Crazy Zero

[Rémi Coulom]

Hi,

Here is a short report about my Crazy Zero experiments. Crazy Zero is a generic implementation of AlphaZero. I already used it for Go, shogi, gomoku, renju, othello, and chess.

The version that played on the Ataxx server (http://server.ataxx.org/) is less than one day old. The network architecture is 10 layers of 64 units. 4 inputs: empty, white, black, and (time-since-last-split / 100). The policy has one output for splits, and 16 outputs for jumps. I have only one convolutional output channel for the value: I simply apply the logistic function to the sum of values of this output. This way the network is purely convolutional, and could work on any board size.

I trained on a single machine (i9 + RTX 2080Ti), by iteratively generating self-play games (400 playouts per move), and training a new network (each time from scratch) with the data generated by the previous network.

At this moment, here is the number of games I generated with each network:

Code: Select all

   1      2048
   2      4096
   3      4096
   4      4096
   5      6144
   6     14336
   7     14336
   8     30720
   9    243712
  10    126976

And here are the self-play match results:

Code: Select all

-----------------------------------
  player opponent wins draws losses
-----------------------------------
1      2        1  738     0     30
2      3        2  718     0     50
3      4        3  715     0     53
4      5        4  613     0    155
5      6        5  543     1    224
6      7        6  489     2    277
7      8        7  458     0    310
8      9        8  591     0    177
9     10        9  569     0    199

So it seems that even this small network still has a big potential for improvement. From the results on the server, it seems already much stronger than the traditional alpha-beta searchers.

I am looking forward to challenging Ataxx Zero. Maybe this evening.

I am glad I discovered this game. It looks to be very rich, and at least as interesting as Othello.