Tetris AI - Genetic Programming Vs Tetris Game

This video shows the power of a genetic algorithm applied to a simple game like tetris. This AI is able to last for more than 2kk+ pieces, i don't really know its limit because i had to stop testing due to leak of time.

I programmed a simple GUI for the AI that shows the usual tetris interface, it's made in C# and it's no longer than 200 lines of code :) This AI has a 5-values genoma that makes it flexible and adaptive without effort.

The AI was created during just one night of evolution, actually 8 hours of elaboration on ~500 generations of AIs.

If you have any doubts, critics or advices, email me to progv@libero.it .

Thanks

[song from Dj Dom Bro L'Eclipse Night-Club Gold Compilation track 10]

Category:

Science & Technology

Tags:

• Tetris
• AI
• Genetic Programming
• Artificial Intelligence
• Heuristics
• Challenge Bot
• Tetris AI

[R4MIN4T0R]

how do we know its not actually a chinese nerd guy playing? haha, just kiddin, great work man :P

[HeDo88TH]

@R4MIN4T0R LOL

[waddlerobloxxxx]

Hmmm.....is this possible for more complex programs like car simulators??

[HeDo88TH]

@waddlerobloxxxx It would be a lot more difficult, but i think that after a good problem formalization everything can be done with GA :)

[vafles]

did you train yours GAs using a specific static sequence of tetris blocks, or will it alwways perform this good with a random block sequence each time you run it? (without having to train it EVERY TIME)?

If it works with a random sequence then well done. I think GAs and NNs are the future and I fully understand the urge to show to ppl the real POWER of this algorithms!

keep it up dude

[HeDo88TH]

@vafles It was trained with pure random sequences :)

[infoburp]

hi, enjoyed your video, i would be interesting to see how it performs against bastet (barstard tetris) which is available for linux, and deliberately gives the player the most awkward shapes. if it can beat this im very impressed !

[HeDo88TH]

ok, i tested the thinking time, it's about 3 and 3.5 ms.

external ai should have been an overkill :)

[fndTenorio]

Nice work!

I'm doing a tutorial on how to implement this in Java.

Check it out: bitsrandomicos dot blogspot dot com

=)

[Agorsan]

Hello there skynet

[fndmiranda]

Hi, amazing work!

How does your board evaluation function works?

[TheBlueShamen]

Mind open sourcing it? :D

[Blackennew]

excelente!!!

[dewinmoonl]

what's your state? Is it just 2^number_of_cell on the tetris board? Like what is your formalized question? Given a shape that's your board, which next state reachable by an action is most good? I guess you can use more efficient state by only remembering the top siluet of the board? I'm curious ;)

[fstratzero]

I want to see Genetic Programming vs BATTLETOADS