I was getting frustrated by long test times, so I did something about it. Previous data showed initialization took upwards of 80% of the total testing time, so that’s where I focused my effort. The results:
What took 2.2 seconds per test before now takes 1.4 seconds! Each test is 128 game playthroughs, which is enough to create visible trends in win percentage. This multiplied by eight different tests is a time savings of six seconds per total test run. While this may not seem like much (and maybe it isn’t), waiting 17 seconds is a lot different than waiting 11 seconds.
Now to resume figuring out how much to cost this tech! Clearly 200 Wood is too much, but what if I tweak the AI…
Update: I recreated the test from the last blog post by running 512 games per test. Compare the two times below.