AI Cars Learn To Drive! 

Agree with this sentiment. Feels like this might have been more of a situation where they are overtrained on the data.

Yeh, I agree. As long as the network was big enough, it'd probably be somewhat easy to train it on new tracks for it to become more robust

ayo i was 69th view too nice👌

Yeah that would be awesome see if they really learn from theirs mistake

Thank you! Will do! Still got to work on that consistency thing

850th sub over here 🙋🏻‍♂️ so many hundreds of thousands to come

It has been a while but I just started a new series for the neural network tutorial! ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-JXGogdI7RIE.html

@@JohnnyCodes Oh my goodness! Thank you so much!

Glad you enjoyed it!

I wan't sure if people would know it but I put it in there for me. His storytelling skills are insane. I watched so many of his videos without even playing the game

@@JohnnyCodes ah I see. Your attention to storytelling shows

unusual because it is not main stream ? 🤣

No, because it doesn't work for bigger networks.

@@VladimirPanteleev but do you really need monolithic large networks ?

@@dm3on Yes, for many classes of problems.

How do you use backpropagation for a system with no obvious direction of improvement? My current understanding of backpropagation only allows it to be applied to situations where you already know the best answer for each of the questions in a training set, so you know which direction the output needs to move in order to improve the network's performance - e.g. the network said 2, the correct answer was 3, so you need to increase the network's output, so you backpropagate through the system to determine how all the weights and biases need to change in order to make the output closer to three. But with driving, it is not always obvious what affect a change to the weights will have on the performance of the system, since you don't necessarily know beforehand whether driving straight or a small adjustment to the left is optimal in a given situation (or if you do there's probably not much point in training a neural network - just hard-code it). How do you resolve this issue?

Yeah in this video I did make the ray casts go through the checkpoint otherwise they would try to avoid the checkpoint thinking it’s a wall (unless you adjust the inputs so it know the difference between a wall ray and a checkpoint ray)

@@JohnnyCodes okay thanks !