fieryrms / estr-metasquares Goto Github PK

The AI doesnt seem to prefer having a high weight for killerheuristics like i expected it to. Most probably the implementation is wrong. Or maybe it doesnt make any difference in the training.

None of the AIs in any generation were able to beat all of the AIs. The weights of the top AI change dramatically in a span of 2-3 generations, and often its an AI you wouldn't expect to perform well. Then in the following generation, that AI is suddenly performing very bad. I expected the AIs to eventually settle at a well rounded AI who is able to get draws against stronger AIs.

My suspicions is that the AIs may have developed a couple of strategies, and each strategy can beat one other, but not every strategy, a non-transitive relationship, much like rock paper scissors. So as the number of one strategy increases, another strategy then beat all those AIs and gets a high position. A visual example could be this meme.

seeing the visualization, you can see that breeding method one is performing terribly, need to find a better way to breed. I found out so late that it would not be worth the effort to fix anymore.

expected:
when restart.bool is set to true, the python stops execution, restart.sh pulls origin and executes start.sh to start the program

observed:
the file does pull and restart, but the nohup.out file is not created when start.sh is run, thus I cannot see live logs unless its through google logging.