PZChessBot

• Don't allow NMP in singular search (10 Elo)
• Compile with LTO (15 Elo)
• Add node time management (24 Elo)
• Lower history values each search (8 Elo)
• Fix node counting (5 Elo)
• Do not store best move on TT fail-low (4 Elo)
• Fix checkmate (1 Elo)
• Disallow IIR at expected all-nodes (12 Elo)
• Raise singular extension beta (8 Elo)
• Disallow corrhist update on mate scores (5 Elo)
• Relax SE bounds (5 Elo)

• Adjust soft bound based on search complexity (9 Elo)
• Only do NMP at sufficient depth (6 Elo)
• Prefetch TT values before recursing (7 Elo)
• Reduce killer moves less (16 Elo)
• Increase overall LMR (7 Elo STC, 9 Elo LTC)
• QS Prior Extensions (5 Elo)
• NMP Improving (4 Elo)

I actually might just be the goat, this is the largest green wave we've seen in a while
- Use optimized move picking in qsearch instead of std::stable_sort (which is slow) (4.5 Elo)
- Optimize capture generation for qsearch (6 Elo)
- Optimize NNUE inference with Finny-like tables (15 Elo)
- Only store upper 32 bits of hash (5.7 Elo)
- Fail-firm beta adjustment when failing high (16.6 Elo)
- Simplify Finny NNUE inference (1 Elo)
- Modify time allocation (8.3 Elo)

• Disallow LMR at low depths (7 Elo)
• Disallow LMR for early moves (5 Elo)

• Allow razoring up to depth 8 (+2 Elo)
• Do more NMP if score is far above beta; depth cap RFP (+5 Elo)
• Add nonpawn corrhist (+8 Elo)

• Re-train NNUE again
• Allow adjustments to NMP margin based on current depth (+10 Elo)

• Generate another 600M positions for the neural network
• Cutnode reductions (4.8 +/- 3.5 Elo)
• Dynamic PVS SEE margin (5.2 +/- 3.8 Elo)
• PrevMove CorrHist (6.7 +/- 4.4 Elo)
• TT-corrected QS evals (8.9 +/- 5.2 Elo)

• Create a cool little visualization thing for the NN weights

• Do some retrains of the main evaluation NNUE
• Continue with self-gen, gain another +20 Elo from WDL updates (0.15 -> 0.40)

• Continue tweaking the new method (normalization helped it gain @ LTC but hindered its performance @ STC), now only gaining +3 Elo
• Not sure if I want to merge this into PZ ... but we'll see

Try a completely new pruning method (cut net) that gained +15 Elo (WTF!!!!!)
Might PR this idea into Stockfish and see if it can also gain from this

• Test the new neural network against v3.0 (+16.0 @ STC, -4.7 @ LTC)
• Test the new neural network against generation 12 of the self-trained neural network (+16.5 @ STC, +30.7 @ LTC)

• Completely overhaul the PVSearch mechanism in PZChessBot
• Fix PV handling (for the most part; still some weird edge cases)
• Modify training schedule and prepare to train next generation of NNUE w/ 2.6 billion positions (78GB of data)

Continuing on with general QoL fixes in PZChessBot since the search is pretty much complete :)
- Add MultiPV support
- Test 3-PV vs 1-PV @ fixed depth=10 (+57 Elo)

• Try adding cutnode reductions (to no avail, again)
• Test disallowing NMP in PV nodes
• I think my nodetype handling is kinda cooked

• Add a web-playable version, compile with WASM

• Continue self-generated NNUE
• Change to using 5 buckets and also shuffle data (+30 Elo)

• Fix the TT issue that has been cooked for a while (+11 Elo) Basically, I was storing everything into the TTable as an UPPER_BOUND (which is obviously wrong), but fixing this bug would always lose Elo. The problem was that I was doing TT cutoffs in PV nodes, which is expressly disallowed; but only recently discovered this problem...

• Test out a bunch of tweaks, including:
• Lowering RFP threshold to 100 (+2.8 Elo)
• Center control evaluation term (+10.2 Elo)
• King safety evaluation term (+11.3 Elo) Around 2500 Elo now for the HCE version of the engine!

• Also use PeSTO values for material, and switch the tapering system to use PeSTO's system (+148 Elo)

• Tweak the eval function to prioritize rooks on 7th (+15 Elo)
• Use PeSTO PSQTs (+120 Elo)

• Add bishop pair bonus (+9 Elo)
• Taper the PSQTs for endgames for pawns and kings (+79 Elo)
• Test adding mobility bonus but my code is too slow for it to gain Elo... maybe in the future

• Begin working on the hand-crafted evaluation part of the engine
• Restarting from scratch (with just material counting), added PSQT (+270 Elo)

• Touch up the data generation script to skip heavily biased starting positions
• Generate another 500M chess positions for training the network!
• Also did a small change where instead of only deepening the search for the hash move, singular extensions actually raised the depth of the entire node (+3 Elo)

Added pretty console output! It looks really great IMO

• Add QSearch Futility Pruning (+7 Elo)
• Reduce less on PV nodes (+5 Elo)
• Reduce more when the TT Move is a capture (+6 Elo)
• Fix killer move logic (+18 Elo)
• Train a new horizontally mirrored network (+28 Elo)
• Add PVS SEE pruning (+24 Elo)

• Rewrite move ordering scheme (+5 Elo)
• Implement LMP (+8 Elo)
• Implement singular extensions (+5 Elo)
• Implement negative extensions (+13 Elo)
• Implement history pruning (+21 Elo)
• Do more RFP when improving (+18 Elo)

• Run an SPSA (+5 Elo)
• Do not reset Transposition Table on position command (+65 Elo!!!!!)

• Only do NMP when eval is above beta (+12 Elo)
• Only call TT once per node (+12 Elo)
• Add razoring (+31 Elo)

• Used TT eval instead of static eval when available (+10 ELO)
• Use lazy sorting (+2 ELO)

• Test various tweaks, including:
• always replace scheme for TTable
• lazy sorting in main
• doing NWS first in PVS, even after score > alpha condition is met
• disallowing RFP in mating positions (for more accurate mating scores)

none of these passed :(

Continuing rewrite:
- Implemented check extensions (50 ELO)
- Implemented proper time management (30 ELO)

Testing:
- Futility Pruning
- QSearch Futility Pruning
- Capture History Heuristic

• Continuing rewrite
• Implemented quiescence, transposition tables, aspiration windows, PVSearch, RFP, NMP, SEE, and FP
• We're already back to 3200 ELO!

To answer @nosrep, I'm rewriting the search because I wrote the bulk of the foundation for the search back when I didn't really know much about chess engines. As a result, my search code is, honestly, extremely messy and difficult to deal with. So, I'm trying to clean it up (along with fixing a few bugs and adding some small microoptimizations)! Also, when I wrote the base search, I didn't know how to properly test code either, so I just made changes and hoped they made the engine stronger. Now, by testing all my changes, I can be sure that my engine is actually stronger, and also I can document my progress and write blog posts on it!

• Begin working on rewrite of search
• Implemented a/b pruning, basic move ordering (MVV-LVA, Killer Moves, HH)
• Also documenting progress on int0x80.ca (other project)

• Add static exchange evaluation in quiescence search (~20 ELO) ~3300 ELO now

• Messed around with trying to add Horizontal Mirroring to optimize my NNUE inference but didn't work (will be revisited soon)
• Flesh out data generation workload, tool is fully working now (~3.5kpos/s)
• Test adding Capture History to QSearch
• Test using a model trained on 0.05WDL instead of 0.15WDL

Start working on self-gen data for self-training network
Tried a variety of small improvements that didn't work :(

• Finally got CorrHist working (+7 ELO)
• Still waiting on other tests that will take eons

• Figured out why my pawn corrhist wasn't working
• Waiting on SPRT results (again)
• Broke the bank to rent a 48CPU Google server to speed up tests
• Please help my sanity

• Various small speed optimizations (~4 ELO)
• Try adding Pawn CorrHist (but it's failing miserably) We're like 3250 ELO right now!

• Completely overhaul the NNUE evaluation system
• Change network to use king input buckets (~54 ELO) Insane change for an insane amount of ELO

• Add extended futility pruning along with another SPSA tune (~11 ELO)
• Testing adding internal iterative deepening, seems to not be doing too well (-0.95 +/- 4.3 ELO)
• Testing increasing the NMP R-value from 4 to 5, awaiting results right now

• Change history heuristic to use history gravity formula (~16 ELO)
• Add capture history heuristic to make MVV-LVA -> MVV+CaptHist (~14 ELO)
• Allow RFP to be performed at all depths (~5 ELO) Engine is now around 3160 ELO, putting it around 140th in global standings

• Added futility pruning (~20 ELO)
• Vectorize nnue_eval() (~7 ELO)
• Run a full SPSA tune of values and use the tuned values (~20 ELO) Didn't have time to run a progtest but it's looking good!

• Improved aspiration window logic to gradually widen instead of defaulting to a full-window search (~20 ELO)
• Improved time management to include a soft bound (~20 ELO) Engine should be around 3120-ish ELO now, putting it around 160th in global rankings (further testing needed)

Finalized some stuff for the v2.0 release! Not too sure if I will continue work (since most of it will be with NNUE stuff and that's really :( and also because engine dev is getting really dry)