jazzzooo / lichess-ordo Goto Github PK

Would you mind commenting to explain what these are. Once I understand that, I think the rest is self-explaining

n_players = int.from_bytes(f.read(4), "little")
n_opponents = np.fromfile(f, dtype=np.uint16, count=n_players).astype(np.intc)
total = np.fromfile(f, dtype=np.uint16, count=n_players).astype(np.float64)
scored = np.fromfile(f, dtype=np.uint16, count=n_players).astype(np.float64) / 2
opponents = np.ascontiguousarray(np.fromfile(f, dtype=np.uint32, count=n_pairings).astype(np.intc))
opp_played = np.ascontiguousarray(np.fromfile(f, dtype=np.uint16, count=n_pairings).astype(np.intc))

Currently you fit using predicted_score = L(eloA-eloB), with L(x)=1/(1+10^(-x/400)). This is the standard ELO model.

Instead you could try adding the eloDraw factor from BayesElo model like that:

P(A wins)=L(eloA-eloB-eloDraw)
P(B wins)=L(eloB-eloA-eloDraw)
P(draw)=1-P(A wins)-P(B wins)

From which you get the predicted_score for the pairing

predicted=P(A wins)+0.5*P(draw)

And fit the elo as you do now.

To make things easier, you can treat eloDraw as a hyper parameter. In fact you can use the global statistics of your total sample and the equation above to calculate eloDraw analytically.

In theory this should improve the quality of fit, and out of sample predictions. In practice this has yet to be demonstrated empirically on such a dataset, and also put in the balance with speed performance