1 Recognizing safe territories Presented by: Xiaozhen Niu Date: 2003/09/22
2 Outline n Introduction u Safety of Blocks and regions n Enhancements n Experimental Results n Future Work
3 Introduction n Recognizing whether territories are safe is a major task in Go evaluation. n Most cases need to be proven by using search, not static evaluation.
4 Safety of Blocks and Regions n Safety of blocks and regions are strongly related.
5 Result of the Example n The left graph shows that the write blocks and territory are all safe, the right graph shows that both are not safe.
6 Safe Blocks but Unsafe Regions n In some cases, Blocks can be safe even regions are not. For example:
7 Enhancements n 1. Recognizing external eyes for boundary blocks of a region. n 2. Merging regions when necessary to simplify the situation. n 3. Evaluation Function. n 3. Move ordering for local search.
8 External Eye Example n An external eye of the boundary block will be very helpful for evaluating region safety.
9 Merging regions n Previous program analyzes regions separately. However, in many cases, adjacent several regions should be treated as a combined region when evaluating.
10 Evaluation Function n Add more Go knowledge into the evaluation function, original values are only safe, unsafe and unknown. n Currently: u Value -= factor1* (Num of boundary blocks) u Value -= factor2* (Num of interior points)
11 Move Ordering n Also, original program generates random moves inside the region. Adding Go knowledge into move ordering should gain speedup in the search. n Adding knowledge such as cutting points, liberties of boundary blocks, liberties of inner attacker’s blocks to score all the points inside the region, then sort them by their scores.
12 Example n Score = f1*(self liberties) - f2*(opponent’s liberties) n Adjust a bit when changing players.
13 Experimental Results n Test sets: 31 games played by amateur dan players. n Step 1: adding external eyes. n Step 2: merging regions. n Step 3: evaluation function and move ordering. Original Step1Step2 Step3 34.9%36.2%41.2%???
14 Future Work n Recognizing saki.
15 Future Work n Decomposition large region to weakly depedant sub- regions, and implement local search for thoes sub- regions.
16 Reference n [1] M. Müller. Playing it safe: Recognizing secure territories in computer Go by using static rules and search. In H. Matsubara, editor, Game Programming Workshop in Japan '97, pages 80-86, Computer Shogi Association, Tokyo, Japan, 1997.