Slimmed replay model

[mrbermell]

Currently replays save the complete game state for every step. It should be possible to identify the state's delta of every step using the forward model which would increase speed and reduce size of replays.

• Replay created through forward model
• Auto save games played in Web UI by default
• Undo many actions in the Web UI

[njustesen]

I agree, but then replays only work for that particular version of FFAI.
I think it's still the way to go, but we need to at least mark the replay with a version number.

Saving just the seed and actions is a common way of saving replays that we should consider.

[mrbermell]

Indeed. Maybe we there's a way to make replays use the forward model and still be less susceptible by tweaks in the framework.

Saving the seed and actions are neat for pure replay i.e. forward but it's not possible to step backwards. Also I wish to convert FUMBBL replays to FFAI replays for imitation learning. That makes the seed thing tricky.