Posts : 2609 Join date : 2020-11-17 Location : Netherlands
Subject: Release Rebel 15x2 Tue May 17, 2022 12:48 pm
REBEL 15x2 Double sized neural network
x2 stands the multiplication of the neural net with a factor of 2 from 22Mb to 44Mb. Rebel 15x2 is not meant as a new elo bomb but as a final touch of Rebel 15 and investment for a future Rebel.
Because of the double sized network (based on 2.75 billion positions) Rebel has much more chess knowledge available which can be easily demonstrated by playing a match on depth instead of regular time control. The match ended in favor for the x2 version with a 58% score indicating a 56 elo increase.
However there is an important drawback, because of the bigger neural network the NPS has dropped significantly, depending on your PC with 25% - 35% which has to be compensated by the better chess knowledge hence the elo gain is not much. On the other hand it makes Rebel 15x2 much more suitable as an analysis engine and an excellent starting point for creating an even better neural network.
I started the first test with Rebel 15x2 tonight, more results tomorrow evening.... It is clear to me that the larger network has to make up for the speed disadvantage regarding Rebel 15.
Rebel 15x2NN x64 1CPU
Pedone 3.1NN x64 1CPU 3335 43,0 3286 [J.B.] (+14 Elo compared with 15.0)
running...
Admin, Mclane, TheSelfImprover, adminx, matejst and Damir Desevac like this post
Damir Desevac
Posts : 330 Join date : 2020-11-27 Age : 43 Location : Denmark
an unofficial test on my very old computer (built in 2014, but with avx2) TC 30s + 2s, all 2 threads, which started at the same time as the official test for CEGT. Again, Rebel 15x2 is ahead of Rebel 15 (about 28 Elo, head calculation due to score difference). According to my tests so far, the larger net scores very balanced against opponents of very different playing strength, no matter on which hardware, which is a good sign. Keep it up... !
Rank
Engine
Rating
Score
%
Dr
Ko
Ro
Se
Ko
Re
Lc
Wa
Fr
Re
Ko
Za
Lc
De
S-B
01
Dragon 3.0
2200
326.0/390
83.5
· ·· ·· ·
3-0-27
11-0-19
13-0-17
20-0-10
24-0-6
18-0-12
22-1-7
26-0-4
26-0-4
24-0-6
26-0-4
24-0-6
26-0-4
58351,00
02
Koivisto 8.0
2200
280.5/390
71.9
0-3-27
· ·· ·· ·
6-2-22
7-1-22
11-1-18
16-1-13
15-0-15
18-0-12
15-0-15
20-1-9
14-1-15
20-0-10
17-0-13
22-0-8
50501,50
03
RofChade 3.0
2200
263.5/390
67.5
0-11-19
2-6-22
· ·· ·· ·
4-3-23
15-4-11
15-2-13
15-0-15
16-2-12
16-0-14
15-1-14
19-0-11
16-1-13
14-1-15
21-0-9
46763,00
04
Seer 2.5.0
2200
243.0/390
62.3
0-13-17
1-7-22
3-4-23
· ·· ·· ·
9-4-17
7-2-21
13-1-16
15-1-14
12-0-18
13-0-17
16-3-11
15-1-14
11-0-19
17-0-13
43293,25
05
Komodo 13.02
2200
204.0/390
52.3
0-20-10
1-11-18
4-15-11
4-9-17
· ·· ·· ·
7-3-20
10-3-17
8-3-19
10-3-17
8-5-17
8-2-20
10-1-19
11-4-15
17-1-12
35996,00
06
Rebel 15x2
2200
177.0/390
45.3
0-24-6
1-16-13
2-15-13
2-7-21
3-7-20
· ·· ·· ·
6-4-20
5-6-19
4-5-21
5-0-25
10-8-12
12-4-14
9-3-18
7-3-20
31399,50
07
Lc0 791814
2200
168.5/390
43.2
0-18-12
0-15-15
0-15-15
1-13-16
3-10-17
4-6-20
· ·· ·· ·
7-8-15
8-5-17
5-7-18
10-6-14
7-10-13
7-0-23
11-3-16
30161,75
08
Wasp 5.50
2200
166.0/390
42.5
1-22-7
0-18-12
2-16-12
1-15-14
3-8-19
6-5-19
8-7-15
· ·· ·· ·
6-5-19
6-4-20
2-6-22
6-5-19
10-3-17
10-5-15
29466,00
09
Fritz 18NN
2200
164.5/390
42.1
0-26-4
0-15-15
0-16-14
0-12-18
3-10-17
5-4-21
5-8-17
5-6-19
· ·· ·· ·
8-7-15
11-6-13
5-3-22
8-11-11
17-4-9
28821,50
10
Rebel 15.0
2200
161.5/390
41.4
0-26-4
1-20-9
1-15-14
0-13-17
5-8-17
0-5-25
7-5-18
4-6-20
7-8-15
· ·· ·· ·
7-6-17
4-4-22
10-3-17
13-7-10
28339,25
11
Komodo 13.02 (MCTS)
2200
154.5/390
39.6
0-24-6
1-14-15
0-19-11
3-16-11
2-8-20
8-10-12
6-10-14
6-2-22
6-11-13
6-7-17
· ·· ·· ·
8-8-14
6-7-17
5-2-23
27642,50
12
Zahak 10.0
2200
149.0/390
38.2
0-26-4
0-20-10
1-16-13
1-15-14
1-10-19
4-12-14
10-7-13
5-6-19
3-5-22
4-4-22
8-8-14
· ·· ·· ·
8-5-17
4-7-19
26407,00
13
Lc0 771721
2200
147.0/390
37.6
0-24-6
0-17-13
1-14-15
0-11-19
4-11-15
3-9-18
0-7-23
3-10-17
11-8-11
3-10-17
7-6-17
5-8-17
· ·· ·· ·
8-6-16
26644,00
14
Deep Hiarcs 15.1
2200
125.0/390
32.0
0-26-4
0-22-8
0-21-9
0-17-13
1-17-12
3-7-20
3-11-16
5-10-15
4-17-9
7-13-10
2-5-23
7-4-19
6-8-16
· ·· ·· ·
22135,25
2730 games played / Tournament finished
Tournament start: 2022.04.26, 12:49:53 Latest update: 2022.05.20, 19:17:40 Site/ Country: isle of Rügen, Germany Level: Blitz 0:30/2 Hardware: Intel(R) Core(TM) i3-7100 CPU @ 3.90GHz mit 7,9 GB Speicher Operating system: Windows 10 Enterprise Professional (Build 9200) 64 bit Table created with:Arena 3.5.1
Admin, Mclane and matejst like this post
Admin Admin
Posts : 2609 Join date : 2020-11-17 Location : Netherlands
Thank you for all your intensive tests. It looks like I am still on the right track, results give hope 1) it will better scale on longer time controls and 2) that the new net I am currently creating will be even better.
There have been no search changes in Rebel 15, it's only the neural net.
Mclane, TheSelfImprover and Dio like this post
Mclane
Posts : 3024 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
I noticed that there are lots of blunders in the lines given by Rebel, already after a dozen plies: e.g. (from today) a missed mate on the first rank, a blundered rook, in relatively simple positions. While I think the net is very good (it gives solutions very similar to accepted theory or to SF at similar depth; I still have not tested it in endings but I will soon), there seem to be some hidden problems with the search.
Admin Admin
Posts : 2609 Join date : 2020-11-17 Location : Netherlands
I will agree with you search is the weak part and my unwillingness to do something about that. In the meantime the [GPL] engine deserves something better and I am looking for volunteers
OTOH Boban, you are probably using the double sized and SSE version and what me wonder is what the average NPS is doing on your laptop.