This page presents research in Tetris done by undergraduate students Lauren Gillespie and Gabriela Gonzalez as
part of Southwestern University's Summer research program
SCOPE. Note that this research has since been
extended and improved upon.
The research makes use of the MM-NEAT software package
(an extension of NEAT),
which has been extended to include HyperNEAT.
State evaluators are evolved to play the game of Tetris using both raw screen inputs (taken from the 10 by 20 grid)
and hand-designed features (commonly used in previous Tetris research) using both HyperNEAT and standard NEAT.
The videos on this page can be viewed in a playlist here.
NEAT Using Raw Screen Inputs
NEAT using raw screen inputs is unable to learn how to play the game. The agent uses a delaying tactic of moving pieces to the sides of the board before filling the center. This tactic results in a very small number of lines being cleared.
HyperNEAT Using Raw Screen Inputs
HyperNEAT using raw screen inputs plays Tetris much better than regular NEAT. The agent is able to clear many more lines than the NEAT agent, though it does lose before too long. Decent performance using raw inputs is impressive, but still cannot compete with using simple hand-designed features.
HyperNEAT Using Hand-Designed Features
HyperNEAT performs much better with hand-designed features than with raw screen inputs. In fact, the geometric awareness of HyperNEAT allows it to quickly shoot up in performance and do better than standard NEAT (with hand-designed features) in early generations. However, HyperNEAT's performance plateaus early and stays steady, even though there is still room for improvement.
NEAT Using Hand-Designed Features
This particular NEAT champion using hand-designed features performs much better than other HyperNEAT champions, though most NEAT champions perform worse. NEAT runs learn slower than HyperNEAT and tend to be worse, but as this video shows, the best NEAT results are better than the best HyperNEAT results when using hand-designed features.