Multipurpose Adversary Planning in the Game of Go Ph.D thesis by Shui Hu Presenter: Ling Zhao Date: November 18, 2002
Outline Motivations Overview Basic structures and concepts Combining goals A concrete example Conclusions
Motivations Go is a strategy game Computer Go programs still have difficulties to convey human’s knowledge Multi-purpose moves are quite common in human-played Go games Traditional multi-valued move is too passive!
Overview Heuristic adversary planning Static analysis and dynamic look ahead Look for combined goals and the steps leading to them Strength: actively search for combined goals Weakness: only a prototype, hard to implement in a real system
Basic structures Hierarchy of objects: group, chain, string, stone Generate goals: 1. high level goals are generated by static analysis (using knowledge base) 2. low level goals are generated by looking ahead Goal structure (see example next next page)
Knowledge base
Goal Tree
Goal relations Goal and subgoals Master and servant goal
Achievability Black to move, and the goal is to kill white group Achievable: Near-achievable:
Decide the achievability Start from leaf goals Generate goal/counter goal pairs Use look ahead search Propagate results upward Note this method can also decide the achievability of combined goals
CP2 search procedure g11 : (g11, c11) (g11, c12) (g11,c21) … (g11, c33) g12 : (g12, c11) (g12, c12) (g12,c21) … (g12, c33) g33 : (g33, c12) (g33, c12) (g33,c21) … (g33, c33)
Interaction of leaf goals If the intersection of moves to realize two near- achievable goals is not empty, we find some multipurpose moves! Combine two goals and use the look ahead to decide if the multipurpose moves can make one of the near-achievable moves achievable. If yes, your multipurpose planning works. The example explains the situation similar to double threats, and there are more situations.
Example
Goal Tree
Conclusions A multipurpose planning framework was brought out Prototype, can only work on very few finely designed example Planning is weak, and almost the same as search.