Pokémon MuZero

The training pipeline is finally done !
now the real challenge can finally start !

A quick rundown of a fun bug I encountered today. The concurrency API doesn’t provide any tools for generating or validating a deck, it’s entirely up to the model to create a valid 60card deck, without any safeguards

I ran into a big problem with ACE SPEC cards: the game rules prohibit having more than one per deck, regardless of the ACE SPEC category. However, the card database is poorly designed, some cards simply weren’t declared as ACE SPEC in the rules column. My code treated them as standard cards and added several to the deck without issue. The C++ game engine, however, wasn’t having any of it: it systematically rejected the deck, and the worker went into a crash loop.

The fix required restarting the detection system from scratch: I now scan all suspicious columns in the CSV and the engine’s API to list absolutely every ACE SPEC in the game. As soon as the Deck Builder draws one, the other 19 are immediately locked for the rest of the draw

So now the error is finally corrected and I can finally focus on the real nightmare : the training of the neural network

fist devlog, for now the pipeline isn’t ready but it will be soon !