zhengwei143 / cs3243-project Goto Github PK

Reference: Tetris: A Heuristic Study

For 5 heuristics parallel 5 11 row jobs

Genetic algorithm runs very slowly, could use multithreading to speed it up.
ref: http://www.michaelnoven.se/reports/tetris.pdf

• Each candidate(vector) should be run many times and either 1) take the average score or 2) take the best of those runs. Currently each candidate only uses the fitness from a single run resulting in highly variable results. This process can be parallelized

Print out each generation's weights and fitness into .txt files for documentation

After optimal heuristic weights are obtained, they need to be tested for their average, min and max fitness for report. Should be started at least 1-2 days before report submission.

• Can be done by running ~100(?) plays in parallel

Current heuristics may not be enough to achieve global maximum.
Other promising heuristics:

• Blockades

Run genetic learner on depth-2 search either using SoC computing clusters or AWS servers to obtain new optimal constants for heuristics

Discussion:

1. Include or not to include
2. If include, penalise linearly (current implementation) or polynomially/exponentially?

• Refine heuristics section especially pictures and captions (well sum, bumpiness, aggregate height) and add more formulas (bumpiness, aggregate height and blockades). ref: https://codemyroad.wordpress.com/2013/04/14/tetris-ai-the-near-perfect-player/
• Speedup section in parallelization
• Results section (include num generations ran, runtime on cluster)
• Conclusion (4heu vs 5heu, para5 vs para10, 11 rows vs 21 rows (hypothesis: 21 allows itself to be more lax, relying on luck to fix itself at high heights), maybe mention about how the quality of init pop matters a lot?)