Game Solving with Online Fine-Tuning
Ti-Rong Wu,1 Hung Guei,1 Ting Han Wei,2 Chung-Chin Shih,1,3 Jui-Te Chin,3 I-Chen Wu1,3
1 Academia Sinica, 2 University of Alberta, 3 National Yang Ming Chiao Tung University
NeurIPS 2023
Paper
Talks
Code
Introduction
BibTex
Solution Trees
This page provides the solution trees (in Smart Game Format; SGF) solved by the BASELINE, ONLINE-SP, ONLINE-CP, and ONLINE-SP+CP solvers. We recommend using GoGui to view these solution trees. Please refer to the bottom of this page for detailed instructions.
JA
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
JB
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
JC
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
JD
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
KA
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
KB
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
KC
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
KD
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
KE
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
KF
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
DA
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
SA
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
SB
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
SC
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
SD
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
SE
BASELINE
ONLINE-SP
ONLINE-CP
ONLINE-SP+CP
Detailed Instructions
To view the solution tree, we recommend using GoGui, which requires installing GoGui and Java on a computer with at least 16G of memory.
Step 1: Launch GoGui and open the solution tree file
First, launch GoGui with the argument -Xmx16g.
For example, in Windows, navigate to the GoGui installation directory and run javaw.exe -Xmx16g -jar GoGui.exe to launch it.
Once GoGui has been launched, use File > Open to open an SGF file to view, e.g., JA-baseline.sgf. Note that SGF files are provided in compressed format. They must be decompressed before use.
Step 2: Navigate to the Root Position
Once the SGF file is loaded, it will display an empty board.
Use Go > Forward to navigate to the next move multiple times until the root position is reached, i.e., until the comment panel displays some information.
The figure shows the root position of JA-baseline.sgf.
Step 3: Traverse the Proof Tree
Then, traverse the proof tree by clicking the intersections on the board.
The information is from the perspective of the player who played the latest move:
solver_status: WIN, LOSS, or UNKNOWN
p: policy probability
v: proof cost value
mean: MCTS Q value
count: MCTS simulation count
equal_loss: whether the node is pruned by the relevance zone
match_tt: whether the node is proven by the transposition table
check_ghi: whether the position has GHI issues to check
More instructions on traversing the proof tree
equal_loss: whether the node is pruned by the relevance zone of a sibling node
For the case of equal_loss = -1: this node is not pruned
For the case of equal_loss not equal to -1: this node is pruned by a sibling node
For example, after playing the move F4 from the positions in step 3, equal_loss = cg (corresponding to the C1 location of the board), meaning the sibling move C1 helps prune the move F4, and the solution tree below C1 is sufficient to replay the solution after playing F4.
Therefore, roll back one move and click C1 to further traverse the tree.
match_tt: whether the node is proven by the transposition table
For the case of match_tt = false:
If it is White to play, you can find at least one winning move leading to a node whose solver_status is WIN.
If it is Black to play, all next moves will lead to corresponding nodes whose solver_status is LOSS.
For the case of match_tt = true:
Press Ctrl + F to search for "rzone_data_index =" with the associated index in the comment, e.g., "rzone_data_index = ~975~", and find the position with the same "rzone_data_index = ~975~" and "match_tt = false", where you can continue to traverse the solution tree.
Note that while searching, you can press function key F3 for the next matching.
