Igel 3.3.0 at
https://github.com/vshcherbyna/igel/releases/tag/3.3.0 (official executable binaries for Windows and IGN net).
This release brings adequate improvement in strength as a result of a switch to a new network architecture (halfkav2) and re-training network from scratch using existing Igel data as well as newly generated data from Igel 3.2.0.
In order to improve the quality of existing data some work was done to 'sanitize' my old data files. In addition to the sanitization, there were also some improvements in Igel's training data generator such as
adding hash table to avoid duplications and
more stricts fen checkes that helped to produce a higher quality of Igel 3.2.0 data used for training.
I also experimented with d5 positions that resulted in ability to generate data quickly to get some elo in initial phases of the training. Network is trained of d5 + d8 Igel positions with d16 validation Igel positions.
This version only supports AVX2 builds but I am going to look at adding AVX512 support in future releases.
What's new:
- new network architecture
- new network ign-2-51ba2968
Regression run against Igel 3.2.0 on Fast Time Control (3+0.03)
Code: Select all
Score of Igel 3.3.0 64 BMI2 AVX2 vs Igel 3.2.0 64 BMI2 AVX2: 1713 - 756 - 2531 [0.596] 5000
... Igel 3.3.0 64 BMI2 AVX2 playing White: 1073 - 258 - 1169 [0.663] 2500
... Igel 3.3.0 64 BMI2 AVX2 playing Black: 640 - 498 - 1362 [0.528] 2500
... White vs Black: 1571 - 898 - 2531 [0.567] 5000
Elo difference: 67.3 +/- 6.8, LOS: 100.0 %, DrawRatio: 50.6 %