SmartGo: How it works and how it doesnt

1
SmartGo How it works(and how it doesnt)

• Anders Kierulf
• Smart Go, Inc.

2
Overview

• How SmartGo chooses a move
• Move selection
• Position evaluation
• What works
• What doesnt

3
Algorithm for Play a Move

• Choose time per move
• Full-board selective search
• Move selection Whatever works
• Position evaluation As well as possible
• Dont care how position is reachedNo urgency
  value for moves
• Pick best move when out of time

4
Full-board search Good

• Ignores small stuff, sees big tradeoffs
• Avoid some blunders
• Lookahead handles interaction between groups,
  multi-purpose moves
• Playing it out rather than guessing effect of a
  move
• Most positions have acceptable move
• Iterative deepening widening

5
Full-board search Bad

• Horizon effect
• Non-local moves
• Quiescence
• Uncertainty in evaluation
• Hard to evaluate positions fast enough
• Only 3-5 ply lookahead

6
Move Selection

• Some patterns (hard-coded)
• Move generators Play in empty corner, extend
  along side,
• Tactical moves Block captures, LD (GoTools),
  connections, semeai
• Pick N moves with highest value (N15 at top
  level, N10 below, N5 deep in tree)

7
Move Selection Good

• Most positions have acceptable move
• Just make sure good move gets generated, dont
  care why
• E.g. generate opponents response
• Move selection independent of evaluation of
  resulting position
• Can easily add new move generators

8
Move selection Bad

• Ad-hoc, significant gaps
• No joseki moves yet
• Hard to tune move values (value inflation)

9
Position evaluation

• Safe areas (Benson Müller)
• Capture tactics
• Influence
• Links (patterns influence)
• Groups (areas inside links)
• Tactics Life death, connection, semeai

10
Position Evaluation Good

• Sequence to position doesnt matter
• Groups are alive, dead, unsettled, or unknown
• Area counted full or moyo (1/3 point)
• Tactical results get cached
• Reduce area around unstable boundaries

11
Position Evaluation Bad

• Hard to evaluate position fast enough
• Dependency area for tactics too large
• No concept of weak group strength
• No idea of good shape
• No idea of ko threats

12
GoTools for Life Death

• GoTools by Thomas Wolf
• Superb on small enclosed problems
• Great when it can solve the problem

13
GoTools Issues

• GoTools takes too long
• Heuristic version not accurate
• Always enclosing problem
• Boundary may affect solution
• Boundary may be poorly chosen
• No boundary change during search
• Poor heuristics if GoTools fails
• Too accurate Doesnt try if dead

14
Future Local search

• Divide board into independent areas
• Search for both players in each area
• Combine using combinatorial game theory
• Faster (fewer moves, smaller area)
• More accurate evaluation
• Look deeper while area unsettled
• But New problems to worry about!