"MarI / O" which clearly clears Mario in a certain sense only by super-powerful genetic algorithm and neural network learning
Evolutionary algorithmIt is one ofGenetic algorithmWhenArtificial neural network (neural network), A program that tried to clear the donut plain 1 of "Super Mario World" only by machine learning without teaching the solution by humans is "Mar I / O"is.
The program was created by Mr. Seth · Bring and the source code is published by Pastebin.
NEATEvolve.lua - Pastebin.com
http://pastebin.com/ZZmSNaHX
Watch this learning AI smash Super Mario World with ease
http://thenextweb.com/insider/2015/06/14/watch-this-learning-neural-network-annihilate-super-mario-world-with-ease/
In fact, the following movie explains where "MarI / O" is clearing Super Mario World, and how it works.
MarI / O - Machine Learning for Video Games - YouTube
The first minute of the movie's beginning seems to be just like normally solving Super Mario World normally.
However, the player is not a human but a program called "MarI / O".
The remaining time at the time of clear is 354, that is, cleared at 46 seconds.
Go back to the starting point and look at Mario's "in the head", this is the structure. A colorful line connecting these rectangles with them is a "neural network" that expresses the function of human brain function.
The upper left part of the screen expresses this world diagrammatically. Red bars are Mario, white squares are fixed objects, black squares are moving objects (enemies), coins that are floating, etc. are ignored.
On the right side is an input button used by the neural network.
In the center, a line connecting them is formed. This time I will aim for the optimal solution that I can think about in this "brain" for 24 hours.
The top notation is "Gen 0", but this is the 0th generation. Still more training stage.
There are several moves in Mario, such as dash and jump, but the neural network first chose to "run continuously to the right." Even if I pick up a mushroom and become a super mario, it hits the enemy and it shrank back to the original ... ...
I was destroyed without taking advantage of enemy attacks. With this, I learned that "you can not just run running to the right".
I will try various ways of going over and over again and again. The number in the column labeled "Fitness" at the top of the screen indicates how far the stage has progressed.
As each generation gets accumulated more and more experience gets accumulated and "How can I do without getting done, I will advance even a little", so the number of lines will gradually grow.
For example, it advanced to the 16th generation, 1020.
It fell into a hole in the 20 th generation, as it advanced to 1268.
Such a technique is "Neuro Evolution of Argumenting Toplogies (NEAT)It is called.
As the number of generations increases, the more solutions you can find the optimal solution, so even in the ridiculous stage of the "Nintendo World Championship 2015" finals held on June 15, if you raise mountains as much as you can, It should clear up with such unthinkable play.
Related Posts: