This blog post has a specific purpose: to teach me how to create a Neural Network from scratch.
By from scratch, I mean no libraries, no hardware acceleration, no optimization. Purely logical programming structure and all machine learning functions customized. Often, Neural networks (NNs) are implemented with matrix multiplications performed on the GPU - with libraries taking care of the heavy lifting. Here is the result (more comments below):
While this can be amazingly effective, it abstracts the "black box" idea of a black box NN even further from the user. As a result, I have come up with a class based NN implementation, where nodes are linked together and all processing is done on the CPU in real time.
The NN needs an objective. As you can see, the arbitrary objective of learning to play my version of the "Google chrome no-internet connectivity" game.
For the time being, the leader of each generation survives and the rest are mutated from the winner. You can see the offspring by the preservation of their last name. The winner has a crown indicator in the top right of their game screen. Sometimes, an offspring will be randomly generated into a totally new NN, indicated by a unique last name.
This algorithm is NOT optimized. Frankly, it takes a very long time to generate an individual who can score better than a 5 year old... Also, occasionally an individual who seems to be "good at the game", i.e. they only jump when a tree comes near or maybe hold duck most of the time, will lose out to an individual who holds the jump button.
In summary, there is a lot of room for improvement. But this was an excellent first step into the world of NNs for me. I encourage you to peer into the code or provide suggestions on how to improve the performance.
Be curious!
© 2021 Sky Hoffert