AI crashes when calling chankan

[mahjongrepository]

AI crashes when calling chankan

[mthrok]

@Nihisil I already downloaded all the logs of tenhou table from 2009 to around 2017 June. If you want to I can share them with you. (It's in AWS s3 storage). I made a simple search engine using these logs for fun https://t.co/05bhyhmixi .

I am thinking to apply Deep Q Learning. So I suggest that we refactor the bot code in the way that different AI functions are pluggable.

A simple approach at top-level is to create a dedicated class which abstracts away the communication to tenhou.net and translate tenhou.net-specific protocol into simpler game state expression, like an array of integers.

tenhou.net <-> Communicator  <-> AI

This way, it is easier to extend the code to run with log files. We can create a new Communicator which actually reads a log file and reproduce the game with the same interface.

[Nihisil]

Issue was more complex than I thought :) I Will try to fix it in next 2-3 days

Thanks for the report

[Nihisil]

Thanks again for the report :)

Issue was fixes in the 0.2.8 release: https://github.com/MahjongRepository/tenhou-python-bot/releases/tag/v0.2.8

My next plans to this bot (maybe you are interested in them):

1. Refactoring. Really heavy refactoring. I have a lot of places to improve in terms of code base
2. Profiling. I want make this bot more faster and make it consume less resources
3. Improve play strategy (It have a lot of places to improve)
4. Add AI based on the machine learning functions (from phoenix replays)

[0xrgb]

Thanks for fix :)

ps. I'm interested in this project, because I'm learning about AI(machine learning) at school.

[mthrok]

Hi @Nihisil
I am also interested in your next plans, as I am also thinking of building AI.
If you have a concrete idea of refactoring, maybe I can help that too.

[Nihisil]

@mthrok @0xrgb Thank you :)

Regarding machine learning AI. I started to collect all phoenix logs with this repo: https://github.com/MahjongRepository/phoenix-logs

And my next step is learning about machine learning, because I don't have experience with this field of knowledge. I know that famous bot ⓝ爆打 was build with Monte Carlo method, so I will start from here.

If you have something suggest to read about methods that can be used for bot, I would like to hear your comments.

[Nihisil]

@mthrok

@Nihisil I already downloaded all the logs of tenhou table from 2009 to around 2017 June. If you want to I can share them with you. (It's in AWS s3 storage). I made a simple search engine using these logs for fun https://t.co/05bhyhmixi .

Thank you, but I almost finish my downloading, so there is no need to share logs.

So I suggest that we refactor the bot code in the way that different AI functions are pluggable.

Yes. I plan to do refactoring for the next a couple of month and my main goal is build pluggable AI. So, anyone can try to implement their own AI and easy run it on the tenhou.net.

Meantime you can try to implement your AI and run it with other copies of your bot with this mahjong-server: https://github.com/mahjong-server/mahjong-server (I didn't try it yet)