I do most of my test at fast time controls. Sometimes game in 10 seconds or 5 seconds plus a small increment. This enables me to evaluate and tune changes using oodles of games.
Based on these super-fast test Maverick 0.51 rating seems to be about 2375 ELO on the CCRL scale. For example, I pitted Maverick against TJ-Chess rated 2367 ELO at 10 seconds per game. After 1000 games Maverick wins by 518 wins 347 loses and 1356 draws. This is +60 ELO. Maverick seems to get similar results against other engines close to TJ Chess’ strength e.g. OliThink and Dragon 4.6.
So when Graham Banks tested Maverick in Division 6 I thought it would do quite well. I was wrong! Maverick ended up in 11th place out of twelve participants (TJ Chess came in 4th):
I thought I’d investigate and run a match at slower time controls. I used the same time controls as the Division 6 tournament (40 moves in 25 minutes) on my a 2.8 GHz i7. The results were not what I expected (or hoped for). Maverick lost with 20 wins 40 draws and 40 loses! This results shows TJ Chess is about 100 ELO better at slower time controls. This is a swing of 160 ELO between slow and fast time controls – far more than I thought.
As a result I’ve revised my estimate of Maverick’s strength based on the time controls. At super fast time controls it’s about 2400 ELO while at longer time controls it’s 2250 ELO!
Why the Difference?
I suspected the branching factor in Maverick is quite high. I ran some tests and indeed it seems to be about +4.5. This is high by today’s standard. I think Stockfish is about +2.0. This means every ply takes 4.5 longer to complete than the previous ply. At super-fast time controls Maverick does quite well it’s a relatively fast searcher (4.2 million nps on my i7). As the time controls get longer the branching factor takes its toll and the more selective engines out-search Maverick.
The high branching factor is almost certainly due to lack of selectivity in Maverick’s search. It does use Null Move but it doesn’t use any form of Late-Move-Reduction, which seems to be the cost common, and most effective, form of selectivity. This is by design. I have a hunch that if I can create a reasonably good evaluation function, I’ll be able to add a more effective form of LMR selectivity, guided by the evaluation function. My aim is to take Maverick up to 2500 ELO before I add this selectivity. It looks like I have 200 ELO to go!