EPD : epd\ProDeo.epd
Time : 250ms
Solving Max Total Time Hash
Engine Score Used Time Found Pos Time Score Rate ms Mb Cpu CCRL
1 time-bonus=on 1710890 06:43:40.3 40733 84617 00:19:22.2 2538510 67.4% 250 128 1 3200
2 time-bonus=off 603229 06:43:40.3 40733 84617 00:19:22.2 846170 71.3% 250 128 1 3200
So, the way I coded things a difference of almost 4% is not noise.
The time bonus is based on the simple observation that Ethereal 12 will find moves much faster than Ethereal 8 because 12 is stronger and thus it should rewarded.
If I understand your table, the result is better without time bonus, so it is doubtful here too.
Maybe the time bonus should only be applied when the best move (from the c0 list) is found, and never on inferior node?
EPD : epd\ProDeo.epd
Time : 250ms
Solving Max Total Time Hash
Engine Score Used Time Found Pos Time Score Rate ms Mb Cpu CCRL
1 time-bonus=on 1710890 06:43:40.3 40733 84617 00:19:22.2 2538510 67.4% 250 128 1 3200
2 time-bonus=off 603229 06:43:40.3 40733 84617 00:19:22.2 846170 71.3% 250 128 1 3200
So, the way I coded things a difference of almost 4% is not noise.
The time bonus is based on the simple observation that Ethereal 12 will find moves much faster than Ethereal 8 because 12 is stronger and thus it should rewarded.
If I understand your table, the result is better without time bonus, so it is doubtful here too.
Maybe the time bonus should only be applied when the best move (from the c0 list) is found, and never on inferior node?
The time bonus is about the red.
One can argue about the height of the bonus currently between 0-20 which makes it a different animal than without.
90% of coding is debugging, the other 10% is writing bugs.
abulmo2 wrote: ↑Thu May 07, 2020 2:11 pm
I just tested on various Amoeba's versions (1.0 to 3.2 in development), and using my own MEA implementation¹ the impact of time bonus. It seems to just add noise to the score.
Let's check if we are talking about the same. Hypothetical example from the start position:
depth-1 bm Nf3 time 5 ms
depth-2 bm Nf3 time 8 ms
depth-3 bm e4 time 15 ms
depth-4 bm e4 time 25 ms
depth-5 bm d4 time 50 ms
depth-6 bm d4 time 100 ms
depth-7 bm d4 time 200 ms
depth-8 bm e4 time 300 ms
depth-9 bm e4 time 400 ms
depth 10 final move e4 600 ms
For applying a time bonus, which milliesecond value do you use? The blue or the red?
The last blue one.
Just for the record, the "bm" can't be trusted, the EPD'S come from various sources, the MEA moves in "c0" are what matters. And secondly, my experience with sfx sets are better than with the lcx sets.
In my experience, bm moves looked better than the MEA scores with time bonus.
I did again the experiment with sfx-1.epd and get similar results.
Note that Amoeba is a strange beast much better at playing games than at solving positions. Maybe the problem is in my program?
I downloaded Amoeba but noticed it has no tuneable UCI parameters, so I can't test it.
A couple of days ago I ran an eval change with Rubichess, 20,000 positions, 250ms and NICE gave a very positive result. When I ran the same eval change with 84,000 positions the result was equal to the default setting. That was a good lesson, from now on I only test with sfx.epd (80,000 positions).
Currently you can not use more than 8 threads to do the job but in the next version (soon) there is support for 64 threads.
90% of coding is debugging, the other 10% is writing bugs.