See topic title and just watch the video!



I find this development very, very interesting (and as with everything AI based, also a little scary potentially) 
It says that the AI is learning to play SMK based on the inputs that the gamer has provided (so it's mimicking + trial and error). I strongly wonder what would happen if we fed inputs of a top player into the initial states...

The AI and video are created by a famous name in gamer country, Seth Bling (https://en.wikipedia.org/wiki/SethBling). He also developed the credits warp in Super Mario World and actually build a working Atari computer within Minecraft!  

That's very true, depending on how well it is configured and how many iterations are done. Potentially this method will deliver the absolute perfect TASes and new techniques that humans may not even be able to execute. But a lot will depend on how the configuration is done though. It's probably very difficult to get it right. Arriving at complex techniques by randomly (unguided) trying out button sequences is something that WILL take a very long time (same problem with the evolution of complex life forms!). The result it gets should then be presented back to the program as a sort of natural selection principle. Therefore, the initial inputs may not be that trivial after all if you want to greatly reduce the number of iterations needed (intelligent design  ).

I find the parallel with the evolution of life very interesting. Evolution by natural selection (which is effectively the neural networks approach) takes an extravagant amount of iterations, whereas guided evolution (humans editing genomes with purpose and direction) is potentially very fast, but may not produce the most optimal results. The combo of the two is most likely the most powerful / efficient.

Creating AI to surpass TAS should be the ultimate goal!

God help us!!!