Acknowledgement

[agemuend]

Hi Jimboom7,

thanks for publishing this work. I'm working on a free Skat game right now and would like to include a computer opponent based on your work. I know that the license allows it, but I still feel like it's more appropriate to ask..Would it be okay if I included your model as an AI opponent called SkatZero? Do you have a publication I could refer to or would you maybe prefer to be cited with your real name? Maybe my game will never be downloaded by anyone, but I still feel it's more appropriate to ask. ;)
I'd also have a few questions regarding the integration, especially regarding the bidding. In the readme you say that you used a conventional (heuristic?) bidding algorithm, but in the sources there seem to be data files for probabilistic distributions? Are they based on the ISS files (I saw parsers for that)? Did I understand it correctly that the model always plays Diamonds, and that I should shuffle the colours on the engine to let it appear as if it would play diamonds for the other suites?
I'd appreciate any help,
Best Greetings
André

[agemuend]

I think I was able to re-implement what the Python code does and the AI is playing now. Is the Python code, e.g. the EV based bidding with the table lookups and the play card code the latest working version, which was used to achieve the results on the ISS? Just wondering what I should expect. I think I still have some issues, as its not as strong as I expected.

[Jimboom7]

Hi, sorry for the late answer.

Would it be okay if I included your model as an AI opponent called SkatZero?

Yes, that's totally fine. There is no publication, but you can cite this repository if you want. No citation is also fine. If you ever put your game online you can send me a link so I can have a look at it :)

I'd also have a few questions regarding the integration, especially regarding the bidding.

The bidding algorithm is not very complicated. The idea is that you can get the EV of the first discard from the AI. You can then simulate all possible hands which could form from picking up the 2 Skat cards and take the average expected value. If performance is a problem, it's actually not necessary to simulate all possible hand combinations, it converges quite fast to the average expected value. But I'm sure you already figured that out.

...but in the sources there seem to be data files for probabilistic distributions?

Indeed. That's solely for the bidding to map the EV output from the AI to percentages for Schneider or Schwarz. For example to make a decision whether to bid a maximum of 18 with single diamonds, or to go up to 27 (expecting Schneider) you need more information than a simple EV. That's where the probabilistic distributions come into play, so you know for example that an EV of 75 leads to Schneider in 50% of the games.

Did I understand it correctly that the model always plays Diamonds, and that I should shuffle the colours on the engine to let it appear as if it would play diamonds for the other suites?

Yes, that's correct.

the EV based bidding with the table lookups and the play card code the latest working version, which was used to achieve the results on the ISS?

Yes, that's the latest version.
If the AI doesn't perform as well as expected there is probably some bit wrong in your features. I also had to debug that a lot when I connected SkatZero to the EuroSkat server via JavaScript.