Posted May 15, 2023 by sulsadood
#genetic algorithm #fitness #ai #machine learning #neural network
Wow, I was watching one of Jabrils old videos on machine learning and implemented one of his solutions. Before the AI getting better relied on random mutations, but now the next generation are the children of the last generation. The parents that are merged to form a child are randomly picked, with a bias towards higher fitness levels. This one change took the max fitness value I recorded, which was ~53.5, to over 400. This change eliminated my problems with consistency and allowed the AIs to make consistent progress over a long period of time. I also made it so the 2 worst AIs of each generation are just randomized again for diversity and the children have a chance to mutate to a degree as well.