Рет қаралды 2,967
This is 30Hz with 300ms reaction time, and is trained with 430-line cap in mind. On double killscreen, the reaction time means it will never make decision based on the next piece, but it can always react to the current piece.
This is about an 1-in-500 seed with 172 hong bars, but well, 3 million is a great number so I just went for it :) I found this seed purely by running BetaTetris on a lot of seeds. The agent didn't know the seed and there was no RNG manipulation.
This is a hybrid agent powered by the new BetaTetris Tablebase and the old-style BetaTetris neural network. For information about the tablebase, please refer to the GitHub page: github.com/adrien1018/betatet...
The bot is designed to minimize the number of adjustment inputs (more specifically, the weighted squared number of adjustment inputs) provided that it can reach its desired placements. So for example if it adjusts from the left all the way to the right, then there must be certain next piece(s) that would make it place it on the left.