Why Patricia 3.1 more aggressive than 4.0?

[Stallion77RepoOfficial]

Is it all about nnue net? Or are there other things? And also my advice to you is to optimize as HCE because networks will lose their profit to you after a while. You will not get better results anymore.

[Adam-Kulju]

3.1 has a higher EAS score than 4.0 because of it's massive strength increase; against the strong opposition that Pohl tested it against, it manages to draw many games that it would otherwise lose, and is able to grind out more wins in the endgame. This actually lowers the EAS score because a high EAS engine wants to win as fast as possible and draw as little as possible, but I'd rather Patricia win a boring game than draw a boring game.

I'm still not sure as to what entirely happened given my own testing on a different opening book has 4.0 being every bit as aggressive, if not moreso, than 3.1, but that's my best guess.

An HCE heavily optimized for aggression would be more aggressive than any neural network could be, yes, but it would also be hundreds of elo weaker. Patricia would be too weak to stand a chance against any strong engine and would never get to show off its style. I have attempted to add HCE elements in, such as explicit bonuses for attacking the king, but none have worked to increase EAS yet - hopefully soon I'll break through.

[Stallion77RepoOfficial]

Thank you but I didn't get an answer to my first question. I guess you didn't see that part. What I mean is, is all this EAS difference due to NNUE or did you make too many code changes in the engine?

[Adam-Kulju]

I have a different neural network for Patricia 4, but it's trained in the same way and is every bit as aggressive as the old net. The difference is all the changes in the search as far as I can tell.