What’s interesting, network is even simpler than previously. I created another simple map using new tile: Green tile - It’s possible to jump on it but it’s not to stand on it. If network performs so great let introduce another tile. Training progress(fitness is percent of level completeness): After 115 generations with 150 specimens in each looks like that(still nothing amazing): Training took couple of seconds on quad-core Xeon. So, network needs 49 input neuron for each tile type(black and red for now) and neural network on start looks like this(I know, nothing impressive) Neural network as input takes 7x7 tiles in front of player(picture below) and output is only one, jump or not. To create AI I used NEAT algorithm using sharpneat library. You can stand on black tiles, red tiles kill. At first, I created simple geometry dash clone and following map. Geometry Dash is platformer game that I really like, but I suck at it, so I will try to create AI to play it for me.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |