I arrive back at the labs only to find that two of my experiments have crashed, due to NoSuchElementExceptions. The two that crashed were the Reintegration and Regeneration experiments. Thankfully, some progress was made on each. The reintegration experiment had completed 22%, thus creating 2 useful files, and the regeneration had completed 14%, creating 1 useful file.
Each of the failed experiments crashed in the removeNWorst method. I’ll have to sort that out.
As for the other two, they are still running. Each has been running for 91 hours, and are at 23% with 307 hours remaining for the hand-coded rules (due to each episode taking longer due to better performance) and 45% with 108 hours remaining for the random-rules.
This is looking at 2 weeks for the hand-coded rules to complete an more than 4 days for the random rules to complete. I wish I could speed up the process, but I cannot. At least I have some performance files to work with.
Glancing at the results, it looks as though regeneration could be a winning strategy. But, once again, I have no proof that the performance file could just be a lucky break. Anyway, the random rules appear to be peaking at around 40K (which seems lower than my initial results. Perhaps due to the addition of the flashing ghosts rule?). However, regenerated rules seem to push above the plateau.
Actually, the reason these particular results plateau at 40K could be precisely because of the rulebase. However, regeneration allows the rulebase to change – improve, preferably. Hence, the plateau may arrive much later, when the rules have become near-optimal. The rulebase writes itself out at the end of every episode, so when the experiment is complete, I will be able to look at the rules.