I’m trying to create a game where AIs learn to do simple Roblox obbies based off some easy gameplay I’ve already created. However, the AIs struggle greatly in learning beyond a certain point. They are able to figure out which direction to walk from the start, and eventually to jump constantly, but they are unable to adapt to their situation for the most part. This is a rough map of their path by the end:
To make matters worse, I once left my computer running for an hour or so while the AIs were training using a checkpoint-based method. They had completed the course, though their skill had waned after leaving it running for at least fifty more generations. I attempted to replicate it twice, however, that was the only time they successfully made it through the entire course, by sheer luck.
The area on the other side of the wall is just the start area (which is a square), so it’s most likely not the cause. I’m using a natural selection algorithm to pick 3 out of 25 AIs who have 4 hidden layers of 32 nodes each. Are there any modifications I can make to this system to improve their training and fix this issue?