ArnoHu wrote:

ArnoHu wrote:

austin_intel001jrahs wrote:

UPDATE #2 — THE FIRST ASCENDANT BASELINE IS COMPLETE

A lot has happened since the challenge started.

The first full improvement cycle is now complete.

After mapping Game of Kings, establishing a controlled baseline, testing individual optimisation ideas and validating the resulting engine, I now have the first version of GoK Ascendant that I am comfortable using as the new development baseline.

Performance result:

In a matched fixed-depth benchmark, the current Ascendant engine completed the same search work in approximately 5.72% less total time than the baseline.

That does NOT automatically mean 5.72% stronger chess.

The important part is that the candidate also passed the bounded correctness tests used during this development cycle.

Engine-v-engine testing:

I also built an automated match system that:

• uses the same Scratch runtime
  
• matches actual wall-clock thinking time
  
• uses paired openings
  
• reverses colours
  
• uses the same opening-book policy
  
• checks moves with an independent chess arbiter
  
• records games and results
  

The long match became much more expensive to run than expected because of Scratch/headless-runtime behaviour, so I eventually stopped it rather than spending several more days testing.

The usable completed sample currently stands at:

GoK Ascendant: 7 wins
Baseline GoK: 4 wins

There was also one unfinished very long game and one baseline timeout/fault.

Important:

I am NOT claiming a +Elo value or statistically proven superiority from 11 completed games.

The result is encouraging, but it is still a relatively small and incomplete sample.

What I can currently claim is:

✓ the first Ascendant optimisation cycle passed its correctness checks
✓ the current engine is measurably faster in the controlled benchmark
✓ limited fair match evidence is currently favourable
~ statistically significant playing-strength superiority is still unproven

Because of that, I have promoted this engine to become the new GoK Ascendant development baseline.

The next stage is NOT release yet.

I'm starting a second improvement cycle to see whether the engine can be pushed further before doing the final strength test.

The challenge continues.

From the time of your original posting I am estimating your baseline GoK NNUE might be around version 7.3383.2318 (you can verify by pressing ‘v’). I played a short 20-games series against that baseline on TurboWarp (1sec), and current GoK NNUE won 12.5 - 7.5. Doesn't matter so much which version it was, what I am saying is the project underwent improvements since, although I will slow it down now and focus on GoK Classic.

Final result 40 - 26. Here is your target.