Probabilistic Ranking: TrueSkill

Transcription

1 Probabilistic Ranking: TrueSkill Zoubin Ghahramani Department of Engineering University of Cambridge, UK Wroc law Lectures 013

2 Motivation for ranking Competition is central to our lives. It is an innate biological trait, and the driving principle of many sports. In the Ancient Olympics (700 BC), the winners of the events were admired and immortalised in poems and statues. Today in pretty much every sport there are player or team rankings. (Football leagues, Poker tournament rankings, etc). We are going to focus on one example: tennis players, in men singles games. We are going to keep in mind the goal of answering the following question: What is the probability that player 1 defeats player? Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking /

3 The ATP ranking system for tennis players Men Singles ranking as of 8 December 011: Rank, Name & Nationality Points Tournaments Played 1 Djokovic, Novak (SRB) 13, Nadal, Rafael (ESP) 9, Federer, Roger (SUI) 8, Murray, Andy (GBR) 7, Ferrer, David (ESP) 4, Tsonga, Jo-Wilfried (FRA) 4, Berdych, Tomas (CZE) 3, Fish, Mardy (USA), Tipsarevic, Janko (SRB), Almagro, Nicolas (ESP), Del Potro, Juan Martin (ARG),315 1 Simon, Gilles (FRA), Soderling, Robin (SWE),10 14 Roddick, Andy (USA) 1, Monfils, Gael (FRA) 1, Dolgopolov, Alexandr (UKR) 1, Wawrinka, Stanislas (SUI) 1, Isner, John (USA) 1, Gasquet, Richard (FRA) 1, Lopez, Feliciano (ESP) 1,755 8 ATP: Association of Tennis Professionals ( Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 3 /

4 The ATP ranking system explained (to some degree) Sum of points from best 18 results of the past 5 weeks. Mandatory events: 4 Grand Slams, and 8 Masters 1000 Series events. Best 6 results from International Events (4 of these must be 500 events). Points breakdown for all tournament categories (01): W F SF QF R16 R3 R64 R18 Q Grand Slams Barclays ATP World Tour Finals *1500 ATP World Tour Masters (5) (10) (1)5 ATP (0) ()0 ATP (5) (3)1 Challenger 15,000 +H Challenger 15, Challenger 100, Challenger 75, Challenger 50, Challenger 35,000 +H Futures** 15,000 +H Futures** 15, Futures** 10, The Grand Slams are the Australian Open, the French Open, Wimbledon, and the US Open. The Masters 1000 Tournaments are: Cincinnati, Indian Wells, Madrid, Miami, Monte-Carlo, Paris, Rome, Shanghai, and Toronto. The Masters 500 Tournaments are: Acapulco, Barcelona, Basel, Beijing, Dubai, Hamburg, Memphis, Rotterdam, Tokyo, Valencia and Washington. The Masters 50 Tournaments are: Atlanta, Auckland, Bangkok, Bastad, Belgrade, Brisbane, Bucharest, Buenos Aires, Casablanca, Chennai, Delray Beach, Doha, Eastbourne, Estoril, Gstaad, Halle, Houston, Kitzbuhel, Kuala Lumpur, London, Los Angeles, Marseille, Metz, Montpellier, Moscow, Munich, Newport, Nice, Sao Paulo, San Jose, s-hertogenbosch, St. Petersburg, Stockholm, Stuttgart, Sydney, Umag, Vienna, Vina del Mar, Winston-Salem, Zagreb and Dusseldorf. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 4 /

5 A laundry list of objections and open questions Some questions: Rank, Name & Nationality Points 1 Djokovic, Novak (SRB) 13,675 Nadal, Rafael (ESP) 9,575 3 Federer, Roger (SUI) 8,170 4 Murray, Andy (GBR) 7,380 Is a player ranked higher than another more likely to win? What is the probability that Murray defeats Djokovic? How much would you (rationally) bet on Murray? And some concerns: The points system ignores who you played against. 6 out of the 18 tournaments don t need to be common to two players. Other examples: Premier League. Meaningless intermediate results throughout the season: doesn t say whom you played and whom you didn t! Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 5 /

6 Towards a probabilistic ranking system What we really want is to infer is a player s skill. Skills must be comparable: a player of higher skill is more likely to win. We want to do probabilistic inference of players skills. We want to be able to compute the probability of a game outcome. A generative model for game outcomes: 1 Take two tennis players with known skills (w i R) Player 1 with skill w 1. Player with skill w. Compute the difference between the skills of Player 1 and Player : s = w 1 - w 3 Add noise (n N(0, 1)) to account for performance inconsistency: t = s + n 4 The game outcome is given by y = sign(t) y =+1 means Player 1 wins. y = -1 means Player wins. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 6 /

7 TrueSkill, a probabilistic skill rating system w 1 and w are the skills of Players 1 and. We treat them in a Bayesian way: prior p(w i ) = N(w i µ i, i) s = w 1 - w is the skill difference. t N(t s,1) is the performance difference. The probability of outcome y given known skills is: ZZ p(y w 1, w ) = p(y t)p(t s)p(s w 1, w )dsdt likelihood The posterior over skills given the game outcome is: p(w 1, w y) = p(w 1 )p(w )p(y w 1, w ) RR p(w1 )p(w )p(y w 1, w )dw 1 dw TrueSkill : A Bayesian Skill Rating System. Herbrich, Minka and Graepel, NIPS19, 007. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 7 /

8 The Likelihood ZZ Z p(y w 1, w ) = p(y t)p(t s)p(s w 1, w )dsdt = p(y t)p(t w 1, w )dt = (y(w 1 - w )). where (a) = Z a -1 N(x; 0, 1)dx (a) is the Gaussian cumulative distribution function, or probit function. density p(x) and cumulative probability Φ(x) density, N(0,1) cumulative probability, Φ x Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 8 /

9 The likelihood in a picture Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 9 /

10 An intractable posterior The joint posterior distribution over skills does not have a closed form: p(w 1, w y) = N(w 1 ; µ 1, 1 )N(w ; µ, ) (y(w 1 - w )) RR N(w1 ; µ 1, 1 )N(w ; µ, ) (y(w 1 - w ))dw 1 dw w 1 and w become correlated, the posterior does not factorise. The posterior is no longer a Gaussian density function. The normalising constant of the posterior, the prior over y does have closed form: ZZ y(µ1 - µ ) p(y) = N(w 1 ; µ 1, 1)N(w ; µ, ) (y(w 1 - w ))dw 1 dw = q This is a smoother version of the likelihood p(y w 1, w ). Can you explain why? Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 10 /

11 Joint posterior after several games Each player playes against multiple opponents, possibly multiple times; what does the joint posterior look like? skill, w skill, w 1 skill, w skill, w 1 The combined posterior is difficult to picture. skill, w skill, w How do we do inference with an ugly posterior like that? How do we predict the outcome of a match? Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 11 /

12 How do we do integrals wrt an intractable posterior? Approximate expectations of a function (x) wrt probability p(x): Z E p(x) [ (x)] = = (x)p(x)dx, where x R D, when these are not analytically tractable, and typically D Assume that we can evaluate (x) and p(x). Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 1 /

13 Numerical integration on a grid Approximate the integral by a sum of products Z TX (x)p(x)dx ' (x ( ) )p(x ( ) ) x, where the x ( ) lie on an equidistant grid (or fancier versions of this). = Problem: the number of grid points required, k D, grows exponentially with the dimension D. Practicable only to D = 4 or so. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 13 /

14 Monte Carlo The foundation identity for Monte Carlo approximations is E[ (x)] ' ˆ = 1 TX (x ( ) ), where x ( ) p(x). T = Under mild conditions, ˆ! E[ (x)] as T! 1. For moderate T, ˆ may still be a good approximation. In fact it is an unbiased estimate with V[ ˆ ] = V[ ] Z, where V[ ] = (x) - p(x)dx. T Note, that this variance in independent of the dimension D of x. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 14 /

15 Markov Chain Monte Carlo This is great, but how do we generate random samples from p(x)? If p(x) has a standard form, oo one-dimensional we may be able to generate independent samples. Idea: could we design a Markov Chain, q(x 0 x), which generates (dependent) samples from the desired distribution p(x)? One such algorithm is called Gibbs sampling: for each component i of x in turn, sample a new value from the conditional distribution of x i given all other variables: x i p(x i x 1,...,x i-1, x i+1,...,x D ). It can be shown, that this will generate dependent samples from the joint distribution p(x). Gibbs sampling reduces the task of sompling from a joint distribution, to sampling from a sequence of univariate conditional distributions. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 15 /

16 Gibbs sampling example: Multivariate Gaussian 0 iterations of Gibbs sampling on a bivariate Gaussian; both conditional distributions are Gaussian. Notice that strong correlations can slow down Gibbs sampling. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 16 /

17 Gibbs Sampling Gibbs sampling is a parameter free algorithm, applicable if we know how to sample from the conditional distributions. Main disadvantage: depending on the target distribution, there may be very strong correlations between consequtive samples. To get less dependence, Gibbs sampling is often run for a long time, and the samples are thinned by keeping only every 10th or 100th sample. It is often challenging to judge the effective correlation length of a Gibbs sampler. Sometimes several Gibbs samplers are run form different starting points, to compare results. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 17 /

18 Gibbs sampling for the TrueSkill model We have g = 1,..., N games where I g : id of Player 1 and J g : id of Player. The outcome of game g is y g =+1 if I g wins, y g = -1 if J g wins. Gibbs sampling alternates between sampling skills w =[w 1,...,w M ] > conditional on fixed performance differences t =[t 1,...,t N ] >, and sampling t conditional on fixed w. 1 Initialise w, e.g. from the prior p(w). Sample the performance differences from p(t g w Ig, w Jg, y g ) / (y g - sign(t g ))N(t g ; w Ig - w Jg,1) 3 Jointly sample the skills from 4 Go back to step. p(w t) {z } / p(w) {z } N(w;µ, ) g=1 N(w;µ 0, 0 ) NY p(t g w Ig, w Jg ) {z } / N(w;µ g, g ) Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 18 /

19 Gaussian identities The conditional for the performance is both Gaussian in t g and proportional to a Gaussian in w p(t g w Ig, w Jg ) / exp - 1 (w I g - w Jg - t g ) / N - 1 w Ig - µ h 1 > 1-1 i w Jg - µ -1 1 w Ig - µ 1 w Jg - µ with µ 1 - µ = t g. Notice that h i µ 1 µ = t g -t g Remember that for products of Gaussians precisions add up, and means weighted by precisions (natural parameters) also add up: N(w; µ a, a )N(w; µ b, b )=z c N w; c ( -1 a µ a + -1 b µ b), c =( -1 a + -1 b )-1 Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 19 /

20 Conditional posterior over skills given performances We can now compute the covariance and the mean of the conditional posterior. NX NX -1 = g µ = -1 0 µ g µ g g=1 {z } -1 To compute the mean it is useful to note that: NX µ i = t g (i - I g ) - (i - J g ) g=1 And for the covariance we note that: NX [ -1 ] ii = (i - I g )+ (i - J g ) g=1 [ -1 ] i6=j = - NX g=1 g=1 {z } µ (i - I g ) (j - J g )+ (i - J g ) (j - I g ) Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 0 /

21 Implementing Gibbs sampling for the TrueSkill model we have derived the conditional distribution for the performance differences in game g and for the skills. These are: the posterior conditional performance difference for t g is a univariate truncated Gaussian. How can we sample from it? by rejection sampling from a Gaussian, or by the inverse transformation method (passing a uniform on an interval through the inverse cumulative distribution function). the conditional skills can be sampled jointly form the corresponding Gaussian (using the cholesky factorization of the covariance matrix). Once samples have been drawn from the posterior, these can be used to make predictions for game outcomes, using the generative model. How would you do this? Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking 1 /

22 (a) is the Gaussian cumulative distribution function, or probit function. Rasmussen & Ghahramani (CUED) Lecture 6 and 7: Probabilistic Ranking / Appendix: The likelihood in detail ZZ Z p(y w 1, w ) = p(y t)p(t s)p(s w 1, w )dsdt = p(y t)p(t w 1, w )dt = = = y = = Z +1-1 Z +1-1 Z +y1 -y1 Z +1-1 Z +1 0 (y - sign(t))n(t w 1 - w,1)dt (1 - sign(yt))n(yt y(w 1 - w ),1)dt (1 - sign(z))n(z y(w 1 - w ),1)dz (use z yt) (1 - sign(z))n(z y(w 1 - w ),1)dz N(z y(w 1 - w ),1)dz = Z y(w1 -w ) -1 N(x 0, 1)dx (use x y(w 1 - w ) - z) = (y(w 1 - w )) where (a) = Z a -1 N(x 0, 1)dx