This is a little project i worked on about a year ago. I made a post on the other forum about it.
This would be my second attempt.
I am running a program called MAR I/O designed to make neural networks and genetic algorithms and is attempting to learn how to play Super Mario bros. I am broadcasting this attempt LIVE
http://www.twitch.tv/cvripley
The left box is the input of the program. This is what the program actually sees. The black boxes are enemies and the white boxes are blocks it can stand on.
The red and green lines are the neural network. This will become extremely complicated as the program learns the game.
Each free floating box are like a neuron and the lines are similar to the axons and dendrites in the human brain
Green lines are positive connections and the red are negative connections.
Its progress is determined by the fitness number at the top which is a formula based on how close to the end of the level he gets in the least amount of time.
The program will start by making some different genomes, then species then generations by breeding the most successful attempts and then adding some random mutations.
As computers are not as fast as the human brain this is a very slow process. The game is speed up hundreds of times faster even then this will take days or weeks.
It will reset itself each time it fails. (If it dies or no progress is made after a certain amount of time normally due to lack of input)
I have seen it beat up to the second level. (YES It will EVENTUALLY learn to jump over the first goomba)
This broadcast will run until it beats the game or until I get bored whatever comes first.
I am dedicating a computer to this process. (its taking 30% of my computer resources believe it or not)
So check it out ever so often and see its progress.
The concepts of this program are based on NEAT or
Neural
Evolution of
Augmenting
Topologies
http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf
This would be my second attempt.
I am running a program called MAR I/O designed to make neural networks and genetic algorithms and is attempting to learn how to play Super Mario bros. I am broadcasting this attempt LIVE
http://www.twitch.tv/cvripley
The left box is the input of the program. This is what the program actually sees. The black boxes are enemies and the white boxes are blocks it can stand on.
The red and green lines are the neural network. This will become extremely complicated as the program learns the game.
Each free floating box are like a neuron and the lines are similar to the axons and dendrites in the human brain
Green lines are positive connections and the red are negative connections.
Its progress is determined by the fitness number at the top which is a formula based on how close to the end of the level he gets in the least amount of time.
The program will start by making some different genomes, then species then generations by breeding the most successful attempts and then adding some random mutations.
As computers are not as fast as the human brain this is a very slow process. The game is speed up hundreds of times faster even then this will take days or weeks.
It will reset itself each time it fails. (If it dies or no progress is made after a certain amount of time normally due to lack of input)
I have seen it beat up to the second level. (YES It will EVENTUALLY learn to jump over the first goomba)
This broadcast will run until it beats the game or until I get bored whatever comes first.
I am dedicating a computer to this process. (its taking 30% of my computer resources believe it or not)
So check it out ever so often and see its progress.
The concepts of this program are based on NEAT or
Neural
Evolution of
Augmenting
Topologies
http://nn.cs.utexas.edu/downloads/papers/stanley.ec02.pdf
