@@JohnnyCodes what I think could be even more interesting is making the plants spawn animals and vice versa in alternating generations, while also allowing both sides to evolve
I'd love to see some good eco system also.. a game engine such as worldbox might be suitable for you.. probably got invented for that reason but nerfed for public consumption but I think it's base is open source. Somehow evolving a neural net to 'best guess' the creation of new neural nets.. an running evolutions in simulations within simulations.. must be the way forward. Thanks for the inspiration an good luck.
Did you reuse weights from the previous video(padded with zeroes for new input & output)? Or train from scratch using a nearly identical network(only adding new input & output)? Do the cars get some sort of penalty from jumping(e.g. inability to steer, converting speed to jumping energy, etc)? I wonder if the 'old' raycasts alone could be used to infer the existence of hurdles, if they also intersect with them
So I would need to double check how I did it in the code but either I made the bottom raycast (the one for detecting hurdles) pass through the walls. Or I just let the network figure it out. Both options would work because the network could figure out the relationship between the top raycast that goes over the hurdle and the bottom one. For example if both the top and bottom rays are short then it is probably a wall but if the bottom is short and the top is long then it is probably a hurdle. It will take longer to train for the second option because that is one more thing the network will need to learn but it is still doable