In 1.9.9

[BDonnot]

Currently beginning of implementation in the "more_than_2" branch:

• ImpossibleTopology in the environment + doc + test
• get rid of the dict in the self.env.action_space
• pandapowerBackend with non contiguous bus index
• fix the last topo registered in simulate (or at least make sure it's correct)
• opponent_attack_line in observation
• opponent_attack_sub in observation
• opponent_attack_duration in obs
• numpy array with type in type hints (probably requires numpy >= 1.21)
• refacto the GymEnv to have private _action_space and _observation_space and properties to set them with the "converters". Ideally when we called gym_env.action_space it should be gymnasium Discrete and not a Converter.
• implement the MultiDiscreteActSpace for the "zonal" control of the grid (one dimension for set_bus for area for example)
• implement the __bool__ (or something) for the action that says if this action is do nothing or not and could be used in if act: ...
• in the doc for "create an environment" explain "how to create some" when you get only a single grid snapshot (vary genp, load_p according to a common factor)
• do a "developer" documentation, starting by the "workflow" when a grid2op environment is created and then proceed with the doc of all major module (this is huge work)
• related to previous, we can also directly do a class that is able to generate that and make it work with grid2op out of the box: TimeSeriesFromOneSnapshot, parametrized by the "daily load pattern", "weekly load pattern" and "yearly load pattern" for example. The "reset" would sample a week, then the daily load pattern would be applied to generate load_p, load_q and gen_p and tadaaaa
• refactorize the documentation of the "modeled element" to have subfolder maybe (see the elements modeled in pandapower for example)
• add the case for handlers when things (eg rewards) need to access the "future", today some stuff check that if insinstance(MultiFolder) and GridStateFromFile but handlers works too !)
• make a "reward" class that looks into the future and is able to compute the load not seen after a game over or something like that. This would be closer to the scores of l2rpn competitions
• make a "reward" that would use an agent and tag some steps as being "hard" and computes (at the end) the number of "hard times" the agent manage to handle
• developer docs for the "time series handlers" (for now each class is not well documented)
• test backward compat with lightsim2grid