Presentation is loading. Please wait.

Presentation is loading. Please wait.

Planning in Go Ling Zhao University of Alberta September 15, 2003.

Similar presentations


Presentation on theme: "Planning in Go Ling Zhao University of Alberta September 15, 2003."— Presentation transcript:

1 Planning in Go Ling Zhao University of Alberta September 15, 2003

2 Outline Motivations Planning Big picture Current work

3 Motivations Exhaustive search not longer effective in Go? - very large branching factor - difficulties in evaluation - search tree is very deep - human is daunting

4 Motivations Solutions: - selective search - embody human knowledge - many other approaches such as neural net - planning systems

5 Planning? Goal-driving approach different from the traditional data-driven approach Included more or less in many state-of-art programs - Use goals to generate moves - but typically do not check goals later or maintain the goals during several moves

6 Problems Two frameworks 1. S. Hu: Multipurpose strategic planning 2. S. Willmott: Adversarial planning Not applied to real Computer Go programs. Lots of work to define goals, refine goals and generate moves.

7 Work to do Embody a planning system in Explorer Expectations (best case) - a module separate from other parts of Explorer - information hidden - useful: better results but not significant overhead

8 Current work - Dynamic look ahead Problems in Explorer Lack of look ahead - mainly static analysis Implement Minimax + selective search

9 Example

10 Other problems Results from selective search can be wrong! How to find a better trade-off?


Download ppt "Planning in Go Ling Zhao University of Alberta September 15, 2003."

Similar presentations


Ads by Google