Turbowarp link: https://turbowarp.org/1391288176?interpolate&fps=60 RANDOMNESS Variable: Start with the randomness slider at a large value, like 90 to 100. As the network learns, decrease it slightly. Once it finishes a lap, a good amount to have it at is 25 to 35. Once the BEST SCORE variable stops increasing every 10 to 15 seconds, it should be around 10. ___________________________________________ Graphing Growth: If you would like to see a graph of its progress, you can go into the project, show the BEST SCORES ARCHIVE list, right click it, and download it to your device. You can then go to: https://scratch.mit.edu/projects/1391696992/ Once in the project above, click SEE INSIDE, show the SCORES list, right click and select IMPORT on it, and load the file your just saved from here. On licking the green flag, the project will graph the growth. ___________________________________________ Info for Fellow NERDS: 5 inputs, each being the distance to the level in a certain direction: 0, 15, -15, 60, and -60 1 hidden layer with 10 neurons 2 outputs: turn right and turn left no backpropagation. just randomly modifies a running best network. |||||||||||||||| Each of the five cars does its own thing. When it dies or finishes a lap, it compares it to the previous best. If it performed better, it replaces the old best with itself. after it compares, it restarts the track, having a slightly randomized version of the weights and biases of the current best network. |||||||||||||||||||||||||| For more, just look in the code. The custom blocks have descriptive names.
Thank you to @iropDa and their Neural Network + project for being a loose model for me to design the game. https://scratch.mit.edu/projects/296745075/ ------------------------------------------------------------------------- MY original neural network: https://scratch.mit.edu/projects/1389926344/
