New feature: Override player input with machine learning models #17407
Conversation
turn off GAME_AI feature by default
|
Seems like the CI checks failed. |
Fixed |
|
As far as I understand it requires a manually launching separate instance from https://github.com/MatPoliquin/stable-retro-scripts? This may not be very user-friendly. Maybe it makes sense to make it as a dynamic library or run/control the subprocess directly from RA? |
It's already a dynamic lib, only this part of the stable-retro-scripts repo is used with retroarch (ef_lib folder): The python scripts you see are used to train and test the models shipped with the lib |
Perhaps it makes sense to place it in |
Added the ef_lib folder (named it game_ai_lib) in deps |
Just saw your edit about online updater downloading the models... that would be very cool! |
ai/game_ai.cpp
Outdated
| BOOL fFreeResult, fRunTimeLinkSuccess = FALSE; | ||
|
|
||
| hinstLib = LoadLibrary(TEXT("game_ai.dll")); | ||
| assert(hinstLib); |
There was a problem hiding this comment.
retro_assert here and below
This PR adds support for overriding player 1 and/or 2 controller input with custom AI.
For example, NHL94 Sega Genesis community wants to be able to play against better AI, this feature address their needs.
The AI/machine learning part is in an external library so that way there is no need to update RA code base every time there is changes made to a game's AI, updates to pytorch, changes in model format, etc
Users can toggle on/off overriding input with the "Game AI" section in the quick menu.
This is an experimental feature at an early stage and should be turned off by default
Only works on Windows and Linux for now (x86), but eventually more platforms will be supported
Question:
This feature needs an external lib that I made + data (models/ram info). Where is best to put the build instructions for this?
https://github.com/MatPoliquin/stable-retro-scripts/tree/main/ef_lib