BikJump V3.1 is now available for download

[abik]

I am of course already dreaming of the next big jump, but I am also running out of ideas...

Any hints from the experienced engine programmers out here?

Without going in the direction of a neural network or even full NNUE (I like my current static evaluation approach), what technique or optimization could bring a 2400 ELO engine into the 2600 ELO range? Is there a single technique that caused a big improvement in other engines? Or is it a slow climb up from here on?

[Aleks Peshkov]

1. NNUE. It should not be complex for the first implementation to get immediate 500 Elo gain. Better evaluation allows more aggressive search selectivity.
2. Search selectivity parameters tuning. Generally means adding more selectivity to search with extra pruning are depth reductions. LMR and NMP are the most strong as they are recursive.
3. Writing bug free code is obvious advice. I regularly discover plenty of small bugs when refactor my old "stable" code.
4. More search and move-ordering heuristics. But quality rules over quantity.

[abik]

Thanks! I may eventually try NNUE also, just out of curiosity. But at the moment I am still using what is now called HCE (hand-crafted evaluation), simply because I enjoy tuning that part of the engine the most. But, in fairness, most of the improvements so far came from pruning, moving ordering, and tuning the parameters, not so much from smarter evaluation.

BikJump always frustratingly hovered around 2000 ELO and only recently did I have the time again to prove there was a strong engine lurking inside that code With the recent success towards 2400 ELO, my next goal is getting close to 2600 with just HCE, but perhaps that is too ambitious. But it has been fun so far for sure!