Seth Bling is a game developer who also happens to be a Super Mario World speedrunner. But in the blistering run shown above, there is no human playing. Instead, we’re seeing the end result of a computer program — called MarI/O — Seth built to learn like a human brain.
So-called “genetic algorithms” are basically computer programs that zero-in on the best solution to a problem via a trial and error approach. However, once a random trial does better than the rest, the algorithms use that solution as their fundamental approach, randomly proceeding forward from this better position. Eventually, the algorithms evolve enough that the optimal solution is reached (like the most efficient way of finding Waldo). Goombas and flame flowers act as natural selection.
The other component of Seth’s work is that these trials and errors are handled in-game by simulated neural networks — logic connections that mimic how neurons in a human brain send signals and coordinate.
Bling explains everything beautifully above, so go ahead and watch a computer probably destroy your best run.
—
HT: YouTube