Atari Go competition in CodinGame

[Rémi Coulom]

Hi,

There is a 9x9 atari-go competition in codingame:
https://www.codingame.com/multiplayer/b ... ari-go-9x9
also on 19x19:
https://www.codingame.com/multiplayer/b ... g/atari-go

I will prepare an entry. The really fun part is that an entry must be a single source-code file with 100k chars max. It is enough to implement a convnet and MCTS, but the net has to be very small. I like this constraint. It is not about who has got the most GPUs, and you have to find clever ideas.

The competition has a lot of participants, many of them very weak. Having plenty of very weak participants makes it more enjoyable for newcomers. I will very soon prepare plenty of weak bots to populate CGOS. I hope it can make it a bit more attractive.

Rémi
(my codingame handle is Crazy_Remi: https://www.codingame.com/profile/bf359 ... 8594717273)

[flok]

cool!

[Rémi Coulom]

So I entered a few codingame competitions last week. All my entries use a small resnet made of 17 layers of 16 channels. The size limit is 100,000 utf8 characters, not the number of bytes. So I used a trick to store one floating point number in a single utf8 character. I made a generic system to export my Crazy Zero bots to codingame, so I could enter plenty of competitions. I also implemented new games: counting tic tac toe, and amazons. Here is a summary of games:

• Atari Go 9x9: 1/160 https://www.codingame.com/multiplayer/b ... eaderboard
  I guess the top opponents were using mcts, maybe without patterns. My resnet beats them easily, but when I watch the games I can see the mistakes my program makes.
• Atari Go: 5/120 https://www.codingame.com/multiplayer/b ... eaderboard
  The rules are a little strange. Because the game has to stop at 200 moves, it is a little broken. But it is interesting to see that I cannot beat those silly heuristics with a neural network trained with the Alpha Zero method. It makes me want to research this more.
• Amazons: 1/102 https://www.codingame.com/multiplayer/b ... eaderboard
  Amazons is really well suited for convolutional neural networks.
• Connect 4: 4/524 https://www.codingame.com/multiplayer/b ... eaderboard
  I did not expect to perform well at that game at all, against fast programs with endgame solvers. I was surprised to reach such a high rank so fast. Complex patterns seem to be important in Connect 4 after all.
• Counting Tic-Tac-Toe: 1/299 https://www.codingame.com/multiplayer/b ... eaderboard
  Not a very interesting game, but the rules are so easy to implement I decided to have a go anyway. It is unfortunate that mirror play is an optimal policy and ensures a draw.
• Othello: 9/421 https://www.codingame.com/multiplayer/b ... eaderboard
  The strongest players are much stronger. I guess they may have opening books, fast endgame solvers, and fast pattern-based evaluation in-between. That makes them difficult to fight.
• Breakthrough: 5/267 https://www.codingame.com/multiplayer/b ... eaderboard
  Here also, the strongest programs are much stronger. Probably also using fast search with simple pattern tables, but I am not sure.
• Twixt: 1/73 https://www.codingame.com/multiplayer/b ... eaderboard
  12x12 is too small. I guess my entry is playing close to perfection. Not many participants.
• Chess: Unfortunately, my chess data structures are too complex, and don't fit in the 100,000 character limit. I could implement a minimalist version of the rules of chess, but I was too lazy to do it.

It was a lot of fun. I enjoy competing on codingame very much.

[zakki]

How about ULTIMATE TIC-TAC-TOE?

https://www.codingame.com/multiplayer/b ... ic-tac-toe
My simple MCTS bot without neural network wasn't good enough.

[Rémi Coulom]

Oh, you rank is very close to mine! I entered that competition long ago with a plain alpha-beta search.

It is interesting to think about a deep-learning solution to this game. I don't expect standard convolutional network would work so well, because there is this local/global duality that is difficult to grasp with 3x3 local convolutions only. I have thought about inserting a special transposition operator in the network that would swap points like this (0 is padding):

Code: Select all

 0  0  0  0  0  0  0  0  0  0  0  0  0
 0 A1 A2 A3  0 B1 B2 B3  0 C1 C2 C3  0
 0 A4 A5 A6  0 B4 B5 B6  0 C4 C5 C6  0
 0 A7 A8 A9  0 B7 B8 B9  0 C7 C8 C9  0
 0  0  0  0  0  0  0  0  0  0  0  0  0
 0 D1 D2 D3  0 E1 E2 E3  0 F1 F2 F3  0
 0 D4 D5 D6  0 E4 E5 E6  0 F4 F5 F6  0
 0 D7 D8 D9  0 E7 E8 E9  0 F7 F8 F9  0
 0  0  0  0  0  0  0  0  0  0  0  0  0
 0 G1 G2 G3  0 H1 H2 H3  0 I1 I2 I3  0
 0 G4 G5 G6  0 H4 H5 H6  0 I4 I5 I6  0
 0 G7 G8 G9  0 H7 H8 H9  0 I7 I8 I9  0
 0  0  0  0  0  0  0  0  0  0  0  0  0

transformed to:

Code: Select all

 0  0  0  0  0  0  0  0  0  0  0  0  0
 0 A1 B1 C1  0 A2 B2 C2  0 A3 B3 C3  0
 0 D1 E1 F1  0 D2 E2 F2  0 D3 E3 F3  0
 0 G1 H1 I1  0 G2 H2 I2  0 G3 H3 I3  0
 0  0  0  0  0  0  0  0  0  0  0  0  0
 0 A4 B4 C4  0 A5 B5 C5  0 A6 B6 C6  0
 0 D4 E4 F4  0 D5 E5 F5  0 D6 E6 F6  0
 0 G4 H4 I4  0 G5 H5 I5  0 G6 H6 I6  0
 0  0  0  0  0  0  0  0  0  0  0  0  0
 0 A7 B7 C7  0 A8 B8 C8  0 A9 B9 C9  0
 0 D7 E7 F7  0 D8 E8 F8  0 D9 E9 F9  0
 0 G7 H7 I7  0 G8 H8 I8  0 G9 H9 I9  0
 0  0  0  0  0  0  0  0  0  0  0  0  0

I also wonder if it is better to represent the input as 9 channels of 3x3, or a single channel of 11x11 (with padding). I will give it a try.

[Rémi Coulom]

So I trained a CNN with the 11x11 board input encoding. I reached 1st place of the Gold League, but did not promote.
https://www.codingame.com/multiplayer/b ... value=gold
I was ahead of the boss for a short while during my run. I am really not far from promotion. I will try harder.