Project Success

Success

So, the goal of the project was to reproduce a0 and beat stockfish. Specifically SF8 and at similar search sizes as a0 and sf8.

I chose to simulate this by fixing the nodes per move at multiples of 80k (with smart pruning off) for Leela and 70m for SF8. I used net 11077, because it was the most recent when I started. I used no openings or egtb.

1. The 1 second simulation (80k vs 70m) had a record of +0-8=2. Not so good. study
2. The 2 second simulation (160k vs 140m) had a record of +1-0=10 (I had a disk unmount problem, so had to restart and got an extra game). study
3. The 4 second simulation (320k vs 280m) had a record of +3-0=7. study
4. The 8 second simulation (640k vs 560m) by @hanse had a record of +3-1=6. study

I think we're there. I think if we ran the sf8 vs leela match on 60s/move a0 equivalent conditions, we would win by similar margins as a0. If anyone has some monster hardware, you're welcome to run longer simulations

As of this time, I consider this project a success.

Congratulations to team and contributors. On to bigger things.