Improve Swap2

[nguyencongminh090]

I have a method to balance opening (swap2 with 3 stones), that is:

• Step 1: Add 1 white move which one has connection with previous white stone. That move should be near with one in two black stone (in range 5 coordinates).
• Step 2: Add black move which one can balance the current state of opening after the white move. And this is a video I made to describe it: Link

Hope it is helpful!

[MaciejKozarzewski]

Will this algorithm work with arbitrary number of stones? For example, to find 4 balancing moves for a 7 stone position?

[nguyencongminh090]

I'm not sure, but it can be more complex. In that case, the first move will depend on the nearby move to make a connection and then the second stone will balance after making step 1. (Swap2 with n stones) What do you mean about "4 balancing moves for a 7 stone position"? I don't understand why 4 balancing moves. I think this algorithm only works with 2 balancing moves.

[MaciejKozarzewski]

This is why I asked. Algorithm currently used in AG is general - can be used to balance arbitrary position using any number of moves. I agree it could be optimized but currently I don't see any possibility for doing so.

[nguyencongminh090]

Hmm... What is your current algorithm? Can you give me some example case? I wanna try, I hope I can help.

[nguyencongminh090]

solution for the standard case. Because it is quite different when swap2 with 3 stones, you only have 3 stones and 1 threat from the connection of 2 stones of black but with more than 3 stones, it has a lot of potential threat, connection. Can we use 2 ways for 2 types? (I want to discuss more this problem via email)

[MaciejKozarzewski]

Neural network calculates probability of winning for given position. During match the search looks for a move that has highest probability of winning. During balancing I change the function to look for move that has probability of winning closest to 0.5. And that's it. This is why I think my algorithm is better as it is provably correct and not based on a heuristic approach. Look into [https://github.com/MaciejKozarzewski/AlphaGomoku/blob/master/src/mcts/EdgeSelector.cpp]() for details.

[nguyencongminh090]

Hmm.... I think we need a test.

[nguyencongminh090]

My testcase base on connection of 2 stones of black. Testcase 1: Long connection:

• h3e8m5 (balanced)
• h4g12c3 (not balanced)
• h8d4i13 (surewin for black)
• f11j6c6 (surewin for white)
• b3k10g2 (balanced)

Testcase 2: Near connection:

• k8g7j12
• h8h7h12
• h8j9f6
• h8h10l10
• h8f8k10

Testcase 3: No connection.

• n4m2k2
• n3o1m2
• c13c12d13
• b3c2g2
• o14m14m12