks so cool!
Mr. Hattori: Yes, we thought it would be interesting for whoever is watching to display what the A.I. was thinking on the screen.
Ken’ichi: We are game developers, however, we are not A.I. specialists. That’s why we thought it would be interesting to see what kind of A.I. could be created by a company that specializes in this field, and what kind of Duels would be like between these A.I.s, and that’s why we created this project. “Who can create the strongest A.I.?” Of course, it’s also possible for A.I. to Duel human players, so we would also like to see what kind of Duels would happen between “the world’s #1 A.I.” and “the champion of the World Championship”.
Q: So an outsider, someone not from Konami, can participate in this as well?
Ken’ichi: Of course, this is not a system that we will just give away for everyone to create their own A.I. At this moment, [protocols/rules/terms of service] are being created, and we are currently moving forward so that within this year, positions for A.I. Duel creators will be posted at the official Master Duel website.
Mr. Hattori: First, we would expect the participation of groups/organizations with A.I. production skills. If possible, competitions between universities or such.
“That would be so cool, with each university teaching their own A.I. and then pitting it against other universities!”
Ken’inchi: Yes, the A.I.s created by universities are very interesting.
“Yes, they have their unique quirks according to who makes them”
.
