I figure that it is time for an update. Zidanax improved Fillmore Act 1 by 20 frames, and we tried hex editing the simulation in. But that won't be possible because the monsters' AI script is RNG dependent. So I have to redo the simulation segments, which is okay since I can now look for improvements. And of course it is always easier the second time around.
I took another look at the pointer movement in cycle 1 to see if there is a better way to avoid building the houses that catch on fire. I determined that the current path is the fastest. In doing that, I determined more information about how pointer movements affect where the people build. It is not frame dependent, which is good for doing test runs. It doesn't matter how many direct the people segments are used. The only pointer movement that matters are the ones that create roads in that cycle, though I haven't checked if the way the roads are built have any effect; any other pointer movement does not matter. This means that square wobbling, backtrack wobbling, and repeating the same sequence have no effect, which simplifies figuring out the optimal path. The pointer position at the end of the cycle is also important. And, of course, where people built previously would have some effect on future cycles. With this knowledge, I will reanalyze every cycle to hopefully save some frames.
While trying to figure out how to get the hex edit to work, I used DarkKobold's lua script to watch the position of the blue dragon that kept moving. I then decided to enhance the script (and fix some bugs) by changing it to a heads up display, as well as adding other important information so that I don't have to use the memory watch. The script turned out really well, and I am proud of it being that it is my first lua script.
http://lua.pastey.net/129238
EDIT: I finally started on the TAS, and I realized that I needed to add one more thing to the heads up display. The lua script link has been updated.