Genetic Algorithm for Minecraft TASing?

I used a genetic algorithm together with a physics simulator to generate a TAS for walking on a path. I show two outputs: one ran over 10000 generations, and the other ran over 100000 generations. The initial population was given random yaw from -180 to 180 degrees for each tick, then evolved to maximize the Z position within 100 ticks.

I find that the run with more generations produced slightly better results, winning by 0.02973 blocks. I also superficially observe that the latter run contained less jittery movement, suggesting more optimized output. Both runs transition between left- and right-facing angles smoothly, which is distinct from a run that greedily cuts corners. This makes sense as a greedy approach would carry extra velocity going over a corner, causing the path to inflate and is therefore not optimal.

While this suggests that GA can be used to TAS Minecraft, this is a very limited case. Even for this simple task, some tunings were needed for it to produce an optimized result. Therefore, GA would not be appropriate to replace input generation for tasks any more complicated. More testing is needed to see its viability in the optimization of manually generated solutions.

Code: https://github.com/Naruyoko/minecraft-autotas/blob/2a73535c1fec38aa23b917d7130871ab3bbd3cae/src/minecraft_autotas/test/GeneticAlgorithmTest1.java