First, I finished rounds 16-20, though I haven’t yet played the winner. It’s getting a bit cumbersome to post images of the output after each round, so instead, I’ll be posting data from my spreadsheets.
Second, I bought an older machine on Craigslist to use as a home media server. It’s got a quad-core AMD processor, so I also grabbed some more RAM and a new graphics card; specifically, the NVidia GTX 1060 with 3GB of VRAM. While the 6Gb model performs better in gaming, the 3GB model performs identically in compute, while still being decent for gaming at 1080p (the TV it’ll be hooked up to). This means I’ll be able to finish training rounds much more quickly.
Also, I’m going to be re-tooling my utilities to output the data in a more friendly format. For now, though, here are the results of rounds 16-20.
Round Robin Tournament:
| Round 16 | Round 17 | Round 18 | Round 19 | Round 20 | |||
| Smart Random | 60 | 60 | 60 | 60 | 60 | ||
| 1 | x | 100 | 30 | 27 | 27 | 38 | 34 |
| 1 | x | 250 | 30 | 29 | 30 | 29 | 27 |
| 1 | x | 500 | 31 | 29 | 30 | 21 | 29 |
| 1 | x | 1000 | 30 | 32 | 30 | 25 | 28 |
| 1 | x | 2500 | 35 | 28 | 28 | 27 | 25 |
| 1 | x | 5000 | 29 | 25 | 26 | 31 | 30 |
| 2 | x | 100 | 29 | 32 | 36 | 29 | 27 |
| 2 | x | 250 | 31 | 25 | 23 | 29 | 28 |
| 2 | x | 500 | 25 | 27 | 34 | 31 | 27 |
| 2 | x | 1000 | 29 | 28 | 34 | 28 | 26 |
| 2 | x | 2500 | 22 | 33 | 26 | 33 | 22 |
| 2 | x | 5000 | 27 | 32 | 26 | 31 | 29 |
| 3 | x | 100 | 33 | 26 | 33 | 32 | 36 |
| 3 | x | 250 | 27 | 30 | 25 | 26 | 27 |
| 3 | x | 500 | 31 | 25 | 27 | 18 | 28 |
| 3 | x | 1000 | 29 | 25 | 35 | 28 | 24 |
| 3 | x | 2500 | 26 | 30 | 27 | 30 | 26 |
| 3 | x | 5000 | 24 | 29 | 29 | 30 | 30 |
| 4 | x | 100 | 35 | 35 | 31 | 34 | 40 |
| 4 | x | 250 | 32 | 30 | 29 | 34 | 30 |
| 4 | x | 500 | 22 | 24 | 22 | 29 | 24 |
| 4 | x | 1000 | 25 | 26 | 32 | 27 | 24 |
| 4 | x | 2500 | 26 | 23 | 24 | 26 | 26 |
| 4 | x | 5000 | 28 | 35 | 32 | 29 | 36 |
| 5 | x | 100 | 35 | 39 | 38 | 39 | 38 |
| 5 | x | 250 | 31 | 34 | 29 | 29 | 30 |
| 5 | x | 500 | 33 | 21 | 22 | 28 | 25 |
| 5 | x | 1000 | 29 | 33 | 26 | 22 | 32 |
| 5 | x | 2500 | 23 | 30 | 27 | 29 | 28 |
| 5 | x | 5000 | 30 | 27 | 29 | 24 | 30 |
Out of 100 games against the SRP:
| Round 16 | Round 17 | Round 18 | Round 19 | Round 20 | |||
| 1 | x | 100 | 21 | 26 | 23 | 22 | 25 |
| 1 | x | 250 | 26 | 30 | 24 | 30 | 34 |
| 1 | x | 500 | 35 | 43 | 35 | 52 | 47 |
| 1 | x | 1000 | 45 | 45 | 47 | 51 | 51 |
| 1 | x | 2500 | 42 | 50 | 53 | 42 | 46 |
| 1 | x | 5000 | 45 | 57 | 44 | 51 | 54 |
| 2 | x | 100 | 34 | 22 | 32 | 25 | 25 |
| 2 | x | 250 | 40 | 35 | 30 | 34 | 37 |
| 2 | x | 500 | 36 | 39 | 38 | 38 | 47 |
| 2 | x | 1000 | 38 | 46 | 48 | 50 | 50 |
| 2 | x | 2500 | 39 | 53 | 41 | 45 | 53 |
| 2 | x | 5000 | 38 | 41 | 38 | 37 | 49 |
| 3 | x | 100 | 32 | 37 | 21 | 27 | 24 |
| 3 | x | 250 | 37 | 34 | 29 | 34 | 29 |
| 3 | x | 500 | 34 | 32 | 35 | 46 | 48 |
| 3 | x | 1000 | 32 | 39 | 28 | 41 | 47 |
| 3 | x | 2500 | 38 | 46 | 35 | 35 | 45 |
| 3 | x | 5000 | 34 | 41 | 39 | 41 | 46 |
| 4 | x | 100 | 10 | 19 | 20 | 14 | 27 |
| 4 | x | 250 | 28 | 24 | 38 | 33 | 28 |
| 4 | x | 500 | 30 | 35 | 25 | 37 | 39 |
| 4 | x | 1000 | 36 | 38 | 37 | 38 | 45 |
| 4 | x | 2500 | 35 | 50 | 37 | 47 | 41 |
| 4 | x | 5000 | 36 | 41 | 52 | 36 | 39 |
| 5 | x | 100 | 3 | 8 | 13 | 13 | 21 |
| 5 | x | 250 | 27 | 28 | 25 | 24 | 29 |
| 5 | x | 500 | 16 | 26 | 34 | 38 | 43 |
| 5 | x | 1000 | 34 | 30 | 40 | 40 | 43 |
| 5 | x | 2500 | 31 | 47 | 40 | 44 | 45 |
| 5 | x | 5000 | 43 | 41 | 39 | 43 | 49 |
Results
Now, grouping the results by width and depth, we have the following for the Round Robin:
| Depth | |||||
| 1 | 51 | 47 | 48 | 48 | 48 |
| 2 | 45 | 49 | 50 | 50 | 44 |
| 3 | 47 | 46 | 49 | 46 | 48 |
| 4 | 47 | 48 | 47 | 50 | 50 |
| 5 | 50 | 51 | 48 | 48 | 51 |
| Width | |||||
| 100 | 54 | 53 | 55 | 57 | 58 |
| 250 | 50 | 49 | 45 | 49 | 47 |
| 500 | 47 | 42 | 45 | 42 | 44 |
| 1000 | 47 | 48 | 52 | 43 | 45 |
| 2500 | 44 | 48 | 44 | 48 | 42 |
| 5000 | 46 | 49 | 47 | 48 | 52 |
Whereas before there was a slight tendency toward deeper networks, it is ever so slightly less pronounced here (the 5 layer networks collectively won 51% of their games instead of 52%). Also, we seem to have reversed the trend of narrow networks winning… With the exception of the 100 node wide ones, that is. They simply cleaned up.
Now, for absolute number of games won out of 100 matches against the SRP:
| 1 | x | 36 | 42 | 38 | 41 | 43 | |
| 2 | x | 38 | 39 | 38 | 38 | 44 | |
| 3 | x | 35 | 38 | 31 | 37 | 40 | |
| 4 | x | 29 | 35 | 35 | 34 | 37 | |
| 5 | x | 26 | 30 | 32 | 34 | 38 | |
| x | 100 | 17 | 19 | 18 | 17 | 20 | |
| x | 250 | 26 | 25 | 24 | 26 | 26 | |
| x | 500 | 25 | 29 | 28 | 35 | 37 | |
| x | 1000 | 31 | 33 | 33 | 37 | 39 | |
| x | 2500 | 31 | 41 | 34 | 36 | 38 | |
| x | 5000 | 33 | 37 | 35 | 35 | 40 |
This is quite promising. Almost all categories collectively improved (the 250 node wide networks had no change as a group). Additionally, the shallow networks’ lead over the deeper networks has narrowed considerably. We’ll see what additional rounds bring, if the deep networks finally catch up.
As for the wide networks, they maintain roughly the same advantage over the narrow networks as before.