ProbCut

From Chessprogramming wiki
Jump to: navigation, search

Home * Search * Selectivity * Pruning * ProbCut

Search Trees with two ProbCut generalizations [1]

ProbCut,
a selective search enhancement of the alpha-beta algorithm created in 1994 by Michael Buro as introduced in his Ph.D. thesis [2]. It permits to exclude probably irrelevant subtrees from beeing searched deeply. ProbCut and its improved variant Multi–ProbCut (MPC) have been shown to be effective in Othello [3] and Shogi [4], and a technique similar to ProbCut is used in the checkers program Chinook by Schaeffer et al. (1992) [5] described their approach in a footnote: Chinook performs forward cuts in positions with a material deficit, where a shallow search does not show an escape. ProbCut is a generalization of this method in that it is game independent. It was tested by incorporating it in Buro's already strong Othello program Logistello [6] and increased the program's playing strength [7]. Despite some promising results reported by Albert Xin Jiang and Michael Buro with Crafty [8], it seemed not that successful in chess programs which already perform Null Move Pruning and Late Move Reductions, until Stockfish proved otherwise as implemened by Gary Linscott [9] [10].

The Idea

ProbCut is based on the idea that the result v of a shallow search with depth d is a rough estimate of the result v of a deeper search with depth d > d'. A way to model this relationship is by means of a linear model:

v = a*v' + b + e

where e is a normally distributed error variable with mean 0 and standard deviation σ (sigma) or variance σ². If the evaluation function is relative stable, the slope a is about 1.0, offset b about 0.0 and a small variance σ². The cutoff condition of depth d

v ≥ β

becomes

(v' + b - β)/σ ≥ -e/σ

since -e/σ is normally distributed with mean 0 and variance 1 (and distribution function Φ, phi), the condition holds true with probability of at least p iff

(v' + b - β)/σ ≥ Φ-1(p)

which is equivalent to

v' ≥ (Φ-1(p) * σ + β - b) / a

Similar for

v ≤ α

the condition becomes

v' ≤ (-Φ-1(p) * σ + α - b) / a

Pseudo Code

This observation immediately leads to the implementation of the ProbCut alpha-beta extension for one depth and reduced depth pair using floats. Before sigma, a and b are estimated by linear regression likely for different game phases, the search depths d and d' < d and cut threshold must be chosen or be determined empirically, by checking the performance of the program with various parameter settings [11] .

int alphaBetaProbCut(int α, int β, int depth) {
   const float T(1.5);
   const int DP(4);
   const int D(8);

   if ( depth == 0 ) return evaluate();

   if ( depth == D ) {
      int bound;

      /* v >= β with prob. of at least p? yes => cutoff */
      bound = round( ( T * σ + β - b) / a );
      if ( alphaBetaProbCut( bound-1, bound, DP) >= bound )
         return β;

      /* v <= α with prob. of at least p? yes => cutoff */
      bound = round( (-T * σ + α - b) / a );
      if ( alphaBetaProbCut( bound, bound+1, DP) <= bound )
         return α;
   }
   /* the remainder of alpha-beta goes here */
   ...
}

Multi–ProbCut

Multi–ProbCut (MPC) enhances ProbCut by

  • Allowing different regression parameters and cut thresholds for different stages of the game
  • Using more than one depth pair
  • Using internal iterative deepening for shallow searches
struct Param {
   int d;         /* shallow search depth */
   float t;       /* cut threshold */
   float a, b, σ; /* slope, offset, standard deviation */
} param[MAX_STAGE+1][MAX_HEIGHT+1][NUM_TRY];


int alphaBetaMPC(int α, int β, int depth)
{
   if ( depth == 0 ) return evaluate();

   if ( depth <= MAX_D ) {
      for (int i=0; i < NUM_TRY; i++) {
         int bound;
         const Param &pa = param[stage][depth][i];

         if (pa.d < 0 )
            break; /* no more parameters available */

         /* v >= β with prob. of at least p? yes => cutoff */
         bound = round( ( pa.t * pa.σ + β - pa.b) / pa.a );
         if ( alphaBetaMPC( bound-1, bound, pa.d) >= bound )
            return β;

         /* v <= α with prob. of at least p? yes => cutoff */
         bound = round( (-pa.t * pa.σ + α - pa.b) / pa.a );
         if ( alphaBetaMPC( bound, bound+1, pa.d) <= bound )
            return α;
      }
   }
   /* the remainder of alpha-beta goes here */
   ...
}

ProbeCut or MPC in Chess

In 2003, Albert Xin Jiang implemented ProbCut and MPC in Crafty by Robert Hyatt. In his thesis he introductory elaborates on ProbCut in Chess [12] :

There has been no report of success for ProbCut or MPC in chess thus far. There are at least two reasons for this:
  1. Null-move is available for chess. Null-move and ProbCut are based on similar ideas, as a result they tend to prune the same type of positions. Part of the reason why ProbCut is so successful in Othello is that null-move does not work in Othello. But in chess, ProbCut and MPC have to compete with null-move, which is way better than brute-force alpha-beta.
  2. Chess searches tend to make more mistakes than Othello searches [13]. This leads to a larger standard deviation in the linear relationship between shallow and deep search results, which makes it harder to get ProbCut cuts.


In his research, Albert Xin Jiang further determined following parameters by linear regression. The about 2700 positions were chosen randomly from some computer chess tournament games and some of Crafty’s games against human grandmasters on internet chess servers:

CraftyLinearReg.JPG

v' versus v for depth pair (4,8) [14]

Linear regression results. The evaluation function’s scale is 100 = one pawn. r is the regression correlation coefficient, a measure of how good the data fits the linear model:

Pairs Stage a b σ r
(3,5) Middlegame 0.998 -7.000 55.80 0.90
(3,5) Endgame 1.026 -4.100 51.80 0.94
(4,8) Middlegame 1.020 2.360 82.00 0.82
(4,8) Endgame 1.110 1.750 75.00 0.90

While ProbCut did not result in better playing strength of Crafty, Albert Xin Jiang and Michael Buro report an improvement with MPC while playing two times three 64-game matches with three Crafty versions and two time controls versus Dieter Bürßner's program Yace [15] :

Pairing Crafty %
2min +
10sec/move
8min +
20sec/move
Crafty - Yace 42.0 50.8
MPC Crafty(1.2, 1.0) - Yace 53.1 56.3
MPC Crafty(1.0, 1.0) - Yace 57.0 55.5

However, Robert Hyatt first stated results were inconclusive [16] , and later that MPC was somewhat worse in every test he tried [17] , also confirmed by Robert Allgeuer, who performed Crafty MPC tests with following conclusion in CCC [18] :

My tests indicate that the overall playing strength of Crafty 18.15 remains more or less unchanged by the addition of Multi-ProbCut. However, the characteristic of the engine changes significantly due to ProbCut: Even though nominal search depth is increased by one to two plies, tactical strength is severely reduced.
Furthermore with ProbCut match results become more unpredictable and inconsistent: Apparently there are types of opponents against which ProbCut works very well and results in significantly improved results, but there are also other opponents (the tactically stronger ones?) where ProbCut has exactly the opposite effect. 

See also

Publications

[19] [20]

Forum Posts

External Links

References

  1. Illustrations of the first two ProbCut generalizations: a) allowing forward cuts at several heights and b) performing a sequence of check searches of increasing depth from Michael Buro (1997). Experiments with Multi-ProbCut and a New High-quality Evaluation Function for Othello. Technical Report No. 96, NEC Research Institute, Princeton, N.J. pdf, 8 Multi-ProbCut, pp 7.
  2. Michael Buro (1994). Techniken für die Bewertung von Spielsituationen anhand von Beispielen. Ph.D. Thesis. Paderborn University, Paderborn, Germany. (German), Kapitel 4. Selektive Suche
  3. Michael Buro (1997). Experiments with Multi-ProbCut and a New High-quality Evaluation Function for Othello. Technical Report No. 96, NEC Research Institute, Princeton, N.J. pdf
  4. Kazutomo Shibahara, Nobuo Inui, Yoshiyuki Kotani (2002). Effect of ProbCut in Shogi - by changing parameters according to position category. 7th Game Programming Workshop
  5. Jonathan Schaeffer, Joe Culberson, Norman Treloar, Brent Knight, Paul Lu, Duane Szafron (1992). A World Championship Caliber Checkers Program. Artificial Intelligence, Vol. 53, Nos. 2-3, pp. 6, "If a line leads to a material deficit and there is insufficient positional compensation (the positional assessment of the position does not reach a fixed threshold) then the remaining search depth is cut in half. If the result of this search does not restore the material deficit, or result in sufficient positional compensation, this line is abandoned. Otherwise, the line is researched to the full search depth".
  6. LOGISTELLO's Homepage
  7. Michael Buro (1995). ProbCut: An Effective Selective Extension of the Alpha-Beta Algorithm. ICCA Journal, Vol 18, No. 2, pdf
  8. Albert Xin Jiang, Michael Buro (2003). First Experimental Results of ProbCut Applied to Chess. Advances in Computer Games 10
  9. Probcut by Gary, CCC, May 24, 2013
  10. Stockfish/search.cpp at master · official-stockfish/Stockfish · GitHub, Step 9. ProbCut
  11. Michael Buro (1995). ProbCut: An Effective Selective Extension of the Alpha-Beta Algorithm. ICCA Journal, Vol 18, No. 2, pdf
  12. Albert Xin Jiang (2003). Implementation of Multi-ProbCut and Chess. CPSC 449 Thesis, pdf, 2.5 ProbeCut in Chess, pp. 6
  13. Andreas Junghanns, Jonathan Schaeffer, Mark Brockington, Yngvi Björnsson, Tony Marsland (1997). Diminishing Returns for Additional Search in Chess. Advances in Computer Chess 8
  14. Image from Albert Xin Jiang (2003). Implementation of Multi-ProbCut in Chess. CPSC 449 Thesis, pdf
  15. Page 30 in Albert Xin Jiang, Michael Buro (2003). First Experimental Results of ProbCut Applied to Chess. Advances in Computer Games 10, pdf
  16. Re: Multi-ProbCut and Crafty : does it work ? by Robert Hyatt, CCC, June 28, 2003
  17. Re: ProbCut: An Effective Selective Extension of the Alpha-Beta Algorithm by Robert Hyatt, CCC, July 21, 2004
  18. Crafty MPC tests (long post) by Robert Allgeuer, CCC, October 18, 2003
  19. Michael Buro's Publication List
  20. Albert Xin Jiang - publications
  21. Fuego, Go playing program by Markus Enzenberger, Martin Müller, Broderick Arneson, Richard Segal, Gerald Tesauro and Arpad Rimmel

Up one Level