Experimental scripts to train an MLP to predict the winning tile from a set of mahjong discards.
simulated_data contains a mini implementation for simulated data that works alright. Currently working on fitting to 2013's data using with data sourced from arcturus.su and DexGroves/tenhou_parse.
- It works on tenhou data
- It beats a GBM
- I need to write a suite to sanity check the predictions it makes against known ponds