So, I had this idea from a heartfelt meme that I saw online many years ago about someone who saved their XBOX because it had a save file from their dead brother. The save file was a replay of their last run around a particular track in the racing game. I was really moved by the meme and wanted to create a game that sort of "immortalizes" you like that. When I learned about machine learning, I identified it as a way to make that happen.
To make the current game, I started with a car driving around a track. The input data for the neural network was the distance from the central spline, tangent angles, and the preceding and the following curvature of the spline. It worked well but I wanted to add different gameplay than just driving around a track.
In the current iteration, the players are chasing a rabbit around the park and the input data is much simpler than in the first iteration - just the distance and angles. I'm only collecting 15000 data points right now, 5 minutes of game time at Unity's 50 physics per second.
I originally wanted the player to be able to play alongside multiple other data models that may choose to do different activities in the park. So, there was a track, a rabbit, and a butterfly - for each target, I collected different data points. But I wasn't able to make a network that accurately determined what they originally chose to do - i.e. which target they preferred - drive around the track or chase the rabbit or butterfly. The goal was to truly capture what a player preferred to do in the game space rather than just how effectively they were able to chase one goal. I'm still working on that.
Comments
So, I had this idea from a heartfelt meme that I saw online many years ago about someone who saved their XBOX because it had a save file from their dead brother. The save file was a replay of their last run around a particular track in the racing game. I was really moved by the meme and wanted to create a game that sort of "immortalizes" you like that. When I learned about machine learning, I identified it as a way to make that happen. To make the current game, I started with a car driving around a track. The input data for the neural network was the distance from the central spline, tangent angles, and the preceding and the following curvature of the spline. It worked well but I wanted to add different gameplay than just driving around a track. In the current iteration, the players are chasing a rabbit around the park and the input data is much simpler than in the first iteration - just the distance and angles. I'm only collecting 15000 data points right now, 5 minutes of game time at Unity's 50 physics per second. I originally wanted the player to be able to play alongside multiple other data models that may choose to do different activities in the park. So, there was a track, a rabbit, and a butterfly - for each target, I collected different data points. But I wasn't able to make a network that accurately determined what they originally chose to do - i.e. which target they preferred - drive around the track or chase the rabbit or butterfly. The goal was to truly capture what a player preferred to do in the game space rather than just how effectively they were able to chase one goal. I'm still working on that.