Presenting A01-Race in 23.80. After ~50-100k brute-force iterations I was getting pretty sceptical that this time would be achieved with the base run of trabadia's WR 23.83 run. There was simply no time to gain at the end and the search simply didn't yield any results. Recently however, I talked with Adrien how it could be possible to "split" the map into parts and optimize these separately.
I then implemented a trigger system into the script that allows to define custom triggers which act very much like fake checkpoints that the script treats like targets to optimize for. These triggers are useful because they can also act like a fake finish line and can be evaluated much precisely using the position of the car instead. This run used one trigger in the middle of the map, just after the first left turn at ~13s. I ran the script repeatedly to optimize for this trigger which is why you see a pretty big difference already after the first downhill compared to trabadia's run.
After optimizing the first part of the map, I then moved on to fixing the remaining part for the proper line and so that it could finish the map. This fixing resulted in a time of 23.82, which was behind 81, but I knew it only lost time at the end. Surely enough, after brute-forcing the fixed 82 run, the script came up with a time of 23.80 after just 1k iterations.
I'm pretty sure that .7x is possible, but it will require some fiddling again to achieve it.
2 окт 2024