Maps the heuristic evaluation score to a [0, 1] win estimate for the
player-to-move, conditional on a decisive outcome: draw rows are excluded
before fitting. Logistic and isotonic fits are supported; both produce
compact, serialisable mxo_calibrator objects. Brier and reliability
summaries describe the fitting data, not held-out predictive performance.
Usage
mxo_fit_calibration(data, type = c("logistic", "isotonic"))Arguments
- data
A tibble of the form produced by
mxo_make_calibration_data().- type
One of
"logistic"(default) or"isotonic".
Examples
cd <- mxo_make_calibration_data(
n_games = 3L, config = mxo_config_default(n = 3L, ply_cap = 12L),
seed = 1L
)
mxo_fit_calibration(cd)
#>
#> ── mxo_calibrator ──────────────────────────────────────────────────────────────
#> ℹ Type: logistic
#> ℹ a = 7e-06, b = 0.222999
#> ℹ n decisive plies: 10
#> ℹ Brier: 0.2226 (baseline 0.25)
mxo_fit_calibration(cd, type = "isotonic")
#>
#> ── mxo_calibrator ──────────────────────────────────────────────────────────────
#> ℹ Type: isotonic
#> ℹ n decisive plies: 10
#> ℹ Brier: 0.2222 (baseline 0.25)