Estimate an opponent's rating from the moves they played
Source:R/fct_rating_estimate.R
estimate_rating.RdScores every move the given side played under each Maia network and returns the posterior over ratings.
Arguments
- pool
A pool from
maia_pool_start().- fens
Positions before each move, as recorded by the game tracker.
- ucis
Moves played, aligned with
fens.- side
Which side to model, "w" or "b";
NULLscores every move.
Value
A list with posterior (named, sums to 1), rating (the MAP
estimate), n_moves scored, evidence in nats, and loglik per rating.
Details
evidence is the part worth reading before the estimate. It is the summed
spread of per-move log-probabilities across networks, in nats: how much the
observed moves actually distinguished one network from another. A game of
forced recaptures and obvious developing moves can run twenty plies and
carry almost none, because every network would have played the same thing.
Move count alone does not tell you that.