How To Rate Bridge Players and Other Paired Competitors T. S. Michael U. S. Naval Academy Annapolis, Maryland tsm@usna.edu http://www.usna.edu/Users/math/tsm/ joint work with Thomas Quint and Jeff Mortensen University of Nevada Joint National Mathematics Meetings, Baltimore, January 2014 T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 1 / 13 Summary We rate individual bridge players using least squares The resulting linear system has combinatorial interest We use Moore-Penrose inverse for singular systems We apply our results to real-world data Our method can be used to rate other paired competitors (doubles tennis) T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 2 / 13 Current Method to Rate Bridge Players (ACBL) Masterpoints individuals accumulate points over their lives ratings never decrease Drawbacks strong, infrequent players have low ratings some weak, frequent players have high ratings Issues individuals play with partners pairs compete against many pairs simultaneously Data Available list of partnerships partnership scores over many sessions T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 3 / 13 Partnership Multigraph G (for a year, say) 1 4 .62 .4 2 Players Partnerships Partnership scores Usually, 3 .5 .5 .5 .5 .6 5 ←→ ←→ ←→ vertices edges edge labels 0.3 ≤ partnership score ≤ 0.7 T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 4 / 13 Averaging Method rating of player i xi player i 1 2 3 4 5 total score si 1.12 1.62 2.50 1.00 1.00 T. S. Michael (U. S. Naval Academy) degree di 2 3 5 2 2 = average partnership score = si di rating xi 0.56 0.54 0.50 0.50 0.50 1 4 .62 2 How To Rate Bridge Players 3 .5 .4 .5 .5 .6 January 2014 - Baltimore JMM .5 5 5 / 13 Averaging Method (continued) Solve the system: Dx = s D = diag[d1 , d2 , . . . , dn ] = diagonal degree matrix x = ratings vector s = total score vector Flaw: Too local If we increase the score for the partnership of players 1 and 2, then 1 4 .62 x1 and x2 increase x3 should decrease, but doesn’t T. S. Michael (U. S. Naval Academy) 2 How To Rate Bridge Players 3 .5 .4 .5 .5 .5 .6 January 2014 - Baltimore JMM 5 6 / 13 Least Squares Method (Quint) Idea: Players with ratings xi and xj should get partnership score = xi + xj 2 Partnership scores p1 , p2 , . . . , pm are observations of the ratings x1 , x2 , . . . , xn . Choose ratings to minimize sum of squared errors m X xik + xjk 2 E= pk − . 2 k =1 sum is over all edges {ik , jk } of G T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 7 / 13 Least Squares Method (continued) Set derivatives of E equal to 0 to get a system of equations. m X xik + xjk ∂E 0= =− pk − ∂xi 2 k =1 (i = 1, . . . , n). i∈{ik ,jk } sum is over all edges {ik , jk } incident with node i of G .. . Solve the system: (D + A)x = 2s D = diag[d1 , d2 , . . . , dn ] = diagonal degree matrix A = adjacency matrix of partnership multigraph G x = ratings vector s = total score vector T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 8 / 13 Least Squares Method: Example 2 1 1 0 0 1 3 2 0 0 1 2 5 1 1 0 0 1 2 1 0 0 1 1 2 x1 x2 x3 x4 x5 = 2 1.12 1.62 2.50 1.00 1.00 1 4 .62 2 3 .5 .4 .5 .5 .5 .6 5 x1 = 0.61 x2 = 0.69 x3 = 0.44 x4 = 0.52 x5 = 0.52 If we increase the score for the partnership of players 1 and 2, then x1 and x2 increase x3 decreases x4 and x5 increase Method is global. It accounts for strengths of partners. T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 9 / 13 When Is D + A Singular? The matrix D + A is the unsigned Laplacian matrix of the (partnership) multigraph G. It is studied in graph theory. Theorem Let G be a connected multigraph with unsigned Laplacian matrix D + A. The following are equivalent. D + A is singular nullity(D + A) = 1 G is bipartite T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 10 / 13 Least Squares Method: A Singular Example | 3 2 1 0 0 2 3 0 1 0 1 0 0 1 3 1 1 2 1 0 {z singular 0 0 1 0 1 x1 x2 x3 x4 x5 = 2 1.50 1.50 1.65 1.10 0.55 } 1 .4 2 .6 .5 3 .55 5 .5 .6 4 x = 0.38 x2 = 0.59 x3 = 0.68 x4 = 0.47 x5 = 0.42 |1 {z } use Moore-Penrose inverse T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 11 / 13 Least Squares Method: Reno We used our least squares method to rate the bridge players in the Reno Bridge Club. The partnership multigraph G had n = 182 vertices (players) one giant (non-singular) component I x = 2(D + A)−1 s sparse linear system 30 small bipartite components I x = 2(D + A)+ s Moore-Penrose inverse The computed ratings seemed reasonable to the players. We can try the same least squares method to assign ratings to individuals in any sport where players compete in pairs. T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 12 / 13 References Quint A new rating system for duplicate bridge Linear Algebra and Its Applications 2007 Michael, Mortensen, and Quint Rating bridge players: The bipartite components in preparation T. S. Michael (U. S. Naval Academy) How To Rate Bridge Players January 2014 - Baltimore JMM 13 / 13