1 Life-and-Death Problem Solver in Go Author: Byung-Doo Lee Dept of Computer Science, Univ. of Auckland Presented by: Xiaozhen Niu
2 Outline Introduction Pattern Clustering Eye Shape Analysis Game Tree Search Conclusions
3 Introduction Problem: life-and-death of groups Major issues: Infeasible by brute-force search Goal: using heuristic model to reduce branching factor!
4 Basic Components of Solver
5 Pattern Clustering Similar life-and-death problems often have similar solutions (a similar first move to kill or live…) Group the input patterns into different clusters (no predefined clusters) Goal: using the first moves of the clusters as the candidate first move
6 Pattern Classifier Three clustering methods: Euclidean distance based Vector product based Kohonen neural network based
7 Euclidean distance based Clustering Calculate distances between the input pattern and the weighted center of each cluster Find the closest cluster within the range of the threshold P
8 Examples
9 Vector Product Based Clustering Calculate similarity degree (cosØ) between instance vector and centroid vector of each cluster cosØ is 1 => same cosØ is -1 => totally different
10 Examples
11 Results Euclidean distance based clustering is the best with lower threshold (<=3)
12 Eye Shape Analysis Basic of Eye shape Heuristic influence function
13 Eye Shape A surrounded group (A, B, C, D: E) A: num of points with 4 neighbors, B: with 3 neighbors, C: with 2 neighbors, D: with 1 neighbor E: Status: Alive, Dead or Unsettled
14 Examples
15 Heuristic Influence Function Surrounding groups and surrounded groups both radiate influence to the surrounded area
16 Basic Steps 1: Find virtual boundary (radial sweep algorithm) 2: Calculate influence of surrounding and surrounded groups 3: calculate the number of neighbors of zero influence points 4: result point set forms the eye shape
17 Example (1)
18 Example (2)
19 Results 30 problems (size <=10, not completely surrounded)
20 Game Tree Search Selective alpha-beta search Using pattern clustering and eye shape analysis to generate a set of first moves Only in depth 1
21 Evaluation Function
22 Examples
23 Results
24 Conclusions Using pattern clustering and eye shape to do selective search Weaknesses: similar patterns often has similar first moves to kill, but NOT always! Eye shape accuracy too low! (36.7%) Size limitation (<=10), not very useful in real games …