Confusion on Observation Bit-Strings

[jdpena]

Hi,

I have recently been reading multiple research papers and projects that use the HLE (including the original HLE paper) but can't seem to find any documentation on what the bit-string representation of the observation space is. I've contacted a few of the authors of the aforementioned projects and they too don't seem to know what the bit-string observation space is - which is crazy to me since the HLE has played an important role in recent RL advances.

The closest to documentation I have found is in PettingZoo (https://www.pettingzoo.ml/classic/hanabi), but even their documentation is wrong when you inspect the observations. For example, there are times when the Fireworks indices have 1s in them at the start of the game. I've looked through the codebase myself and tried to decipher the CanonicalObservationEncoder in canonical_encoders.cc and the HanabiObservation in hanabi_observation.cc but keep getting different results each test I run.

Can you explain the bit-string representation of the observation space or point me to where I can find one?

Thanks